"Front Matter". In: Handbook of Machine Olfaction

T. C. Pearce, S. S. Schiffman, H.T. Nagle, J.W. Gardner

Handbook of Machine Olfaction

Handbook of Machine Olfaction: Electronic Nose Technology. Edited by T.C. Pearce, S.S. Schiffman, H.T. Nagle, J.W. Gardner Copyright ª 2003 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim ISBN: 3-527-30358-8 Related titles:

J.W. Gardner, V.K. Varadan Microsensors, MEMS and Smart Devices ISBN 0-471-86109-X

H.K. To¨nshoff, I. Inasaki Sensors in Manufacturing ISBN 3-527-29558-5

O. Gassmann, H. Meixner Sensors in Intelligent Buildings ISBN 3-527-29557-7

H. Baltes, G.K. Fedder, J.G. Korvink Sensors Update ISSN 1432-2404 T. C. Pearce, S. S. Schiffman, H.T. Nagle, J.W. Gardner

Handbook of Machine Olfaction

Electronic Nose Technology Tim C. Pearce, PhD This book was carefully produced. Nevertheless, Department of Engineering authors, editors and publisher do not warrant the University of Leicester information contained therein to be free of errors. Leicester LE1 7RH Readers are advised to keep in mind that statements, U.K. data, illustrations, procedural details or other items may inadvertently be inaccurate. Prof. Susan S. Schiffman Department of Psychiatry Library of Congress Card No. applied for. Duke University Medical School 54212 Woodhall Building British Library Cataloguing-in-Publication P.O. Box 3259 Data: Durham, NC 27710 A catalogue record for this book is available USA from the British Library.

Prof. H. Troy Nagle Bibliographic information published by Die Deutsche Department of Electrical Bibliothek and Computer Engineering Die Deutsche Bibliothek lists this publication in North Carolina State University the Deutsche Nationalbibliografie; detailed 432 Daniels Hall bibliographic data is available in the Internet at Raleigh, NC 27695-7911 . USA ª 2003 WILEY-VCH Verlag GmbH & Co. KGaA, Prof. Julian W. Gardner Weinheim Division of Electrical & Electronic Engineering All rights reserved (including those of translation The University of Warwick into other languages). No part of this book may be Coventry CV4 7AL reproduced in any form – by photoprinting, U.K. microfilm, or any other means – nor transmitted or translated into a machine language without written permission from the publishers. Registered names, trademarks, etc. used in this book, even when not specifically marked as such, are not to be considered unprotected by law.

Printed in the Federal Republic of Germany

Printed on acid-free paper

Typesetting Mitterweger & Partner, Kommunikationsgesellschaft mbH, Plankstadt Printing and Bookbinding Druckhaus Darmstadt GmbH, Darmstadt

ISBN 3-527-30358-8 Contents

1 Introduction to Olfaction: Perception, Anatomy, Physiology, and Molecular Biology 1 1.1 Introduction to Olfaction 1 1.2 Odor Classification Schemes Based on Adjective Descriptors 4 1.3 Odor Classification Based on Chemical Properties 7 1.3.1 History of Structure-activity Studies of Olfaction 8 1.3.2 Odor Structures Associated with Specific Odor Classes Based on Qualitative Descriptors 8 1.3.3 Relationship of Physicochemical Parameters to Classifications of Odor Based on Similarity Measures 11 1.3.3.1 Study 1: Broad Range of Unrelated Odorants 12 1.3.3.2 Study 2: Pyrazines 14 1.3.4 Molecular Parameters and Odor Thresholds 16 1.3.5 Conclusions Regarding Physicochemical Parameters and Odor Quality 16 1.4 Physiology and Anatomy of Olfaction 17 1.4.1 Basic Anatomy 17 1.4.2 Transduction and Adaptation of Olfactory Signals 20 1.5 Molecular Biology Of Olfaction 21 1.6 Taste 23 1.6.1 Taste Classification Schemes Based on Sensory Properties 23 1.6.2 Physiology and Anatomy of Taste 23 1.6.3 Transduction of Taste Signals 25 1.6.4 Molecular Biology of Taste 25 1.7 Final Comment 26

2 Chemical Sensing in Humans and Machines 33 2.1 Human Chemosensory Perception of Airborne Chemicals 33 2.2 Nasal Chemosensory Detection 34 2.2.1 Thresholds for Odor and Nasal Pungency 35 2.2.2 Stimulus-Response (Psychometric) Functions for Odor and Nasal Pungency 37 2.3 Olfactory and Nasal Chemesthetic Detection of Mixtures of Chemicals 38 2.4 Physicochemical Determinants of Odor and Nasal Pungency 39

2.4.1 The Linear Solvation Model 39 2.4.2 Application of the Solvation Equation to Odor and Nasal Pungency Thresholds 40 2.5 Human Chemical Sensing: Olfactometry 42 2.5.1 Static Olfactometry 42 2.5.2 Dynamic Olfactometry 44 2.5.3 Environmental Chambers 45 2.6 Instruments for Chemical Sensing: Gas Chromatography-Olfactometry 47 2.6.1 Charm Analysis 48 2.6.2 Aroma Extract Dilution Analysis (AEDA) 49 2.6.3 Osme Method 50

3 Odor Handling and Delivery Systems 55 3.1 Introduction 55 3.2 Physics of Evaporation 56 3.3 Sample Flow System 57 3.3.1 Headspace Sampling 57 3.3.2 Diffusion Method 60 3.3.3 Permeation Method 61 3.3.4 Bubbler 61 3.3.5 Method using a Sampling Bag 62 3.4 Static System 64 3.5 Preconcentrator 65 3.5.2 Sensitivity Enhancement 65 3.5.2 Removal of Humidity 66 3.5.3 Selectivity Enhancement by Varying Temperature 66 3.5.3.1 Selectivity Enhancement using a Preconcentrator 66 3.5.3.2 Autonomous System with Plasticity 67 3.5.3.3 Experiment on Plasticity 69 3.6 Measurement of Sensor Directly Exposed to Ambient Vapor 70 3.6.1 Analysis of Transient Sensor Response using an Optical Tracer 70 3.6.2 Homogenous Sensor Array for Visualizing Gas/Odor Flow 72 3.6.3 Response of Sensor Mounted on an Odor-Source Localization System 74 3.7 Summary 74

4 Introduction to Chemosensors 79 4.1 Introduction 79 4.2 Survey and Classification of Chemosensors 79 4.3 Chemoresistors 81 4.3.1 MOS 81 4.3.2 Organic CPs 84 4.4 Chemocapacitors (CAP) 87 4.5 Potentiometric Odor Sensors 88 4.5.1 MOSFET 88 4.6 Gravimetric Odor Sensors 89 Contents VII

4.6.1 QCM 90 4.6.2 SAW 92 4.7 Optical Odor Sensors 93 4.7.1 SPR 93 4.7.2 Fluorescent Odor Sensors 94 4.7.3 Other Optical Approaches 95 4.8 Thermal (Calorimetric) Sensors 96 4.9 Amperometric Sensors 96 4.10 Summary of Chemical Sensors 98

5 Signal Conditioning and Preprocessing 105 5.1 Introduction 105 5.2 Interface Circuits 106 5.2.1 Chemoresistors 106 5.2.1.1 Voltage Dividers 106 5.2.1.2 The Wheatstone Bridge 108 5.2.1.3 AC Impedance Spectroscopy 109 5.2.2 Acoustic Wave Sensors 110 5.2.3 Field-Effect Gas Sensors 112 5.2.4 Temperature Control 113 5.3 Signal Conditioning 114 5.3.1 Operational Amplifiers 114 5.3.2 Buffering 116 5.3.3 Amplification 116 5.3.4 Filtering 116 5.3.5 Compensation 118 5.3.5.1 Linearization of Resistance Measurements 118 5.3.5.2 Miscellaneous Functions 119 5.4 Signal Preprocessing 120 5.4.1 Baseline Manipulation 120 5.4.2 Compression 122 5.4.3 Normalization 123 5.4.3.1 Local Methods 123 5.4.3.2 Global Methods 125 5.5 Noise in Sensors and Circuits 125 5.6 Outlook 128 5.6.1 Temperature Modulation 128 5.7 Conclusions 129 5.8 Acknowledgements 130

6 Pattern Analysis for Electronic Noses 133 6.1 Introduction 134 6.1.1 Nature of Sensor Array Data 135 6.1.2 Classification of Analysis Techniques 136 6.1.3 Overview 137 VIII Contents

6.2 Statistical Pattern Analysis Techniques 138 6.2.1 Linear Calibration Methods 139 6.2.2 Linear Discriminant Analysis (LDA) 140 6.2.3 Principal Components Analysis (PCA) 141 6.2.4 Cluster Analysis (CA) 143 6.3 ‘Intelligent’ Pattern Analysis Techniques 145 6.3.1 Multilayer Feedforward Networks 146 6.3.2 Competitive and Feature Mapping Networks 150 6.3.3 ‘Fuzzy’ Based Pattern Analysis 152 6.3.4 Neuro-Fuzzy Systems (NFS) 154 6.4 Outlook and Conclusions 155 6.4.1 Criteria for Comparison 155 6.4.2 Intelligent Sensor Systems 157 6.4.3 Conclusions 158

7 Commercial Electronic Nose Instruments 161 7.1 Introduction 161 7.1.1 Geographical Expansion 162 7.1.2 Scientific and Technological Broadening 162 7.1.3 Conceptual Expansion 163 7.2 Commercial Availability 164 7.2.1 Global Market Players 164 7.2.1.1 Alpha M.O.S. 165 7.2.1.2 AppliedSensor Group 165 7.2.1.3 Lennartz Electronic 167 7.2.1.4 Marconi Applied Technologies (now ELV Technologies) 167 7.2.1.5 Osmetech plc 168 7.2.2 Handheld Devices 170 7.2.2.1 AppliedSensor Group 170 7.2.2.2 Cyrano Sciences, Inc. 170 7.2.2.3 Microsensor Systems, Inc. 171 7.2.3 Enthusiastic Sensor Developers 171 7.2.3.1 Bloodhound Sensors Ltd. 171 7.2.3.2 HKR Sensorsysteme GmbH 171 7.2.3.3 OligoSense n.v. 172 7.2.3.4 Quality Sensor Systems Ltd. 172 7.2.3.5 Quartz Technology Ltd. 172 7.2.3.6 Technobiochip 173 7.2.4 Non-Electronic Noses 173 7.2.4.1 Laboratory of Dr. Zesiger 173 7.2.4.2 Agilent Technologies, Inc. 174 7.2.4.3 Illumina, Inc. 174 7.2.4.4 Electronic Sensor Technology, Inc. 174 7.2.5 Specific Driven Applications 175 7.2.5.1 Astrium 175 Contents IX

7.2.5.2 Element Ltd. 175 7.2.5.3 Environics Industry Oy 175 7.2.5.4 WMA Airsense Analysentechnik GmbH 175 7.3 Some Market Considerations 176

8 Optical Electronic Noses 181 8.1 Introduction 181 8.1.1 Optical Sensors 181 8.1.2 Advantages and Disadvantages of Optical Transduction 182 8.2 Optical Vapor Sensing 183 8.2.1 Waveguides 183 8.2.2 Luminescent Methods 183 8.2.3 Colorimetric Methods 185 8.2.4 Surface Plasmon Resonance (SPR) 187 8.2.5 Interference and Reflection-Based Methods 189 8.2.6 Scanning Light-Pulse Technique 191 8.3 The Tufts Artificial Nose 191 8.4 Conclusion 198

9 Hand-held and Palm-Top Chemical Microsensor Systems for Gas Analysis 201 9.1 Introduction 201 9.2 Conventional Hand-held Systems 203 9.2.1 Hardware Setup 203 9.2.2 Fundamentals of the Sensing Process 205 9.2.3. Commercially Available Instruments Based on Conventional Technology 206 9.2.3.1 Hand-held Units Based on Mass-Sensitive Sensors 207 9.2.3.2 Hand-held Units Based on Chemoresistors 210 9.3 Silicon-Based Microsensors 211 9.3.1 Micromachining Techniques 212 9.3.1.1 Bulk Micromachining 212 9.3.1.2 Surface Micromachining 213 9.3.2 Microstructured Chemocapacitors 213 9.3.3 Micromachined Resonating Cantilevers 216 9.3.4 Micromachined Calorimetric Sensors 219 9.3.5 Single-Chip Multisensor System 221 9.3.6 Operation Modes for CMOS Microsystems 223 9.3.6.1 Reverse Mode of Operation (RMO) 224 9.4 Summary and Outlook 226

10 Integrated Electronic Noses and Microsystems for Chemical Analysis 231 10.1 Introduction 231 10.2 Microcomponents for Fluid Handling 233 10.2.1 Microchannels and Mixing Chambers 233 10.2.2 Microvalves 238 X Contents

10.2.2.1 Active Microvalves 238 10.2.2.2 Passive Microvalves (Check Valves) 240 10.2.3 Micropumps 241 10.2.3.1 Mechanical Micropumps 241 10.2.3.2 Nonmechanical Micropumps 245 10.3 Integrated E-Nose Systems 245 10.3.1 Monotype Sensor Arrays 245 10.3.2 Multi-type Sensor Arrays 250 10.4 Microsystems for Chemical Analysis 251 10.4.1 Gas Chromatographs 251 10.4.2 Mass Spectrometers 255 10.4.3 Optical Spectrometers 258 10.5 Future Outlook 260

11 Electronic Tongues and Combinations of Artificial Senses 267 11.1 Introduction 267 11.2 Electronic Tongues 269 11.2.1 Measurement Principles 269 11.2.2 Potentiometric Devices 270 11.2.2.1 The Taste Sensor 271 11.2.2.2 Ion-Selective Electrodes 273 11.2.2.3 Surface Potential Mapping Methods 274 11.2.3 Voltammetric Devices 275 11.2.3.1 The Voltammetric Electronic Tongue 277 11.2.3.2 Feature Extraction 279 11.2.3.3 Industrial Applications using the Voltammetric Electronic Tongue 280 11.2.3 Piezoelectric Devices 283 11.3 The Combination or Fusion of Artificial Senses 284 11.3.1 The Combination of an Electronic Nose and an Electronic Tongue 285 11.3.2 The Artificial Mouth and Sensor Head 286 11.4 Conclusions 287

12 Dynamic Pattern Recognition Methods and System Identification 293 12.1 Introduction 293 12.2 Dynamic Models and System Identification 294 12.2.1 Linear Models 295 12.2.2 Multi-exponential Models 297 12.2.3 Non-linear Models 300 12.3 Identifying a Model 304 12.3.1 Non-Parametric Approach 304 12.3.1.1 Time-Domain Methods 305 12.3.1.2 Frequency-Domain Methods 307 12.3.2 Parametric Approach 308 12.4 Dynamic Models and Intelligent Sensor Systems 309 12.4.1 Dynamic Pattern Recognition for Selectivity Enhancement 311 Contents XI

12.4.2 Calibration Time Reduction 314 12.4.3 Building of Response Models 315 12.4.4 Drift Counteraction 317 12.5 Outlook 319

13 Drift Compensation, Standards, and Calibration Methods 325 13.1 Physical Reasons for Drift and Sensor Poisoning 325 13.2 Examples of Sensor Drift 329 13.3 Comparison of Drift and Noise 331 13.4 Model Building Strategies 332 13.5 Calibration Transfer 332 13.6 Drift Compensation 333 13.6.1 Reference Gas Methods 335 13.6.2 Modeling of Sensor Behavior 339 13.6.3 Pattern-Oriented Techniques for Classification 340 13.6.4 Drift-Free Parameters 343 13.6.5 Self-Adapting Models 343 13.7 Conclusions 344

14 Chemical Sensor Array Optimization: Geometric and Information Theoretic Approaches 347 14.1 The Need for Array Performance Definition and Optimization 347 14.2 Historical Perspective 349 14.3 Geometric Interpretation 351 14.3.1 Linear Transformations 352 14.4 Noise Considerations 355 14.4.1 Number of Discriminable Features 355 14.4.2 Measurement Accuracy 357 14.4.3 2-Sensor 2-Odor Example 360 14.5 Non-linear Transformations 363 14.6 Array Performance as a Statistical Estimation Problem 366 14.7 Fisher Information Matrix and the Best Unbiased Estimator 367 14.8 FIM Calculations for Chemosensors 369 14.8.1 2-Sensor 2-Odor Example 370 14.9 Performance Optimization 370 14.9.1 Optimization Example 371 14.10 Conclusions 373 14.A Overdetermined Case 375 14.B General Case with Gaussian Input Statistics 375 14.C Equivalence Between the Geometric Approach and the Fisher Information Maximization 375

15 Correlating Electronic Nose and Sensory Panel Data 377 15.2 Sensory Panel Methods 378 15.2.1 Odor Perception 378 XII Contents

15.2.2 Measurement of Detectability 379 15.2.3 Transforming the Measurement of the Subject to the Subject’s Measurement of an Odor 379 15.2.4 Assessor Selection 380 15.2.5 Types of Dynamic Dilution Olfactometry 380 15.2.5.1 Choice Modes 380 15.2.5.2 Yes/No Mode 381 15.2.5.3 The Forced Choice Mode 381 15.2.5.4 Laboratory Conditions 382 15.2.5.5 Laboratory Performance Quality Criteria 382 15.2.5.6 Compliance with the Quality Criteria 383 15.2.6 Assessment of Odor Intensity 384 15.2.7 Assessment of Odor Quality 386 15.2.8 Judgment of Hedonic Tone 387 15.3 Applications of Electronic Noses for Correlating Sensory Data 387 15.4 Algorithms for Correlating Sensor Array Data with Sensory Panels 388 15.4.1 Multidimensional Scaling 389 15.4.2 Regression Methods 390 15.4.3 Principal Components Regression 391 15.4.4 Partial Least Squares Regression 391 15.4.5 Neural Networks 392 15.4.6 Fuzzy-Based Data Analysis 392 15.5 Correlations of Electronic Nose Data with Sensory Panel Data 393 15.5.1 Data from Mouldy Grain 394 15.6 Conclusions 396

16 Machine Olfaction for Mobile Robots 399 16.1 Introduction 399 16.2 Olfactory-Guided Behavior of Animals 400 16.2.1 Basic Behaviors Found in Small Organisms 400 16.2.2 Plume Tracking 400 16.2.3 Trail Following by Ant 402 16.3 Sensors and Signal Processing in Mobile Robots 403 16.3.1 Chemical Sensors 403 16.3.2 Robot Platforms 404 16.4. Trail Following Robots 404 16.4.1 Odor Trails to Guide Robots 404 16.4.2 Robot Implementations 406 16.4.3 Engineering Technologies for Trail-Following Robots 406 16.5 Plume Tracking Robots 407 16.5.1 Chemotactic Robots 408 16.5.2 Olfactory Triggered Anemotaxis 410 16.5.3 Multiphase Search Algorithm 411 16.6 Other Technologies in Developing Plume Tracking Systems 413 16.6.1 Olfactory Video Camera 413 Contents XIII

16.6.2 Odor Compass 414 16.7 Concluding Remarks 416

17 Environmental Monitoring 419 17.1 Introduction 419 17.1.1 Water 419 17.1.2 Land 421 17.1.3 Air 421 17.2 Special Considerations for Environmental Monitoring 425 17.2.1 Sample Handling Problems 425 17.2.1.1 Sample Lifetime 425 17.2.1.2 Humidity 425 17.2.1.3 Extraction of volatiles 425 17.2.1.4 Tubing system 425 17.2.1.5 Temperature 425 17.2.2 Signal Processing Challenges 426 17.3 Case Study 1: Livestock Odor Classification 426 17.3.1 Background 426 17.3.2. Description of the problem 427 17.3.3. Methods 427 17.3.4 Signal Processing Algorithms 428 17.3.4.1 Bias Removal 428 17.3.4.2 Humidity 428 17.3.4.3 Concentration 428 17.3.4.4 Dimensionality Reduction 428 17.3.5. Results 429 17.3.6. Discussion 429 17.4 Case Study 2: Swine Odor Detection Thresholds 430 17.4.1. Description of the Problem 430 17.4.2 Methods 431 17.4.3 Results 431 17.4.4 Discussion 431 17.5 Case Study 3: Biofilter Evaluation 432 17.5.1 Description of the Problem 432 17.5.2 Methods 432 17.5.3. Results 434 17.5.4 Discussion 436 17.6 Case Study 4: Mold Detection 437 17.6.1 Background 437 17.6.2 Description of the Problem 437 17.6.3 The NC State E-Nose 437 17.6.4 Methods 440 17.6.5 Results 440 17.6.6 Discussion 441 17.7 Future Directions 441 XIV Contents

18 Medical Diagnostics and Health Monitoring 445 18.1 Introduction 445 18.2 Special Considerations in Medical/Healthcare Applications 449 18.3 Monitoring Metabolic Defects in Humans Using a Conducting Polymer Sensor Array to Measure Odor 450 18.3.1 Background 450 18.3.2 Methodology 451 18.3.3 Results 452 18.4 The Use of an Electronic Nose for the Detection of Bacterial Vaginosis 454 18.4.1 Background 454 18.4.2 Methodology 456 18.4.3 Results 456 18.4.4 Discussion 457 18.4.5 Conclusion 458 18.5 Conclusion 458

19 Recognition of Natural Products 461 19.1 Introduction 461 19.2 Recent Literature Review 462 19.3 Sampling Techniques 462 19.3.1 Sample Containment 462 19.3.2 Sample Treatments 468 19.3.2.1 Heating 468 19.3.2.2 Cooling 468 19.3.2.3 Removal of Base Component 468 19.3.2.4 Preconcentration 469 19.3.2.5 Grinding 469 19.3.3 Instrument and Sample Conditioning 469 19.3.3.1 Modifying Baseline 470 19.3.3.2 Purge Technique 470 19.3.3.3 Temperature Control 470 19.3.4 Sample Storage 470 19.3.5 Seasonal Variations 471 19.3.6 Inherent Variability of Natural Products 471 19.4 Case Study: The Rapid Detection of Natural Products as a Means of Identifying Plant Species 471 19.4.1 Wood Chip Sorting 472 19.4.2 Experimental Procedure 472 19.4.3 SPME-GC Analysis of the Sapwood of the Conifers Used in Pulp and Paper Industries 473 19.4.4 Conclusion: Wood Chip Sorting 475 19.5 Case Study: Differentiation of Essential Oil-Bearing Plants 475 19.5.1 Golden Rod Essential Oils 475 19.5.2 Essential Oils of Tansy 477 Contents XV

19.5.3 Conclusion: Essential Oils 478 19.6 Conclusion and Future Outlook 478

20 Process Monitoring 481 20.1 Introduction 481 20.1.1 On-line Bioprocess Monitoring 482 20.1.2 At-line Food Process Monitoring 483 20.2 Previous Work 483 20.2.1 Quantitative Bioprocess Monitoring 483 20.2.2 Qualitative Bioprocess Monitoring 485 20.2.3 At-line Food Process Monitoring 486 20.3 Special Considerations 487 20.4 Selected Process Monitoring Examples 487 20.4.1 On-line Monitoring of Bioprocesses 487 20.4.2 At-line Monitoring of a Feed Raw Material Production Process 488 20.4.3 Monitoring Setup 489 20.4.4 Signal Processing 489 20.4.5 Chemometrics 491 20.4.5.1 Study 1: Estimation of Cell growth in Escherichia coli Fermentations 491 20.4.5.2 Study 2: Physiologically Motivated Monitoring of Escherichia coli Fermentations 493 20.4.5.3 Study 3: Quality Control of a Slaughter Waste Process 496 20.4.5.4 Discussion 500 20.5 Future Prospects 501

21 Food and Beverage Quality Assurance 505 21.1 Introduction 505 21.2 Literature Survey 507 21.3 Methodological Issues in Food Measurement with Electronic Nose 510 21.4 Selected Case 511 21.4.1 LibraNose 511 21.4.2 Case Study: Fish Quality 515 21.5 Conclusions 520 21.6 Future Outlook 521

22 Automotive and Aerospace Applications 525 22.1 Introduction 525 22.2 Automotive Applications 525 22.3 Aerospace Applications 526 22.4 Polymer Composite Films 529 22.5 Electronic Nose Operation in Spacecraft 530 22.5.1 The JPL Enose Flight Experiment 532 22.5.2 Data Analysis 533 22.5.2.1 Data Pre-Processing 534 22.5.3 Pattern Recognition Method 536 XVI Contents

22.6 Method Development 536 22.6.1 Levenberg-Marquart Nonlinear Least Squares Method 537 22.6.2 Single gases 539 22.6.3 Mixed Gases 541 22.6.4 STS-95 Flight Data Analysis Results 541 22.7 Future Directions 543 22.7.1 Sensors 543 22.7.2 Data Acquisition 543 22.7.3 Data Analysis 544 22.8 Conclusion 545

23 Detection of Explosives 547 23.1 Introduction 547 23.2 Previous Work 548 23.3 State-of-the-art of Various Explosive Vapor Sensors 549 23.4 Case Study 557 23.5 Conclusions 559 23.6 Future Directions 559

24 Cosmetics and Fragrances 561 24.1 Introduction 561 24.2 The Case for an Electronic Nose in Perfumery 562 24.3 Current Challenges and Limitations of Electronic Noses 563 24.4 Literature Review of Electronic Noses in Perfumery and Cosmetics 564 24.5 Special Considerations for using Electronic Noses to Classify and Judge Quality of Perfumes, PRMs, and Products 566 24.6 Case Study 1: Use in Classification of PRMs with Different Odor Character but of Similar Composition 567 24.6.1 The Problem 567 24.6.2 Methods 568 24.6.3 Results 568 24.6.4 Conclusions for Case Study 1 570 24.7 Case Study 2: Use in Judging the Odor Quality of a Sunscreen Product 570 24.7.1 Background 570 24.7.2 The Problem 572 24.7.3 Equipment and Methods 573 24.7.3.1 Equipment 573 24.7.4 Results 574 24.7.4.1 Sensory Correlation and Long Term Repeatability 574 24.7.4.2 Database transfer from Du¨bendorf to Vernier 574 24.7.5 Conclusions for Case Study 2 575 24.8 Conclusions 575 24.9 Future Directions 576

Index 579 XVII

Preface

In the past decade, electronic nose instrumentation has generated much interest internationally for its potential to solve a wide variety of problems in fragrance and cosmetics production, food and beverages manufacturing, chemical engineering, environmental monitoring, and more recently, medical diagnostics and bioprocesses. Several dozen companies are now designing and selling electronic nose units globally for a wide variety of expanding markets. An electronic nose is a machine that is designed to detect and discriminate among complex odors using a sensor array. The sensor array consists of broadly tuned (non-specific) sensors that are treated with a variety of odor-sensitive biological or chemical materials. An odor stimulus generates a characteristic fingerprint (or smellprint) from the sensor array. Patterns or fingerprints from known odors are used to construct a database and train a pattern recognition system so that unknown odors can subsequently be classified and identified. Thus, electronic nose instruments are comprised of hardware components to collect and transport odors to the sensor array – as well as electronic circuitry to digitize and store the sensor responses for signal processing. This book provides a comprehensive and timely overview of our current state of knowledge of the use of electronic sensors for detection and identification of odorous compounds and mixtures. The handbook covers the scientific principles and technologies that are necessary to implement the use of an electronic nose. A comprehensive and definitive coverage of this emerging field is provided for both academic and prac- ticing scientists. The handbook is intended to enable readers with a specific background, e.g. sensor technology, to become acquainted with other specialist aspects of this very multidisciplinary field. Following this Preface, Part A covers the fundamentals of the key aspects related to electronic nose technology, from the biological olfactory system that has inspired the development of electronic nose technology, through to sensor materials and pattern analysis methods for use with chemical sensor arrays. This section provides a valuable tutorial for those readers who are new to the field before delving into the more specialist material in later chapters. More advanced aspects of the technology are dealt with in Parts B and C, which provide an up-to-date survey of current research directions in the areas of instrumentation (Part B) and pattern analysis (Part C). Advanced instrumentation issues include

novel sensing materials through to handheld chemical sensing devices and distributed chemosensory systems. Recent topics in pattern analysis include on-line learning methods to extend calibration life-span, dynamic signal processing methods that exploit sensor transient behavior and optimization strategies for chemical sensor arrays. An important element of the handbook is the inclusion of case studies of various applications of the electronic nose (Part D). Leading manufacturers of electronic nose equipment and key end-users have provided most of the chapters covering several interesting application areas.

Part A Overview: Fundamentals of Odor Sensing

Part A of the book is an overview of the fundamental key aspects of biological and machine olfaction. The section begins with two chapters that review the field of biological olfaction. The next four chapters cover the basic functional components of electronic noses including the sample handling system, gas sensor arrays and types, and signal processing systems for classification and identification of odorous compounds. The first chapter by Schiffman and Pearce describes how the biological sense of smell utilizes a remarkable sensor array of neurons that detects and discriminates among a vast number of volatile compounds (and mixtures of compounds) present in minute concentrations. This exquisite sensitivity is the reason why scientists and engineers have developed and begun to market machines that mimic this biological apparatus to detect and discriminate among volatile chemicals. The initial chapter provides an overview of the physicochemical and molecular properties of odorous molecules (called odorants) along with a description of odor classification and its limitations. It also provides an introduction to the biological olfactory pathway including descriptions of the olfactory epithelium, olfactory sensory neurons, seven-membrane-spanning receptors, the olfactory bulb, and the olfactory cortex. The chapter emphasizes that as few as 40 molecules of some compounds (e.g. mercaptans) are sufficient for humans to perceive an odor. Second, the range of distinctive odor sensations is vast, and a skilled perfume chemist can recognize and distinguish 8000 to 10 000 different substances on the basis of their odor quality. The remarkable discriminability is achieved by a coding scheme in which different odor stimuli are recognized by different combinations of olfactory receptors. That is, the biological olfactory system uses a combinatorial receptor coding scheme such that the specific patterns of activation across many neurons induced by an odor stimulus makes it possible to discriminate among the vast number of distinct smells. The second chapter of Part A by Cometto-Muniz expands on the first chapter with additional details of human olfactory perception and an overview of the topic of chemesthesis (the common chemical sense). Olfactory perception is achieved by stimulation of the olfactory nerve (cranial nerve I), which allows us to discriminate between odor stimuli such as chocolate and coffee. Chemesthetic sensations, on the other hand, include piquancy, prickling, stinging, burning, freshness, tingling, and irritation, which are grouped under the term pungency and are mediated by a different Preface XIX nerve called the trigeminal nerve (cranial nerve V). Airborne compounds elicit odor sensations at concentrations below those that induce pungency. Methods for quantifying odor and pungency in humans are described including the determination of thresholds, the relationship between concentration and perceived intensity, and the sensory consequences of adding multiple compounds together in a mixture. Ap- proaches for quantifying odor with static olfactometry, dynamic olfactometry, and environmental chambers are explained. In static olfactometry the vapor stimulus is drawn from an enclosed container in which the liquid and odorous vapor of the chemical(s) are in equilibrium with one another. In dynamic olfactometry, the vapor flows continuously in a carrier-gas stream, typically odorless air or nitrogen. A mathematical model is presented that can be used to predict odor and pungency threshold concentrations from physicochemical determinants. Instrumentation currently used by the flavor industry to analyze odorous mixtures including gas chromatography and mass spectrometry (GC/MS) is described. Overall, the sense of smell outperforms conventional analytic instruments (specifically GC/MS) in detecting and identifying odorous substances. The third chapter by Nakamoto covers basic principals of odor handling and delivery of samples to electronic noses with two main types of systems (flow and static) described. In flow systems, the sensors are placed in the vapor flow of the sampling system so that the vapor around the sensors is constantly exchanged. Several flow systems are described, including headspace sampling, diffusion and permeation methods, a bubbler, and sampling bags. In static systems there is no vapor flow around the sensors but rather the sensors are exposed to vapor with a constant concentration. For static systems, the steady-state response of the sensors is measured. An open system is also illustrated in which a sensor is directly exposed to a vapor without a sensor chamber. Because different types of sensors vary widely in their sensitivity, methods for increasing the sensitivity are described using a preconcentrator tube. The physics of evaporation are also covered because most samples submitted to electronic noses are liquids from which odorants are evaporated. Issues of removal of humidity from samples are also described. The fourth chapter by Nanto and Stetter is an overview of chemosensors that can be used in electronic nose systems to convert chemical information into an electrical signal. The chapter describes conductometric chemosensors (metal-oxide semiconductors (MOS) and conducting polymers (CPs)), chemocapacitors, potentiometric chemosensors (e.g. MOS field-effect transistors (MOSFETs)), gravimetric chemosensors (quartz crystal microbalance (QCM), surface acoustic wave (SAW)), optical chemosensors (surface plasmon resonance (SPR), fluorescent sensors), calorimetric sensors, and amperometric sensors. The underlying principle of conductometric sensors (also called chemoresistors) is the conductivity change that occurs when gaseous molecules react chemically with MOS or organic CPs. These are the simplest of type of gas sensors and are widely used to make arrays for gas and odor measurements. In chemocapacitor (CAP) devices, a polymer adsorbs the gaseous analyte, which alters the electrical (e.g. dielectric constant e) and physical properties (e.g. volume V) of the polymer relative to the baseline capacitance of the polymer when no gaseous analyte molecules are present. Potentiometric chemosensors of the MOSFET type utilize a XX Preface

gate that is made of a gas sensitive metal as a catalyst for gas sensing. Gravimetric odor sensors detect the effect of sorbed molecules on propagation of acoustic waves. The two main types of gravimetric sensors include QCM and SAW devices that are configured as mass-change sensing devices in the electronic nose. Optical chemosensors have several principals of operation. SPR is a physical process that can occur when plane-polarized light hits a metal film under total internal reflection conditions. In order to utilize this system as a gas sensor, a very thin film of methylmethacrylate, polyester resin or propylene ether as a sensing membrane can be deposited on gold metal thin film, and the angle of the reflected light is measured. Another type of chemosensor consists of optical fibers deposited with a fluorescent indicator dye in polymer matrices of varying polarity, hydrophobicity, pore size, elasticity, and swelling tendency to create unique sensing regions that interact differently with vapor molecules. Thermal sensors record the heat of solution of an analyte in the coating, with greater heat generated by larger amounts of absorbed analyte. The principle of amperometric gas sensors is the electrochemical oxidation or reduction of the analyte gas at a catalytic electrode surface that generates electrical current proportional to the concentration of the analyte. The next chapter by Gutierrez-Osuna, Nagle, Kermani, and Schiffman covers interface circuits, signal conditioning electronics, and pre-processing algorithms; topics that serve as a bridge between the previous chapter on odor sensors (see Nanto and Stetter Chapter 4) and the following chapter on pattern analysis techniques (Hines and colleagues Chapter 6). The chapter presents a review of interface circuits for the most widely used odor sensors (chemoresistive, acoustic wave, and field effect), as well as an introduction to analog conditioning circuits for signal amplification, filtering, and compensation. Signal preprocessing algorithms commonly used prior to pattern analysis, including baseline manipulation, compression, and normalization, are also reviewed. The final chapter in Section A by Hines, Boilot, Gardner, Gongora, Llobet deals with pattern analysis for electronic noses. There is an introduction into the nature of sensor array data and classification of analysis techniques including conventional statistical methods as well as biologically motivated technologies. This is followed by a more detailed discussion of statistical techniques such as principal components analysis (PCA), discriminant function analysis (DFA), partial least squares (PLS), multiple linear regression (MLR), and cluster analysis (CA) including nearest neighbor (NN). The discussion of biologically motivated technologies covers artificial neural networks (ANN), fuzzy inference systems (FIS), self-organizing map (SOM), radial basis function (RBF), genetic algorithms (GA), wavelets, neuro-fuzzy systems (NFS), and adaptive resonance theory (ART). Biologically motivated technologies for pattern analysis are especially attractive for use with electronic nose technology because they have the potential to perform incremental learning and offer self-organizing and self-stabilizing potential. Preface XXI

Part B Overview: Advanced Instrumentation

Part B of the book describes in some detail sensor technologies and instrumentation for electronic nose systems. The section begins with a chapter that reviews the field of electronic nose instruments that are currently available. These commercial instruments are predominantly large desktop-based systems that require an automated headspace sampler and a personal computer to operate the whole system. More recent instruments may be described as handheld but tend to have a limited battery life caused by either the need for the sensors to be held at a constant (elevated) temperature or high computing power. The next chapter considers the development of optical rather than solid-state electronic noses. In this type of instrument, chemically sensitive materials are used as the sensing elements. For example, Dickinson et al. describe the operation of an optical ‘smell camera’ based upon the 2D raster scanning of the surface of a distributed capacitor, in order to read out the charge generated by a local catalytic reaction with the odor molecule. The composition and temperature of the catalyst, making up one electrode of the capacitor, is varied to generate a 2D image of the smell. In a different approach, Walt et al. coat a large number of small glass beads with a variety of fluorescent indicator dyes and these are used to create pixels in a composite image of an odor. This involves the fixing of the beads on to the end of optical fibers to complete the transducer. The process has been simplified more recently by Suslick et al. who have created a small rectangular array of porphyrin based sensing elements that change their chromatic properties when exposed to reactive gases. This colorimetric electronic nose can work from an ordinary light source and CCD array, and so is quite similar in technology to a commercial color flatbed scanner. The concept of an opto-electronic electronic nose is an attractive one and it remains to be seen how this technology stands against the alternatives. The chapter by Baltes et al. explores the current research being undertaken in the development of small palm-top electronic noses. The approach focuses on the use of CMOS technology to fabricate a low-cost, low-power and miniature electronic nose. This necessitates the use of room-temperature gas-sensitive materials that can be deposited at a low temperature (compared with CMOS processes). Consequently, the chapter describes the development of capacitors, resistors, calorimeters, and cantilever beams predominantly coated with compounds used as the stationary phase in gas chromatography, i.e. rubbers and polymers. The fabrication of CMOS sensors permit the integration of CMOS or even BiCMOS circuitry next to the sensing elements and thus produce simple voltage read out. It is thus an attractive technology for the production of electronic noses at high volume, e.g. millions of units per year. Gardner et al. expands upon the concept of a micro nose and investigates the possible development of an electronic nose that has integrated mechanical as well as electrical components. There has been rapid progress in the field of micro electro mechanical systems in recent years and this chapter considers related advances in the fabrication of micro valves, micro pumps and other micro-fluidic components. The challenges associated with making an analytical instrument on a chip are also presented with a description of work being carried out to make micro gas chromatographs and XXII Preface

micro mass spectrometers. This approach poses a number of technological challenges because it has to deal with the transportation of the odor through the nose as well as the sensing elements. However these analytical micro noses may well compete with solid- state noses in terms of discriminating power. The final chapter describes the advances taking place to create another sensory instrument, namely, the ‘electronic tongue’. Clearly, an instrument that can mimic both the sense of smell and taste would provide valuable information on the nature of the flavor of a compound. In some ways the electronic tongue, as described, here behaves as an electronic nose under water – in other words the chemical sensors work in the liquid rather than gaseous phase. Thus the sensors are not specific to detecting the four primary tastes, salty, bitter, sweet, and sour (or putative additional taste primaries such as metallic and monosodium glutamate) but will provide signals that can be correlated with them. For example, the bitterness of a compound can be related to the acidity (i.e. pH value) while the sweetness will relate to the conductivity. The specificity of electrochemical sensors may be enhanced through the use of biological coatings of, for example, shear-horizontalmode SAW (SH-SAW) devices. Unfortunately, this type of biosensor tends to suffer (like all biosensors) from a short life when exposed to the environment. Nevertheless the development of electronic tongue technology could well lead to further advances in electronic nose technology.

Part C Overview: Advanced Signal Processing and Pattern Analysis

The foundations of signal processing strategies for chemical sensor array systems were provided in Chapter 6, which outlined the fundamentals of applying signal processing (predominantly pattern recognition based) techniques to chemical sensor arrays, for recognizing and discriminating specific ‘fingerprints’ of sensor array response that correspond to distinct categories of odor stimuli. This section of the handbook continues this theme by considering more advanced or, perhaps more accurately, specialized aspects of signal processing related to chemical sensor arrays – each chapter exploring fertile areas for future research in machine olfaction. A key theme here is the technological advantage that can be achieved in these systems through the development of their integral signal/information processing system. The chapters in this section are representative of current trends in research in this area that appear to emphasize two distinct aspects. First, the improvement in system performance through advances in information processing strategies applied to chemical sensor arrays, for example by considering transient sensor response (as opposed to the single-valued steady-state response) to enhance discrimination or the detection threshold of these instruments. Second, widening the scope of applications of such systems and solving novel chemosensory detection problems, for example by correlating quantitative electronic nose data with qualitative human sensory panel information in an attempt to achieve automated sensory panel analysis through technological means. The first of these themes looks more to the past, in terms of refining and improving on what has gone before, whereas the second theme is firmly looking to the future of this technology, in terms of opening up new domains in which the technology may be Preface XXIII applied. For this reason this section of the handbook provides a taste (!) of some exciting prospects for the future of electronic nose technology as we move further into the 21st century, which will be driven by parallel developments in sensor technology and information processing capability. The performance of electronic nose systems depends greatly on each of its components: from the odor delivery system; through to the choice and diversity of chemosensor materials; the interface circuitry; as well as the computational subsystem for discriminating between array responses. The first three chapters relate to the first theme – that is, how to improve system performance by developing signal-processing strategies that may be applied to machine olfaction. Although perhaps at first sight not quite as groundbreaking in its ambition as the second theme, the topics covered in these chapters are vital to the future welfare of this field as a commercial, scientific, and technological endeavor. Key issues are covered here that are important for overcoming existing technological barriers to the take-up and deployment of the technology. The first chapter in this section, by Llobet, covers aspects of dynamical model approaches for interpreting chemical sensor response information. Shifting the emphasis from steady-state sensor response information to transient sensor response promises less sensitivity to drift, the possibility of yielding additional discrimination of stimuli, and becomes essential when environmental conditions vary on a similar time scale as sensor response. An overview of a number of dynamical models and system identification techniques are provided alongside an example of how these might be applied to a specific sensing problem. In many cases the practical performance of chemical sensor array systems is limited by changes in characteristics of sensor response over time or with chemical exposure. Commercial systems require frequent calibration against known standardized samples in order to minimize these effects and assure some minimum measurement accuracy. In many cases, recalibration may be required on a daily basis in order to maintain acceptable performance in the field. Therefore, the development of signal-processing strategies that counteract the affect of these shifts in sensor characteristics to repeated and identical stimuli are of considerable importance to the practitioner and researcher. A true understanding of temporal drift in sensor characteristics will only ultimately be found through a detailed physical understanding of interaction of chemicals with sensing materials. Even then, only if the mechanisms involved are purely deterministic will it be possible to eliminate their effects entirely. In the mean- time, empirical methods for compensation can be developed and these are considered by Artursson and Holmberg in Chapter 13 as practical strategies for coping with this phenomenon in working instruments. Due to the distributed nature of chemical sensor arrays it is not simple to define their sensing performance in terms of the properties of the underlying chemical sensors. However, this is vital if a rigorous approach to specification of sensor performance and future optimization of sensor arrays is ever going to be achieved. Pearce and Sanchez-Montanes (Chapter 14) describe recent work on quantifying sensor array performance for multidimensional stimuli such as odors that allows the system detection performance to be predicted given the tuning and noise properties of the under- XXIV Preface

lying chemosensors. This allows the selection of chemosensors for specific detection tasks to be made, which until recently has been achieved by ad hoc means. In this chapter the theory of performance definition is applied to consider the practical issue of optimizing detection thresholds in artificial olfactory systems. The final two chapters of this section describe new domains where artificial olfactory systems find application. New areas of application open up to this technology all the time but future challenges will also require new and refined signal-processing strategies. Here we consider two areas where the signal-processing subsystems play a key part in this development. The first of these considers signal-processing strategies for correlating human-defined sensory panel information with chemical sensor-array responses. This has important consequences, particularly in the food and beverage industry where millions of dollars are spent each year on both instrumental analyses (mostly GC and MS-based methods) and sensory panel investigations. Neither of these approaches in isolation offers a complete picture of odor or flavor quality. By applying multivariate statistical analysis techniques to chemical sensor array data there is the possibility for artificial olfactory systems to provide the missing link between instrumental and sensory-based investigations. Some of these methods and an example of an environmental monitoring problem is provided by Sneath and Persaud in Chapter 15. Finally a promising new area of research in machine olfaction is presented – applying chemical sensor systems to mobile robotic systems. Ishida and Moriizumi consider the possibilities for mobile chemosensory systems. Two possible modes of operation are considered here: relatively straightforward chemical trail following and the far more complex problem of chemical source localization in turbulent odor plumes. Insect models are used as the inspiration for the approach – the ant for trail following behavior and the moth for chemotaxis within airborne odor plumes. Although their experiments are preliminary and work in this area is at an early stage, there are many exciting research challenges that will need to be considered in the future.

Part D Overview: Applications and Case Studies

This final section of the Handbook presents a variety of areas in which electronic nose technology has been applied. In each application, the tools and techniques of Parts A, B, and C are selectively employed to achieve specific performance goals. In the first chapter, Nagle, Gutierrez-Osuna, Kermani, and Schiffman examine environmental applications. Examples of water, land, and air monitoring experiments reported in the open literature are examined, followed by four case studies of work done by the authors. The first three demonstrate the ability of the AromaScan A32S electronic nose to classify odors from animal confinement facilities. In the first, the A32S was employed to classify the source of an odor emission as being from the lagoon, the confinement building exhaust fan, or a downwind ambient air. In the second, the A32S was used to determine the detection threshold concentration for acetic acid, a major individual constituent in swine slurry odor. In the third case study, the A32S was used to evaluate the performance of a biofilter of earth, wood Preface XXV chips, small twigs, and straw on the confinement building exhaust as an odor remediation measure. In the fourth case study, the NS State Electronic nose, a prototype unit with fifteen commercially available MOSs, demonstrated that an electronic nose can differentiate between five types of fungi that commonly lower indoor air quality in office buildings and industrial plants. These four case studies demonstrate that the electronic nose can perform well in various environmental monitoring applications. The next chapter by Persaud, Pisanelli, and Evans gives a summary of medical diagnostics and health-monitoring applications. Many diseases and intoxications are accompanied by characteristic odors, and their recognition can provide diagnostic clues, guide the laboratory evaluation, and affect the choice of immediate therapy. After reviewing the history of electronic nose uses in this area, two case studies are introduced. In the first, metabolic changes due to myopathies are detected by urine odor. The electronic nose was able to differentiate the normal population from that with myopathies. In the second case study, an electronic nose was employed to detect bacterial vaginosis. Success in this area led Osmetech to seek federal drugs administration (FDA) approval of one of their instruments for this application. Next, Deffenderfer, Feast, and Garneau provide a comprehensive overview of the electronic nose as an analytical tool for applications in natural products ranging from identifying solvents and the discrimination of spirits, to beverage and grain quality. Following this overview, they then illustrate two specific case studies. In the first, the Cyranose 320 is used to identify trees of different species for the pulp and paper industries in eastern Canada. In the second case study, the Cyranose 320 is employed to differentiate essential oil-bearing plants. Their results indicate that the electronic nose has great potential in these industries. Process monitoring is the subject of the fourth applications chapter. Haugen and Bachinger give an overview of the fundamentals of non-invasive on-line monitoring of biological processes, followed by two case studies. The electronic nose in their studies used a set of 10 MOSFETs sensors, up to 19 MOS sensors and 1 CO2-monitor based on infrared adsorption. The MOSFET sensors were produced in-house at Linko¨ping Uni- versity (Sweden) with different catalytic metal gates of Pd, Pt, and Ir. The MOS sensors used were commercially available sensors of Taguchi (TGS) or fuzzy inference systems (FIS) type. The electronic nose was used to monitor the aroma of cell cultures to gain insight into cell and process state changes as well as to identify process faults. In their first case study, ANN technology was used successfully to relate the gas sensor signal pattern to the cell biomass from Escherichia coli fermentations. The second case study focused on using an electronic nose to monitor the composition of the bioreactor headspace gas, and thus to track physiological state changes. Fast cell transition states were monitored in a semiquantitative approach appropriate for on-line and non-invasive control of industrial bioprocesses. The next applications chapter focuses on food and beverage quality assurance. In this chapter, DiNatale states that ‘the analysis of foodstuff is one of the most promising and also the most traveled road towards industrial applications for this technology.’ After a review of the literature in this field, a case study in fish freshness is detailed. The study uses a prototype instrument called the LibraNose from the Univer- sity of Rome ‘Tor Vergata’. The LibraNose is based on an array of QCM sensors whose XXVI Preface

chemical sensitivity is given by molecular films of metalloporphyrins and similar compounds. Spoilage in fish can be detected through the measure of the amount of amines, such as trimethylamine, in the headspace of storage containers. In the study, the LibraNose was able to track two important parameters indicating that the electronic nose is a good candidate for future use in food freshness applications. The next chapter focuses on automotive and aerospace electronic nose applications. Automotive applications include monitoring the exhaust for combustion efficiency, monitoring the engine compartment for leaking oil or other fluids, and monitoring the cabin air for passenger safety (offgassing of fabrics and materials, leaks of coolant from the air-conditioning system, and intake of air from the roadway and the engine compartment). Aerospace applications vary from the addition of an electronic nose to study the variations in atmosphere over days or seasons on other planets, to monitoring air quality in human habitats. The electronic nose developed at the Jet Propulsion Laboratory (JPL) was designed to detect a suite of compounds in the crew habitat of a spacecraft, an enclosed space where air is recycled and it is unlikely that unknown and unexpected vapors will be released. In this chapter, Ryan and Zhou present a case study in which the JPL ENose in a flight experiment on the Space Shuttle flight STS-95 (October–November 1998) was tested as a continuous air quality monitor to distinguish among, identify and quantify 10 common contaminants which may be present as a spill or leak in the recirculated breathing air of the space shuttle or space station. The JPL ENose has an array of 32 sensors, coated with 16 polymers/carbon composite sensing films developed at Caltech. In the study, the JPL ENose was trained to 12 compounds, the 10 compounds most likely to leak or spill and the other two being humidity change and vapor from a medical swab (2-propanol and water) used daily to confirm that the device was operating properly. For all cases except one (formaldehyde), the JPL ENose was able to detect the compound at or below the expected levels. Pamula investigates the use of the electronic nose for the detection of explosives. After reviewing the literature in this important application of electronic nose technology, the author reviews progress of the defense advanced research projects agency (DARPA) program to detect explosive mines by their chemical signatures. The chapter concludes with a case study of the Nomadics’ Fido (Fluorescence Impersonating Dog Olfaction) device. The device uses fluorescent polymer beads to detect trace amounts of TNT emanating from landmines. This technology shows great promise for future deployment in demining applications. In the final applications chapter, Rodriguez, Tan, and Gygax survey electronic nose applications in cosmetics and fragrances. Even though the use of electronic noses in the cosmetic and fragrance industry has been more limited than in many other areas, the published literature shows that, with optimization, many cosmetic and fragrance related analytical tasks can be solved. After the literature review, this chapter presents two case studies. In the first, eight fragrant samples with distinct odor characters but similar bulk composition were tested. Samples were analyzed by an HP 4440 Chemical Sensor and by capillary GC/FID. Both approaches were successful in classifying and differentiating the odorous samples. In the second study, an Alpha MOS Fox4000 electronic nose with 18 chemical sensors and a human panel were used to judge Preface XXVII the odor quality of a sunscreen product. The product samples had already passed analytical tests prior to undergoing sensory evaluation. Expert panel evaluations were made on 150 samples judged to fall in three categories: meets sensory standard, does not meet sensory standard but can be used as a ‘diluent’ when adjusting bulk quality, and does not meet sensory standard and is rejected. Over a six-month evaluation period, the Fox4000 demonstrated its ability to carry out sensory analyses by accurately classifying ‘good’ and ‘bad’ batches of the tested product. We believe that the material presented in the Handbook of Electronic Noses should not only help readers to find out more about this new and challenging subject, but also act as a useful reference in the future.

November 2002

T. C. Pearce, S. S. Schiffman, H.T. Nagle, J.W. Gardner XXIX

List of Contributors

Thomas Bachinger Todd Dickinson Independent Consultant Illumina, Inc. St. Larsgatan 8B/6.3 9390 Towne Centre Drive S-582 23 Linko¨ping Suite 200 Sweden San Diego CA 92121 Henry Baltes USA Institut fu¨r Quantenelektronik Dept. Physik (D-PHYS) Corrado Di Natale HPT H 7 Department of Electronic Engineering ETH Ho¨nggerberg University of Rome Tor Vergata CH-8093 Zu¨rich via di Tor Vergata 110 Switzerland 00133 Roma Italy Marina Cole Division of Electrical & Electronic Philip Evans Engineering Osmetech PLC School of Engineering Electra House Coventry CV4 7AL Electra Way UK Crewe CW1 6WZ J. Enrique Cometto-Muniz UK Chemosensory Perception Laboratory Dept. of Surgery (Otolaryngology) Julian W. Gardner University of California, San Diego Division of Electrical & Electronic Mail Code 0957 Engineering La Jolla, CA 92093-0957 School of Engineering USA Coventry CV4 7AL UK

Ricardo Gutierrez-Osuna Eduard Llobet 401 Russ Engineering Center Dept. of Electronic Engineering Computer Science and Engineering Universitat Rovira i Virgili Wright State University Autovia de Salou s/n Dayton, OH 45435 43006, Tarragona USA Catalonia Spain John-Erik Haugen MATFORSK Toysaka Moriizumi Osloveien 1 Faculty of Engineering N-1430 As Tokyo Insititute of Technology Norway Ookayama, Meguro–Ku Tokyo 152 Andreas Hierlemann Japan Physical Electronics Laboratory ETH Hoenggerberg, HPT-H 4.2, IQE H. Troy Nagle CH-8093 Zurich Department of Electrical and Computer Switzerland Engineering 432 Daniels Hall Evor L. Hines North Carolina State University Electrical & Electronic Engineering Box 79 11 Division Raleigh School of Engineering NC 27695-7911 University of Warwick USA Coventry CV4 7AL Takamichi Nakamoto UK Department of Physical Electronics Graduaute school of Science and Martin Holmberg Engineering S-SENCE and Applied Physics Tokyo Institute of Technology IFM 2-12-1, Ookayama, Meguro-ku Linko¨ping University Tokyo 152-8552 S-581 83 Linko¨ping Japan Sweden Hidehito Nanto Bahram G. Kermani Chair Division of Materials Science Illumina, Inc. Advanced Materials Science Research & 9390 Towne Centre Drive, Suite 200 Development Centre San Diego Kanazawa Institute of Technology CA 92121-3015 3-1 Yatsukaho USA Matto Ishikawa 924-0838 Japan List of Contributors XXXI

Vamsee K. Pamula Robert Sneath Electrical Engineering Silsoe Research Institute Duke University Wrest Park Durham, NC Silsoe USA Bedford MK45 4HS UK Tim C. Pearce Department of Engineering Joseph Stetter University of Leicester Department of Biological, Chemical & University Road Physical Sciences Leicester LE1 7RH Life Sciences Building, room 182 UK 3101 South Dearborn St. Chicago, IL 60616 Krishna C. Persaud USA DIAS UMIST Tsung Tan PO BOX 88 Alpha MOS Add. Sackville Street 20 avenue Didier Daurat Manchester M60 1QD 31400 Toulouse UK France

M. A. Ryan Emmanuel Vanneste Mail Stop 303–308 University of Antwerpen Jet Propulsion Laboratory Universiteitsplein 1 C2.28 4800 Oak Grove Drive B-2610 Wilrijk Pasadena CA 91109 Belgium USA David Walt Manuel A. Sa´nchez-Montane´s Department of Chemistry ETS de Informa´tica Tufts University Universidad Auto´noma de Madrid Pearson Lab Madrid 28049 Medford, MA 02155 Spain USA

Susan S. Schiffman Udo Weimar Department of Psychiatry Institute of Physical Chemistry 54212 Woodhall Building University of Tu¨bingen Box 32 59 Auf der Morgenstelle 8 Duke University Medical School D-72076 Tu¨bingen Durham, NC 27710 Germany USA XXXII List of Contributors

Fredrik Winquist Francois-Xavier Garneau Division of Applied Physics and De´partement des Sciences the Swedish Sensor Center Fondamentales Department of Physics and Universite´ du Quebec a Chicoutimi Measurement Technology 555 Boulevard de l’Universite´ Linko¨ping University Chicoutimi (Quebec) S-581 83 Linko¨ping G7H 2B1 Sweden Canada

Otto Wolfbeis Hanying Zhou University of Regensburg MS 303 –300 Institute of Analytical Jet Propulsion Laboratory Chemistry 4800 Oak Grove Drive DE-93040 Regensburg Pasadena CA 91109 Germany USA Part A Fundamentals of Odor Sensing

1 Introduction to Olfaction: Perception, Anatomy, Physiology, and Molecular Biology

Susan S. Schiffman, Tim C. Pearce

Abstract Odors are sensations that occur when compounds (called odorants) stimulate receptors located in the olfactory epithelium at the roof of the nasal cavity. Odorants are hydrophobic, volatile compounds with a molecular weight of less than 300 daltons. Humans can recognize and distinguish up to 10 000 different substances on the basis of their odor quality. Odorant receptors (ORs) in the nasal cavity detect and discriminate among these thousands of diverse chemical ligands. An individual odorant can bind to multiple receptor types, and structurally different odorants can bind to a single receptor. Specific patterns of activation generate signals that allow us to discriminate between the vast number of distinct smells. The physicochemical attributes of odorants that induce specific odor sensations are not well understood. The genes that code for ORs have been cloned, and results from cloning studies indicate that ORs are members of a superfamily of hundreds of different G-protein-coupled receptors that possess seven transmembrane domains. A complete knowledge of structure- odor relationships in olfaction awaits the three-dimensional analysis of this large family of ORs. Ultimately, simultaneous knowledge of the three-dimensional structure of ORs as well as odorants will allow us to develop a pattern recognition paradigm that can predict odor quality.

1.1 Introduction to Olfaction

All living organisms from simple bacteria to complex mammals including humans respond to chemicals in their environment. Chemical signals play a major role in feeding (e.g. nutrients), territorial recognition, sexual behavior, and detection of potentially harmful conditions such as fire, gas, and rancid food. In higher organisms, special chemical sensing systems (smell and taste) have developed that are distinguished anatomically by the location of their receptors in the nasal and oral cavities, respectively. This chapter will focus on the nature of odors (sensations) and odorants (odorous molecules) that are relevant to human smell perception. The physiology

and biochemistry of olfaction will be addressed as well. Taste will also be reviewed briefly. Odor sensations are induced by the interaction of odorants with specialized receptors in the olfactory epithelium in the top of the nasal cavity. In air-breathing animals including humans, odorants are volatile, hydrophobic compounds that have molecular weights of less than 300 daltons. The largest known odorant to date is labdane that has a molecular weight of 296 [1]. Chemical reactivity has little to do with odor potential since odorant molecules are uncharged. Odorants vary widely in structure and include many chemical classes including organic acids, alcohols, aldehydes, amides, amines, aromatics, esters, ethers, fixed gases, halogenated hydrocarbons, hydrocarbons, ketones, nitriles, other nitrogen-containing compounds, phenols, and sulfur-containing compounds. The signals induced by the interaction of odorants with olfactory receptors (ORs) in the olfactory epithelium are transmitted to the olfactory bulb and ultimately to the brain (see Fig. 1.1 and Section 1.4). The sense of smell is a remarkably sensitive system that responds to very low concentrations of chemicals. It is estimated that only 2 % of the volatile compounds avail-

Fig. 1.1 Cross-section of the skull, showing the location of the olfactory epithelium, olfactory sensory neurons, cribriform plate, olfactory bulb, and some central connections 1.1 Introduction to Olfaction 3 able in a single sniff will reach the olfactory receptors, and as few as 40 molecules of some mercaptans are sufficient to perceive an odor [2, 3]. The exquisite sensitivity of the smell system is illustrated by the human detection thresholds given in Table 1.1 (data from ref. [4]). It can be seen that these compounds can be detected at concentrations in the low parts-per-billion (ppb) and even low parts-per-trillion (ppt) range as in the case of thiophenol, thiocresol, and propyl mercaptan. Over the course of a day, odorants have enormous opportunities to reach olfactory receptors during the process of inhalation and exhalation. An average person breathes 15 times per minute (or 21 600 times per day) moving an average of 0.5 liters of air per breath (or 10 800 liters of air per day). Most odor sensations are produced by mixtures of hundreds of odorants rather than by a single compound. Individual components tend to harmonize or blend together in mixtures leading to perceptual fusion. Humans have limited capacity to identify single odorants in mixtures with three to four components being maximum [5]. This limitation in the ability to identify the individual components of mixtures appears to be an inherent property of olfaction since it is unrelated to the experience of the subjects or the type of odorants. Odor sensations are characterized by general descriptors, such as sulfurous, fruity, floral, and earthy, or by their source such as banana or orange. The range of distinctive odor sensations is enormous, and a skilled perfume chemist can recognize and distinguish 8000 to 10 000 different substances on the basis of their odor quality [6, 7] and even respond to chemicals never before encountered in our environment. The olfactory system detects and discriminates among this immense number of odorant types due to the broad repertoire of olfactory receptor proteins that are encoded by a large olfactory gene family [8–10] (see Section 1.5). Humans have several hundred distinct genes that encode olfactory receptor proteins and rodents have upwards of 500 to 1000 separate genes, that is, as much as 1% of the genome [9, 10]. This extremely broad range of receptor types permits the detection of odor sources comprised of unpredict- able mixtures of molecular species, and even allows detection of newly synthesized compounds with no known function.

Table 1.1 Odor thresholds of representative sulfur compounds [44].

Compound Characteristic odor Odor Threshold

Allyl mercaptan Garlic-coffee 0.05 ppb Amyl mercaptan Unpleasant strong 0.3 ppb Benzyl mercaptan Unpleasant strong 0.19 ppb Crotyl mercaptan Skunk-like 0.029 ppb Dimethyl sulfide Decayed vegetables 0.1 ppb Ethyl mercaptan Decayed cabbage 0.19 ppb Hydrogen sulfide Rotten eggs 1.1 ppb Methyl mercaptan Decayed cabbage 1.1 ppb Propyl mercaptan Unpleasant 0.075 ppb t-butyl mercaptan Skunk, unpleasant 0.08 ppb Thiocresol Skunk, rancid 0.062 ppb Thiophenol Putrid, garlic-like 0.062 ppb 4 1 Introduction to Olfaction: Perception, Anatomy, Physiology, and Molecular Biology

1.2 Odor Classification Schemes Based on Adjective Descriptors

Classification systems based on adjective descriptors have been used historically to organize the many thousands of different odor sensations into a limited number of groups. Table 1.2 shows some of the early schemes for classification of odor sensations. Modern olfactory specialists including perfumers who work with fragrances, however, find the small number of categories of early classification schemes to be inadequate for describing the odors that they encounter in their work. Over the last half century there has been a movement away from trying to classify odors in a few limited classes but rather to develop an extensive vocabulary that is relevant for use with the odor stimuli being examined. Hence, modern odor classification methods are based on an extensive array of adjective descriptors selected for their relevance for specific applications. Modern descriptive classification methods can be general (e.g. for the broad range of odors encountered in everyday life) or specific (e.g. relevant to particular applications in the food or fragrance industry). In the food industry, the odors of chemical compounds are often categorized by the identity of the edible material of which they are reminiscent. Sample odor classes for foods include caramel, honey, vanilla, citrus, and butter. Fragrance odors are often classified by floral and herbal groupings, such as jasmine, rose, balsam, or pine. Table 1.3 presents a series of 146 adjective descriptors developed by the American Society for Testing and Materials [20] for general classification of odors commonly encountered in everyday life. Table 1.4 gives a more specific list of descriptors used by the fragrance industry [21]. Other odor descriptors can be found in flavor and fragrance catalogs (Aldrich [22], for example) as well as on technical web sites (for example, ref. [23]).

Table 1.2 Descriptive categories proposed historically for smell sensations.

Number of Categories Category Classification Reference

6 Sweet, acid (sour), harsh, rich/fat, astringent, fetid 11 7 Aromatic, fragrant, ambrosial (musk-like), alliaceous 12 (garlic-like), hircine (goat-like), foul, nauseating 9 Aromatic, ethereal, fragrant, ambrosial, empyreumatic 13 (burnt), alliaceous, hircine, repulsive, nauseous 6 Flowery, fruity, spicy, resinous, burnt, putrid 14 8 Flowery, fruity, herbaceous (green), animal/ambrosial/ 15 human flesh aura, spicy/minty/camphoric, earthy/ fungoid, woody/balsamic/nut-like, Disagreeable: acrid/ phenolic/burnt/nauseating 7 Ethereal, floral, pepperminty, camphoraceous, musky, 16–18 pungent, putrid 9 Etherish, fragrant, sweet, spicy, oily, burnt, sulfurous, 19 rancid, metallic 1.2 Odor Classification Schemes Based on Adjective Descriptors 5

Table 1.3 ASTM descriptive categories used for general odor quality characterizations.a)

(01) Fragrant (50) Vanilla-like (99) Alcohol-like (02) Sweaty (51) Fecal (like manure) (100) Dill-like (03) Almond-like (52) Floral (101) Chemical (04) Burnt, smoky (53) Yeasty (102) Creosote (05) Herbal, green, cut grass (54) Cheesy (103) Green pepper (06) Etherish, anesthetic (55) Honey-like (104) Household gas (07) Sour, acid, vinegar (56) Anise (licorice) (105) Peanut butter (08) Like blood, raw meat (57) Turpentine (pine oil) (106) Violets (09) Dry, powdery (58) Fresh green vegetables (107) Tea-leaves-like (10) Like ammonia (59) Medicinal (108) Strawberry-like (11) Disinfectant, carbolic (60) Orange (fruit) (109) Stale (12) Aromatic (61) Buttery (110) Cork-like (13) Meaty (cooked) (62) Like burnt paper (111) Lavender (14) Sickening (63) Cologne (112) Cat-urine-like (15) Mushy, earthy, moldy (64) Caraway (113) Pineapple (fruit) (16) Sharp, pungent, acid (65) Bark-like, birch bark (114) Fresh tobacco smoke (17) Camphor-like (66) Rose-like (115) Nutty (18) Light (67) Celery (116) Fried fat (19) Heavy (68) Burnt candle (117) Wet paper-like (20) Cool, cooling (69) Mushroom-like (118) Coffee-like (21) Warm (70) Wet wool, wet dog (119) Peach (fruit) (22) Metallic (71) Chalky (120) Laurel leaves (23) Perfumery (72) Leather-like (121) Scorched milk (24) Malty (73) Pear (fruit) (122) Sewer odor (25) Cinnamon (74) Stale tobacco smoke (123) Sooty (26) Popcorn (75) Raw cucumber-like (124) Crushed weeds (27) Incense (76) Raw potato-like (125) Rubbery (new rubber) (28) Cantalope, honey dew melon (77) Mouse-like (126) Bakery, fresh bread (29) Tar-like (78) Black pepper-like (127) Oak wood, cognac-like (30) Eucalyptus (79) Bean-like (128) Grapefruit (31) Oily, fatty (80) Banana-like (129) Grape-juice-like (32) Like mothballs (81) Burnt rubber-like (130) Eggy (fresh eggs) (33) Like gasoline, solvent (82) Geranium leaves (131) Bitter (34) Cooked vegetables (83) Urine-like (132) Cadaverous, dead animal (35) Sweet (84) Beery (beer-like) (133) Maple (syrup) (36) Fishy (85) Cedar wood-like (134) Seasoning (for meat) (37) Spicy (86) Coconut-like (135) Apple (fruit) (38) Paint-like (87) Rope-like (136) Soupy (39) Rancid (88) Seminal, sperm-like (137) Grainy (as grain) (40) Minty, peppermint (89) Like cleaning fluid (138) Clove-like (41) Sulphidic (90) Cardboard-like (139) Raisins (42) Fruity (citrus) (91) Lemon (fruit) (140) Hay (43) Fruity (other) (92) Dirty linen-like (141) Kerosene (44) Putrid, foul, decayed (93) Kippery (smoked fish) (142) Nail polish remover (45) Woody, resinous (94) Caramel (143) Fermented fruit (46) Musk-like (95) Sauerkraut-like (144) Cherry (berry) (47) Soapy (96) Crushed grass (145) Varnish (48) Garlic, onion (97) Chocolate (146) Sour milk (49) Animal (98) Molasses a) Odor Quality characterizations. Each sample is where 0 indicates no odor and 5 indicates extre- rated on 146 adjectives using a five-point scale mely strong odor. 6 1 Introduction to Olfaction: Perception, Anatomy, Physiology, and Molecular Biology

Table 1.4 Odor descriptors in Allured’s Perfumer’s Compendium [21] used by perfumers and flavorists.

(1) Agrumen (50) Gardenia (99) Ozone (fresh air, marine) (2) Aldehydic (51) Geranium (100) Patchouli (3) Almond (52) Ginger (101) Peach (4) Amber (53) Grape (102) Pear (5) Ambergris (54) Grapefruit (103) Pepper (6) Animal (55) Grass (104) Peppermint (7) Anisic (56) Green (105) Petal (8) Apple Blossom (57) Hawthorne (106) Petitgrain (9) Apple Fruity (58) Hay (107) Pimento (10) Armoise (59) Herbal (108) Pine (11) Balsamic (60) Honey (109) Pineapple (12) Banana (61) Honeysuckle (110) Plum (13) Basil (62) Hyacinth (111) Powdery (14) Bay (63) Incense (112) Raspberry (15) Bergamot (64) Jasmin (113) Rooty (16) Camphoraceous (65) Juicy (114) Rose (17) Cardamom (66) Juniper (115) Sage (18) Carnation (67) Kiwi (116) Sandalwood (19) Cassie (68) Labdanum (117) Sappy-green wood (20) Cassis (69) Lactonic (118) Smokey (21) Castoreum (70) Lavender (119) Spicy (22) Cedar (71) Leafy (120) Strawberry (23) Celery (72) Leather (121) Styrax (24) Chamomile (73) Lemon (122) Sweet (25) Cherry (74) Lilac (123) Sweet pea (26) Chocolate (75) Lime (124) Tagette (27) Chrysanthemum (76) Mandarin (125) Tangerine (28) Cinnamon (77) Medicated (126) Tea (29) Citrus (78) Melon (127) Thyme (30) Civet (79) Metallic (128) Tobacco (31) Clary sage (80) Mimosa (129) Tolu (32) Clove (81) Minty (130) Tonka (33) Coconut (82) Moss (131) Tuberose (34) Cognac (83) Muguet (132) Vanilla (35) Coriander (84) Mushroom (133) Verbena (36) Costus (85) Musk (134) Vetivert (37) Cumin (86) Myrrh (135) Violet (38) Dry (87) Narcisse (136) Waxy (39) Earthy (88) Nasturtium (137) Wintergreen (40) Eucalyptus (89) Neroli (138) Woody (41) Fatty (90) Nutmeg (139) Ylang (42) Fecal-animal (91) Nutty (140) Zesty, peely (citrus) (43) Fig (92) Oily (44) Floral (93) Olibanum (45) Floral bouquet (94) Opoponax (46) Fougere (95) Orange flower (47) Freesia (96) Orange fruit (48) Fruity (97) Oriental (49) Galbanum (98) Orris 1.3 Odor Classification Based on Chemical Properties 7

Classification schemes for odor quality are beset, however, by a variety of limitations. First there are inherent interindividual differences in the emotional and hedonic properties of odors. Labels such as pleasant, delightful, disgusting, and revolting are common associations with odors, and these subjective evaluations can influence the choice of descriptors of odor quality. Emotional responses to odors probably derive from the fact that olfaction is a primal sense that is used in the animal kingdom to identify food, mates, predators, and warnings of danger. Second, there are individual differences in the actual perception of odor based on genetic differences [24–26]. Third, there are individual differences in the use of odor descriptors even among trained panelists. Fourth, the vocabulary of most languages lacks words that describe the full range of odor sensations. For this last reason, measures of similarity rather than adjective descriptors have been used to quantify odor quality by arranging odor sensations in multidimensional spaces (to be described in the next section).

1.3 Odor Classification Based on Chemical Properties

Although much progress has been made in our knowledge of olfactory physiology and biochemistry, the fundamental relationship between odor quality and molecular properties is still poorly understood. Even slight alterations in the chemical structure of an odorant can induce profound changes in odor quality. Current structure-activity models in olfaction are, for the most part, simply collections of disparate facts with no unifying theme; furthermore, they have inadequate predictive accuracy [27]. As a consequence, the basic logic necessary to develop a comprehensive odor classification scheme based on particular features of molecules remains elusive. There are several reasons for the lack of progress in classifying odors on the basis of chemical properties. First, it is not yet possible to model odorant-receptor interactions because the three-dimensional (3D) protein structures of the receptor sites are not known. Second, unlike structure-activity counterparts in pharmacology, there are vast numbers of agonist types (thousands of odorant structures) as well as thousands of different odor sensations. Third, identical molecules may activate different receptor types depending on the orientation of the molecule at the receptor. Beets [6] empha- sized that identical molecules arrive near receptor sites at different orientations and with different conformations. Thus, a given odorant would be expected to interact with a variety of receptor types, and odor quality must be encoded by a pattern of information from multiple receptors (rather than activation of a single receptor type). A fourth problem is that there are no standard methods for quantifying odor quality for use in structure-activity studies. 8 1 Introduction to Olfaction: Perception, Anatomy, Physiology, and Molecular Biology

1.3.1 History of Structure-activity Studies of Olfaction

In spite of these limitations, a variety of attempts have been made to relate odors to physicochemical parameters. Amoore [16] proposed that the shape and size of a molecule are the physicochemical parameters that determine odor quality, i.e. odorants fit into receptor sites in a ‘lock and key’ fashion with molecules of similar size and shape expected to have similar odor quality. In support of this theory, Amoore and Venstrom [28] reported significant correlations between odor quality of 107 odorants and a hand- calculated index of molecular size and shape for five classes of odors (ethereal, camphoraceous, musky, floral, and minty). Amoore [29] also reported a correlation of 0.90 between odor quality and a computer-generated molecular shape index when 25 substances were compared with benzaldehyde (almond odor). Wright [30] challenged Amoore’s results, indicating that it is inappropriate to represent a complex 3D molecular shape by an index consisting of a single number because many different 3D profiles could share the same molecular shape index. Wright [31] suggested that the mechanism for stimulation of olfactory receptors is low-energy molecular vibrations, and that molecules with similar vibrational frequency patterns should have similar odor quality. Wright and Robson [32] supported their hypothesis with the finding of similarity between the pattern of frequencies in the far infrared spectra for odorants with a bitter almond odor. Dravnieks and Laffort [33] suggested that four factors related to intermolecular interaction forces (an apolar factor, a proton receptor factor, an electron factor, and a proton donor factor) could predict both quantitative and qualitative odor discrimination in human beings. In spite of many attempts in addition to those just described, no general structure-activity model or theory has yet been proposed that accurately pre- dicts odor quality of molecules a priori from physicochemical parameters [1, 6, 27, 34, 35].

1.3.2 Odor Structures Associated with Specific Odor Classes Based on Qualitative Descriptors

Figures 1.2 to 1.6 provide examples of chemical structures for compounds classified by experienced odor specialists as having musk, ambergris, muguet, green, and bitter almond odors. Each figure gives the structure of representative chemicals within each specific odor quality. These figures illustrate that compounds with widely varying chemical structures can have similar odor qualities. Musk is an odor category that is used in fragrance with its original source being the glandular secretions of the male musk deer. Molecules with this odor quality are very diverse in structure as shown in Fig. 1.2; they include steroidal, linear, macrocyclic, nitro, as well as bi- and tricyclic compounds. Ambergris is an odor quality used in fragrance that originally derived from the sperm or cachalot whale. Muguet is a lily-of-valley odor. Green is the odor of natural green vegetable products. Bitter almond is an odor quality of an essential oil obtained by stem distillation of kernals from bitter almond (P. amygda- lus). The types of molecules within each odor quality can vary considerably in structure. 1.3 Odor Classification Based on Chemical Properties 9

Fig. 1.2 Compounds with musk odor: a) androst-16-en-3b-ol, b) ethyl citronellyl oxalate, c) cyclopentadecanolide, d) musk ketone, e) Traesolide, f) Galaxolide

Studies of enantiomers have also been used to gain insight into the relationship between physicochemical properties and odor quality. These studies reveal that enantiomers of chiral odorants may or may not show differences in odor quality [1, 6, 27, 34, 35]. There are significant differences in the two enantiomers of carvone with R-()-carvone having an odor of spearmint oil and S-(þ)-carvone having an odor of caraway oil (see Fig. 1.7). Significant differences in the odor quality of enantiomers of nootkatone have also been reported. However, enantiomers of 2-octanol and carbi- naol were not found to differ in odor quality. Overall, Figs. 1.2–1.7 demonstrate that structurally unrelated chemicals can yield similar odor qualities. Furthermore, differences in the odor quality of certain enantiomers indicate that very subtle differences in structure are capable of producing very different and distinct odors. A better understanding of the physicochemical para- 10 1 Introduction to Olfaction: Perception, Anatomy, Physiology, and Molecular Biology

Fig. 1.3 Compounds with ambergris odor: a) oxalactone, b) cyclohexyltetrahydrofuran, c) Karanal, d) timberol, e) cedramber

Fig. 1.4 Compounds with muguet odor: a) lilial, b) mugetanol

Fig. 1.5 Compounds with green odor: a) cis-3-he- xen-1-ol (leaf alcohol), b) Ligustral, c) nonadienal 1.3 Odor Classification Based on Chemical Properties 11

Fig. 1.6 Compounds with bitter almond odor: a) benzaldehyde, b) hydrogen cyanide

Fig. 1.7 Enantiomers of carvone. a) R-() carvone which has a spearmint-like odor, b) S-(þ) carvone which has a caraway-like odor

meters responsible for specific odor qualities requires more knowledge about the 3D structure of ORs.

1.3.3 Relationship of Physicochemical Parameters to Classifications of Odor Based on Similarity Measures

The methodology of multidimensional scaling has also been used to better understand the relationship between odor quality and physicochemical variables [36, 37]. Multi- dimensional scaling (MDS) procedures represent odor sensations in spatial maps. The input for multidimensional scaling procedures consists of quantitative measures of similarity between pairs of odors. For example, if two odors are judged by human subjects to have similar odor quality, they will be positioned near each other in the multidimensional quality space. Stimuli judged to be dissimilar are located distant from one another. Two examples of studies that relate physicochemical properties to odor quality as defined by multidimensional maps are given below. The mathema- 12 1 Introduction to Olfaction: Perception, Anatomy, Physiology, and Molecular Biology

tical procedures used to relate the physicochemical properties to the multidimensional maps are given in the Appendix.

1.3.3.1 Study 1: Broad Range of Unrelated Odorants A group of 50 odorants (5 of which were duplications) that ranged widely in quality and structure were arranged in a two-dimensional (2D) space by MDS on the basis of odor similarity [38]. The 2D space that is shown in Fig. 1.8 accounts for 91 % of the human similarity data. The odor stimuli were roughly positioned by MDS into two groups; the larger subset on the left is affectively more pleasant than the one on the right. Because the spatial arrangement could not be accounted for by a single physicochemical variable (such as chemical structure, molecular weight, number of double bonds, or dipole moment), a series of physicochemical variables were weighted in an attempt to regenerate the space. A mathematical technique generated weights for a series of physicochemical variables such that the distances and thus the spatial arrangements among the stimuli in Fig. 1.8 were regenerated. The mathematical procedure used to maximize the configurational similarity of the psychologically determined space in Fig. 1.8, with a space generated by weighted physicochemical parameters was based on a least-squares method (see Appendix). The physicochemical parameters that were weighted to reconstruct the 2D space in Fig. 1.8 as well as the means for these physicochemical variables are shown in Ta- ble 1.5. Functional groups were coded according to their number in a particular molecule; for example, benzaldehyde has one aldehyde group and the mean number of

Fig. 1.8 Two-dimensional solution for a broad range of odor stimuli. Compounds with similar odor qualities are located near each other in space. The more pleasant stimuli are located in the subset to the left, and the more unpleasant stimuli are in the subset on the right (modified after Schiffman [38]) 1.3 Odor Classification Based on Chemical Properties 13 aldehyde groups for all the molecules in the space in Fig. 1.8 is 0.10. Cyclic compounds were coded ‘1’ and noncyclic compounds ‘0.’ Raman spectra from 100 to 1000 cm1 were included because they contain much information that could be correlated with the pleasantness or unpleasantness of the molecules (i.e. that they fell to the left or right in the space). A large weight will expand the difference between these two stimuli more than a small weight, such that physicochemical variables with large weights are of greater importance in discrimination among the odor stimuli. Although this methodology was successful in relating strict quantitative measures of olfactory quality with quantitative physicochemical measures (i.e. 84 % of the variance was accounted for), the number of physicochemical variables needed to account for odor quality were too large to be of practical value. That is, the success in correlating physicochemical properties to odor quality did not improve the ability to predict or design of molecules with specific odor qualities.

Table 1.5 Weights that were applied to standard scores for physicochemical variables to regenerate the space in Fig 1.8. Functional groups were coded by their number in a molecule, thus, benzaldehyde was coded ‘1’. Cyclic compounds were coded ‘1’ while noncyclic compounds were coded ‘0.’

Physicochemical variable Weight Mean

Molecular weight 6.24 116.57 Number of double bonds 0.51 0.74 Phenol 2.33 0.13 Aldehyde 3.21 0.10 Ester 0.24 0.05 Alcohol 2.54 0.26 Carboxylic acid 5.50 0.13 Sulfur 3.44 0.08 Nitrogen 3.15 0.08 Benzene 0.14 0.33 Halogen 0.34 0.03 Ketone 0.19 0.03 Cyclic 4.56 0.31 Mean Raman intensity Below 175 cm1 0.01 0.51 176–250 cm1 3.57 2.36 251–325 cm1 0.75 1.65 326–400 cm1 3.81 1.56 401–475 cm1 1.65 2.10 476–550 cm1 3.63 1.54 551–625 cm1 0.69 2.07 626–700 cm1 1.16 1.07 701–775 cm1 0.07 2.36 776–850 cm1 3.04 4.36 851–925 cm1 0.24 3.44 926–1000 cm1 0.36 2.06 14 1 Introduction to Olfaction: Perception, Anatomy, Physiology, and Molecular Biology

1.3.3.2 Study 2: Pyrazines Multiple physicochemical parameters were also necessary to account for an odor space representing the similarity among related compounds (pyrazines). The pyrazines were ordered in the 3D space in Fig. 1.9 on the basis of similarity of odor quality [39]. Next, a set of descriptors (see Table 1.6) was generated by the automated data analysis and pattern recognition toolkit (ADAPT), a computer system for automated data analyses by pattern recognition techniques [40, 41]. The substructures used to generate the environment descriptors are given in Fig. 1.10. Canonical regression, a common statistical technique [42], was used to relate the descriptors in Table 1.6 to the 3D arrangement in Fig. 1.9. Canonical analysis extends multiple regression analysis from one criterion variable to a set of criterion variables. For simple multiple regression, the relationship of a set of predictors to a single criterion variable is analyzed. In the current application, canonical regression was used to determine the relationships between two sets of variables, that is, the stimulus coor-

Table 1.6 Descriptors generated by ADAPT [40] for analysis of pyrazines.

1 Number of atoms except hydrogen 2 Number of carbon atoms 3 Number of oxygen atoms 4 Number of bonds 5 Number of single bonds 6 Number of double bonds 7 Molecular weight 8 Path 1 molecular connectivity for all bonds in the structure 9 Path 1 molecular connectivity corrected for rings 10 Path 1 molecular connectivity calculated using the valences of heteroatoms and corrected for rings 11 Path 2 molecular connectivity 12 Path 3 molecular connectivity 13 Path 4 molecular connectivity 14 Molecular volume 15 Number of substructure 1 (see Fig. 1.11) 16 Environment-substructure 1 (calculates connectivity for substructure 1 and nearest neighbors) 17 Number of substructure 2 18 Environment-substructure 2 19 Number of substructure 3 20 Environment-substructure 3 21 Number of substructure 4 22 Environment-substructure 4 23 Number of substructure 5 24 Environment-substructure 5 25 Number of substructure 6 26 Environment-substructure 6 27 Number of substructure 7 28 Environment-substructure 7 29 Number of substructure 8 30 Environment-substructure 8 1.3 Odor Classification Based on Chemical Properties 15

Fig. 1.9a and 1.9b Two-dimensional cross-sections through the three-dimensional space for pyrazines [39]. Duplicate samples of the same stimulus are represented by two datapoints. 16 1 Introduction to Olfaction: Perception, Anatomy, Physiology, and Molecular Biology

Fig. 1.10 The substructures utilized by ADAPT for generating environment descriptors for analysis of pyrazines

dinates of the 3D MDS space and the physicochemical parameters in Table 1.6. The equations for canonical correlation used are given in the Appendix. Small subsets of the physicochemical parameters were used in the tests because in canonical correlation analysis, the number of stimuli should be greater than the number of dimensions and physicochemical parameters combined. The analysis revealed that a linear combination of two ADAPT parameters in Table 1.6 (number of oxygen atoms and chemical environment of substructure 7) in addition to a concentration variable accounted for 63 % of the arrangement of the pyrazine odor space in Fig. 1.9. This study, along with Study 1, again illustrates the difficulty in relating quantitative physicochemical parameters with odor quality.

1.3.4 Molecular Parameters and Odor Thresholds

In addition to odor quality, attempts have been made to determine the relationship between odor thresholds (or suprathreshold intensity) and molecular parameters. Variables that have been related to thresholds and intensities include molecular weight, cross-sectional area, adsorption constants at an oil-water interface, hydrophobicity, molar volume, pKa, saturated vapor pressure, polarizability, hydrogen bonding ability, air/water partition coefficients, log P (octanol-water partition coefficient), parameters derived from gas chromatograpy, Taft polar constants, and various steric parameters [34]). Like structure-activity studies of odor quality, there appear to be no rules that can be generalized for the entire range of odorous compounds.

1.3.5 Conclusions Regarding Physicochemical Parameters and Odor Quality

Although it is possible to develop techniques that weight a series of parameters to predict odor quality, this is of little practical use in understanding the physiological 1.4 Physiology and Anatomy of Olfaction 17 basis of odor quality. A more complete understanding of structure-activity relationships in olfaction will occur when the molecular structure of the odorant receptor (including the stereoelectronic arrangements of binding sites) is brought into the equation along with the structure of the odorant.

1.4 Physiology and Anatomy of Olfaction

1.4.1 Basic Anatomy

The functional organization of the olfactory system is similar to other sensory systems (e.g. vision) but, in this case, the sensory input is provided by molecules (i.e. odorants). Odorants are recognized by specific receptor proteins situated on the ciliary membranes of olfactory sensory neurons located in the olfactory epithelium at the top of the nasal cavity (see Fig. 1.1). The olfactory epithelium is comprised of three cell types as shown in Fig. 1.1: the bipolar olfactory sensory neurons (primary sensory neurons) with dendritic cilia projecting from their terminal ends in a thin mucus layer (10–100 lM thick); supporting or sustentacular cells (a type of glial cell) that terminate in microvilli; and basal cells (like stem cells) which make new olfactory receptor cells. The olfactory epithelium is a thin tissue in the nasal cavity that is easily distinguishable bilaterally in rats and dogs due to its yellowish color. In humans, however, the two small patches (about 2 square inches or 6.5 square centimeters) are more difficult to visualize because their pinkish hue blends with the respiratory epithelium that lines

Fig. 1.11 Olfactory epithelium showing three cell types: olfactory sensory neurons (also called receptor cells), supporting (or sustenacular cells), and basal cells 18 1 Introduction to Olfaction: Perception, Anatomy, Physiology, and Molecular Biology

the rest of the nasal cavity. Odorants can reach the olfactory receptors via orthonasal transport through the nares (e.g. when sniffing) or via retronasal transport from the oral cavity (e.g. when chewing food). In orthonasal transport, the turbinates (bones in the nose) create turbulent airflow patterns that direct volatile compounds to the olfactory receptor cells in the top of the nasal cavity. Inhaled odorants have been shown to reach the blood and brain after three hours of exposure [43], and as a consequence, olfactory receptors can also respond to blood-borne odorants [44]. There are vast numbers of olfactory sensory neurons with estimates between 106 and 108 in man [45, 46]. These olfactory neurons turn over continuously with an average time for replacement of approximately 30 days. This neurogenesis is active throughout the lifespan, and arises from basal cells deep in the epithelium [47]. Olfactory sensory neurons consist structurally of a soma (cell body), a peripheral dendritic knob with fine, long cilia that project into the watery mucus that protects the nasal cavity, and an unmyelinated axon that projects centrally from the soma and propagates action potentials to the olfactory bulb. Specific receptor subtypes are expressed in subsets of olfactory sensory neurons spatially distributed in distinct zones of the olfactory epithelium, and only one odorant receptor type is expressed on the vast majority of individual olfactory sensory neurons [9, 10]. Yet, single olfactory cells respond to a range of compounds with a variety of olfactory qualities because individual olfactory receptors have relatively broad molecular receptive ranges [48]. Axons of the bipolar olfactory sensory neurons fasiculate together and coarse through tiny holes in the cribriform plate of the ethmoid bone to the olfactory bulb where they make their first synapses with second-order neurons in intricate sphe- rical masses of neuropil called glomeruli (see Fig. 1.12). The axons of the bipolar cells constitute the fibers of the olfactory nerve. The neuropil of glomeruli consists of the axons of incoming olfactory sensory neurons and the dendrites of the mitral cells on which they synapse. Olfactory sensory neurons that express a specific odor receptor type converge upon a common glomerulus in the olfactory bulb [9, 10, 49]. In humans, axons from thousands of olfactory sensory neurons expressing a single odorant receptor type are thought to converge onto two or three glomeruli in the olfactory bulb, with each glomerulus receiving input from a single type of olfactory receptor. Local neuronal circuits in the bulb provide the first tier of central processing of odors with olfactory signals sharpened via lateral inhibition among glomerular modules [50]. As a result of this neural processing, mitral cells have narrower molecular receptive ranges than olfactory receptor neurons [48]. Because individual olfactory sensory neurons can respond to multiple odorants, it follows that the pattern across multiple glomeruli provides the basis for discrimination of olfactory quality. The distinct spatial patterns of glomerular activation by specific odorants can be visualized using optical imaging techniques [51, 52]. Olfactory information from the olfactory bulb is next transmitted by the olfactory tract to the anterior olfactory nucleus, the olfactory tubercle, the prepyriform cortex, and the amygdala, and ultimately to higher brain centers that process the olfactory signals. The prepyriform cortex and the amygdala are brain structures that are part of the limbic system, which processes emotions and memories in addition to olfactory signals. Olfactory information is ultimately transmitted to the hypothalamus (which 1.4 Physiology and Anatomy of Olfaction 19

Fig. 1.12 Cross-section of the olfactory bulb. A.C. indicates anterior commissure.

mediates food intake) and to the neocortex. Non-invasive imaging techniques such as electroencephalography, positron emission tomography, and functional magnetic resonance spectrometry have found that the degree of activation of the pyriform cortex, orbitofrontal areas, and parts of the parietal and temporal cortices is dependent on the odor quality and pleasantness of the stimuli (for example see refs. [53] and [54]). Age- related losses occur in the olfactory epithelium, olfactory bulb and nerves, hippocam- pus and amygdaloid complex, and hypothalamus, and these changes parallel perceptual losses in the olfactory system during the aging process. At elevated concentrations, odorants can also stimulate free nerve endings of the trigeminal nerve in the nose. Trigeminal stimulation by odorous chemicals induces sensations such as irritation, tickling, burning, stinging, scratching, prickling, and itching [55, 56]. Sensory information transmitted by the trigeminal nerve is not considered an ‘odor’ because the trigeminal nerve is not directly stimulated by electrical signals from olfactory receptor neurons; rather trigeminal stimulation involves a different sense called chemesthesis which is related to nociception (e.g. pain). 20 1 Introduction to Olfaction: Perception, Anatomy, Physiology, and Molecular Biology

1.4.2 Transduction and Adaptation of Olfactory Signals

Odorants first traverse the aqueous interphase that lines the surface of the olfactory epithelium in order to interact with the olfactory receptors in the ciliary membranes. This process is facilitated by soluble odorant binding proteins that ‘shuttle’ the hydrophobic odorants through the aqueous mucus layer towards specific odorant receptors. Odorant receptors are members of a superfamily of up to 1000 different G-protein- coupled receptors that possess seven transmembrane (7TM) domains. The location of odorant binding is thought to be a hydrophobic pocket in transmembrane regions 3, 4, and 5 of the seven-membrane-spanning receptor. Olfactory signaling is initiated by the interaction of an odorant molecule with a protein receptor on the ciliary surface. This ligand binding triggers multistep intracellular reaction cascades that open an ion channel in the cell membrane leading to depolarization by a few tens of millivolts [8, 57]. Figure 1.13 illustrates the binding of an odorant to a G-protein-coupled receptor in olfactory neurons that results in activation of cAMP. Odorant binding involves a sig-

naling pathway that includes a Gs-like protein (Golf) that activates a specific adenylyl cyclase leading to generation of cyclic AMP (cAMP). cAMP binds directly to a cyclic nucleotide-gated (CNG) ion channel in the cell membrane that increases the probability of positive ions flowing into the cell, leading to depolarization and action potentials.

Figure 1.13 a. and b. In most olfactory neurons, an odorant binds to an odorant receptor (OR) leading to an exchange of GTP (guanosine triphosphate) for GDP (guanosine diphosphate) on the heterotrimeric

G-protein (Golf ). c. The a subunit of Golf activates adenylyl cyclase leading to generation of cAMP. d. Cyclic AMP binds directly to a cyclic nucleotide-gated (CNG) ion channel in the cell membrane that increases the probability of positive ions flowing into the cell. This causes depolarization of the cell membrane and transmission of a signal along the axon to the bulb. CNG channels are nonselective and permeable to cations including Naþ and Ca2þ 1.5 Molecular Biology Of Olfaction 21

Another intracellular second messenger, inositol triphosphate may also mediate changes in the conductance in some olfactory neurons, leading to depolarization of olfactory cells in response to odorant-receptor binding. Olfactory signaling is terminated when receptors are phosphorylated via a negative feedback reaction catalyzed by two types of kinases [57]. The large family of G-protein-coupled 7TM receptors just described may not be the only odorant receptors. An alternate signaling pathway for olfactory transduction has recently been proposed by Gibson and Garbers [58]. They have found a large family of olfactory neuron-specific guanylyl cyclases that are membrane-bound and contain extracellular domains that may constitute a second family of odorant receptors. Activa- tion of guanylyl cyclase elevates cyclic GMP (cGMP) that converges on the same CNG channels as cAMP to generate action potentials. Repeated stimulation of olfactory receptor neurons leads to decrements in the neural responses, i.e. adaptation. Three forms of olfactory adaptation can take place in olfactory receptor neurons: two rapid forms and one persistent form. These three different adaptation phenomena are controlled, at least in part, by separate molecular mechanisms. These mechanisms involve Ca2þ entry through CNG channels, Ca2þ- dependent CNG channel modulation, Ca2þ/calmodulin kinase II-dependent attenuation of adenylyl cyclase, and the activity of the carbon monoxide/cyclic GMP second messenger system [59].

1.5 Molecular Biology Of Olfaction

The molecular era of olfaction began in 1991 with the discovery by Buck and Axel of a multigene family of G-protein-coupled ORs with a 7TM-spanning typology. Buck and Axel [8] obtained complementary DNA (cDNA) utilizing olfactory epithelial RNA from rat in conjunction with an amplification process called the polymerase chain reaction (PCR). (Complementary DNA is a copy of a messenger RNA). They found a PCR product (PCR 13) that contained multiple species of DNA that are representative of a multiple gene family that encodes transmembrane domain proteins that are restricted to the olfactory epithelium. Further work has shown that there is a conservation of certain amino acid motifs within OR gene sequences that distinguish ORs from other 7TM proteins [8, 60, 61]. There are also hypervariable regions within certain membrane regions of ORs (i.e. TMs 3, 4, and 5) that provide a diversity of ligand- binding pockets [61]. A single amino acid substitution in the hypervariable region can change ligand-binding specificity [62]. This diversity in ligand-binding domains is necessary to accommodate the enormous number of structurally diverse volatile chemicals that can activate the olfactory sensory neurons. Early estimates suggested that there are approximately 500 to 750 genes that encode ORs in humans with an estimated 1000 genes in mouse and rat [7, 9, 10, 63]. However, there appears to be a high frequency of pseudogenes (genes with defects that are incompatible with receptor function) in the human but not rat OR repertoire; between 38 % to 76 % of the human sequences do not encode full-length polypeptides 22 1 Introduction to Olfaction: Perception, Anatomy, Physiology, and Molecular Biology

[9, 10, 64, 65]. One recent report claims to have identified and physically cloned 347 human OR genes that they believe represent the complete repertoire of functional human ORs [66]. This reduction in the functional fraction of olfactory receptor genes in humans compared to rats implies that olfaction may have become less important during the course of evolution, perhaps due to relaxed selective constraints [65]. The decrease in viable odorant genes may also be the cause of specific anosmias in humans (inabilities to smell a particular odorant). OR genes are typically organized in clusters of ten or more and are distributed across numerous chromosomes [9, 10, 66–68]. The 347 olfactory genes identified by Zozulya et al. [66] were located on all human chromosomes, except for 2, 4, 18, 20, 21, and Y, with the majority (155 ORs) on chromosome 11 followed in frequency by chromosome 1 (42 ORs), 9 (26 ORs) and 6 (24 ORs). The average human OR is approximately 315 amino acids long. In general, only one OR gene is expressed in a single olfactory sensory neuron [69], and olfactory sensory neurons (OSNs) that express a single OR converge on the same glomerulus in the olfactory bulb. Thus for the adult mouse which has 1000 OR types and 1800 glomeruli [70], each OR may be associated with only two specific glomeruli. However, it should be noted that one recent study reported that there may be a subset of OSNs that expresses two distinct receptor types [71]. Knowledge of the physiological functioning of specific ORs is still in its infancy. That is, we know very little about the range of ligands that interact with each of the particular odorant receptors. This is due in part to the large number of odorant receptors and the enormous repertoire (many thousands) of odorous compounds. Experimental approaches in which ORs are functionally expressed in olfactory sensory neurons are necessary to determine the tuning of a specific OR. Functional expression of a specific ORs is achieved experimentally when a given receptor type is inserted into the plasma membrane, couples with the second messenger system, and produces a measurable response to an odorant ligand. Direct functional proof that the 7TM receptors cloned by Buck and Axel [8] were actually odor receptors was obtained by Zhao et al. [72] who inserted a gene discovered by Buck and Axel into the rat olfactory system, producing electrical activity in olfactory neurons to specific odorant chemicals. Zhao et al. functionally expressed an OR in olfactory sensory neurons of rat in vivo using an adenovirus-mediated gene transfer of a cloned OR, I7 (see ref. 8 for nomenclature). They inserted the I7 genes into an adenovirus vector linked to a gene for green fluorescent protein (GFP) that is used to mark genetically altered cells. (Disabled adenovirus vectors are used as a tool to transfer genes into mammalian cells. A viral gene can be replaced with another gene that encodes an OR protein.) Cells that carried the rat I7 gene also carried the GFP gene, and thus could be visualized because they glowed bright green when exposed to blue light. Extracellular transepithelial potential recordings from summed activity of many olfactory neurons (called an electro-olfactogram) in the infected epithelium were ele-

vated to heptaldehyde (C7), octyl aldehyde (C8), nonyl aldehyde (C9), and decyl aldehyde (C10) when compared with uninfected epithelium [72]. However, electro-olfactogram amplitudes were not elevated for hexaldehyde (C6) or undecylic aldehyde (C11). These findings suggested that the response profile of the 17 receptor is relatively specific for 1.6 Taste 23

C7 to C10 saturated aliphatic aldehydes at least within the limited set of 74 odorants that was tested. Heptaldehyde (C7), octyl aldehyde (C8), nonyl aldehyde (C9), and decyl aldehyde (C10) can be differentiated on the basis of odor quality so that a single receptor type does not code for a specific odor quality. Malnic et al. [69] used a combination of calcium imaging and single-cell RT-PCR (PCR with reverse transcription) to identify ORs for odorants with related structures but varied odors. Their results indicate that one OR recognizes multiple odorants, one odorant is recognized by multiple ORs, but that different odorants are recognized by different combinations of ORs. They concluded that the olfactory system uses a combinatorial receptor coding scheme to encode odor identities.

1.6 Taste

A brief overview of taste will also be given here because some of the sensors described in this book are ‘taste sensors.’

1.6.1 Taste Classification Schemes Based on Sensory Properties

Historically, the taste literature often suggests that there are only four (or possibly five) basic taste qualities (sweet, sour, salty, and bitter, and possibly ‘umami’ which is the taste of glutamate salts). All other tastes have been presumed to be combinations of these basic tastes. However, data are now accumulating that the range of taste sensations is much broader and includes qualities such as astringency, metallic, fatty, and calcium-like (e.g. chalky) [73–78].

1.6.2 Physiology and Anatomy of Taste

The receptor cells for taste are neuroepithelial cells that are clustered into buds and distributed on the dorsal surface of the tongue, tongue cheek margin, base of the tongue near ducts of the sublingual glands, the soft palate, pharynx, larynx, epiglot- tis, uvula, and first third of the esophagus (see Schiffman and Warwick [79] for an overview of anatomy). Taste sensations are induced by the interaction of chemicals (e.g. from food) with taste-buds during ingestion, chewing, and swallowing. Indivi- dual taste cells generally respond to more than one type of taste. Taste buds consist of approximately 50–100 cells that arranged in an onion-like structure (see Fig. 1.14). Individual cells in a taste-bud undergo continuous renewal every 10 to 10 1/2 days. Taste-buds on the tongue are positioned on specialized epithelial projections termed papillae. There are three different kinds of lingual papillae that contain taste-buds: fungiform papillae (which are shaped somewhat like mushrooms), foliate papillae 24 1 Introduction to Olfaction: Perception, Anatomy, Physiology, and Molecular Biology

Fig. 1.14 Taste bud

(which consist of linear depressions or vertical folds), and circumvallate papillae (which are surrounded by deep moats). The entire tongue is sensitive to all taste qualities but there are regional differences in sensitivity; for example, buds on fungiform papillae are more sensitive to sodium salts, foliate papillae to acids, and circumvallate to bitter compounds.

Fig. 1.15 Anatomy of taste showing the cranial nerves and nucleus of the solitary tract 1.6 Taste 25

Taste bud cells form direct neural connections called synapses with three cranial nerves: the facial nerve (VII), glossopharyngeal nerve (IX), and vagus nerve (X). These three cranial nerves relay signals from taste receptor cells to the rostral portion of the nucleus of the solitary tract located in the medulla in the brain stem (see Fig. 1.15). Signals are ultimately transmitted to the thalamus and gustatory cortex. Electrophy- siological studies indicate that individual taste neurons have broad, overlapping response patterns (i.e. they are broadly tuned) so that an individual fiber is non-specific but collectively the pattern of activity across multiple neurons is unique for a given stimulus [77, 80].

1.6.3 Transduction of Taste Signals

Taste stimuli interact with taste proteins (e.g. taste receptors) or with ion channels on the surface of taste cells, which induces electrical signals that ultimately reach the brain to register a taste. The salty taste of sodium salts is produced when Naþ ions traverse sodium channels in the membranes of taste cells [81]. The taste of potassium salts, like sodium salts, involves conductance of Kþ cations through taste cell membranes [82] Most studies indicate that the detection of bitter and sweet by tastants receptor cells involves G-protein-coupled receptors. Some but not all sweet compounds appear to bind to 7TM-spanning cell-surface receptors that activate the ade- nylate cyclase second messenger cascade [83]. At least two pathways are involved in bitter taste transduction: 1) the phosphatidylinositol second messenger cascade, and 2) the alpha-gustducin/phosphodiesterase pathway [86].

1.6.4 Molecular Biology of Taste

At current writing, two families of G-protein-coupled receptors designated as T1R (taste receptor family 1) and T2R (taste receptor family 2) are known to be selectively expressed in subsets of taste receptor cells. In 1999, Hoon et al. [84] cloned and characterized two novel 7TM domain proteins T1R1 and T1R2 (taste receptor family 1, members 1 and 2) that are expressed in topographically distinct subpopulations of taste receptor cells and taste buds. The receptors were localized to the taste pore. The following year, a novel family of receptors T2R were identified [85–87], and like T1R1 and T1R2, the T2R genes were selectively expressed in taste receptor cells. The T2R family consists of 40-80 proteins that appear to code specifically for bitter tastants. A candidate sweet receptor gene, called T1R3 (taste receptor family 1, member 3) was also been identified [88–91]. Further research has shown that receptors T1R2 and T1R3 combine by dimerization producing heterodimers (T1R2 þ 3) to recognize sweet-tasting molecules with different structures such as sucrose and saccharin [92]. Receptors T1R1 and T1R3 combine by dimerization producing heterodimers (T1R1 þ 3) that are broadly tuned to recognize L-amino acids [93]. A receptor 26 1 Introduction to Olfaction: Perception, Anatomy, Physiology, and Molecular Biology

that binds the amino acid L-glutamate called mGluR4 has also been cloned and characterized [94]. Cells expressing T1R2 þ 3 are found predominantly on the posterior tongue, which is innervated by the glossopharyngeal nerve [92]. Cells expressing T1R2 þ 3 are also located on the palate. Cells expressing T1R1 þ 3 are found predominantly on the front of the tongue, which is innervated by the chorda tympani nerve. Even though these different taste receptor types appear to be segregated anatomically, electrophysiological experiments indicate that individual taste cells and nerve fibers respond to stimuli having multiple taste qualities [77, 80]. Thus, further research is needed clarify the full range of taste receptors as well as elucidate how this taste information is coded by the nervous system.

1.7 Final Comment

The biological chemosensory systems just described share many analogies to examples of machine olfaction described in this book. For example, both the human olfactory system and machines have mechanisms for sample handing. In humans, a sniff is initiated when the diaphragm creates a relative negative pressure in the lungs and forces an air sample to be drawn through the nostrils and directed by the curved turbinates onto the sensory layer of the olfactory epithelium. In a typical electronic nose, a vacuum pump produces a negative relative pressure to draw the air sample through a tube (plastic or stainless steel) in a small chamber housing the electronic sensor array. Both biological systems and machines have far fewer sensors than the thousands of known odorants. Humans have several hundred different receptor types while the electronic nose typically has only 5 to 32 sensors. Both biological and machines send their responses into multilevel neural networks that identify and characterize the odor being produced by the odorant sample. Future advances in the molecular biology of smell and taste will undoubtedly impact the development of new electronic nose and electronic tongue devices.

References 1 G. Ohloff. Chemistry of odor stimuli. Ex- 5 A. Jinks, D. G. Laing. The analysis of odor perientia 1986, 42, 271–279. mixtures: evidence for a configurational 2 M. Stuiver. Biophysics of the sense of smell. process. Physiol. Behav. 2001, 72, 51–63. PhD Thesis. Groningen. 1958. 6 M. G. J. Beets. Structure-activity relationships 3 H. DeVries, M. Stuiver. The absolute sen- in human chemoreception. Applied Science sitivity of the human sense of smell. in Publishers Ltd., London, 1978. Sensory Communication (Ed.: W. A. Rosen- 7 R. Axel. The molecular logic of smell. blith), John Wiley and Sons, New York, 1961, Sci. Am. 1995, 273, 154–159. Chapter 9, pp. 159–167. 8 L. Buck, R. Axel. A novel multigene family 4 USEPA (United States Environmental Pro- may encode odorant receptors: a molecular tection Agency). Odor and corrosion control in basis for odor recognition. Cell 1991, 65, sanitary sewerage systems and treatment plants. 175–87. USEPA, Cincinnati, OH EPA/625/1-85/018, 1985. 1.7 Final Comment 27

9 P. Mombaerts. Molecular biology of odorant 26 J. E. Amoore, Specific anosmia and the receptors in vertebrates. Annu. Rev. Neurosci. concept of primary odors. Chem. Senses 1999, 22, 487–509. Flavor 1977, 2, 267–281. 10 P. Mombaerts. Seven-transmembrane 27 L. Turin, F. Yoshii. Structure-odor relations: proteins as odorant and chemosensory A modern perspective. in Handbook of Ol- receptors. Science 1999, 286, 707–711. faction and Gustation (Ed.: R. L. Doty), 11 Aristotle, De Anima, translated by Marcel Dekker, New York, 2002. W. S. Hett, Heinemann, London, Revised 28 J. E. Amoore, D. Venstrom. Correlations and reprinted, 1957. between sterochemical assessment and or- 12 C. Linnaeus (C. Von Linne´) Odores medica- ganoleptic analysis of odorous compounds. mentorum in Amoenitates Academicae, vol. 3, In Olfaction and Taste, vol. 2 (Ed.: T. Lars Salvius: Stockholm, 1752, p. 183. Hayashi), Pergamon, Oxford, 1967, p. 3–17. 13 H. Zwaardemaker. Die Physiologie des 29 J. E. Amoore. Stereochemical and vibrational Geruchs, translated from Dutch by A. J. theories of odour. Nature 1971, 233, von Langegg, W. Engelmann: Leipzig, 1895, 270–271. p. 324. 30 R. H. Wright. Stereochemical and vibratio- 14 H. Henning. Der Geruch I, Z. Psychol. nal theories of odour. Nature 1972, 239, 226. Physiol. Sinnesorgane 1915, 73, 161–257. 31 R. H. Wright. Odour and molecular vibra- 15 S. Klein. Primary odour element classification. Nature 1966, 209, 571–572. tion. Amer. Perfum. Essent. Oil Rev. 1947, 50, 32 R. H. Wright, A. Robson. Basis of odour 453–454. specificity: homologues of benzaldehyde and 16 J. E. Amoore. The stereochemical specifici- nitrobenzene. Nature 1969, 222, 290–292. ties of human olfactory receptors. Perfum. 33 A. Dravnieks, P. Laffort. Physico-chemical Essent. Oil Rec. 1952, 43, 321–323, and 330. basis of quantitative and qualitative odor 17 J. E. Amoore. The stereochemical theory of discrimination in humans. in Olfaction and olfaction. 1. Identification of seven primary Taste, vol. 4, (Ed.: D. Schneider), Wissen- odours. Proc. Sci. Sect. Toilet Goods Assoc. schaftliche Verlagsgesellschaft, Stuttgart, New York 1962, 37, S1–S12. 1972, pp. 142–148. 18 J. E. Amoore. The stereochemical theory of 34 M. Chastrette. Trends in structure-odor olfaction. 2. Elucidation of the stereoche- relationships. SAR QSAR Environ. Res. 1997, mical properties of the olfactory receptor 6, 215–254. sites. Proc. Sci. Sect. Toilet Goods Assoc. 35 K. J. Rossiter. Structure-Odor Relationships. New York 1962, 37, S13–S23. Chem. Rev. 1996, 96, 3201–3240. 19 H. G. Schutz. A matching standards method 36 S. S. Schiffman, M. L. Reynolds, F. W. for characterising odour qualities. Ann. Young. Introduction to multidimensional sca- N. Y. Acad. Sci. 1964, 116, 517–526. ling: Theory, methods, and applications. New 20 American Society for Testing and Materials York: Academic Press, New York, 1981. (ASTM), Atlas Of Odor Character Profiles. DS 37 S. S. Schiffman. Future design of flavour 61. ASTM, Philadelphia, 1992. molecules by computer. Chem Ind. 1983,3, 21 www.AlluredCompendium.com. Allured 39–42. Publishing Corp., 362 S. Schmale Road, 38 S. S. Schiffman. Physicochemical correlates Carol Stream, IL USA 60188-2787, 2001. of olfactory quality. Science 1974, 185, 22 Aldrich Flavors and Fragrances http:// 112–117. www.sigma-aldrich.com. Aldrich, 39 S. S. Schiffman, J. C. Leffingwell. Perception 1001 St. Paul Avenue, Milwaukee, of odors of simple pyrazines by young WI 53233 USA. and elderly subjects: A multidimensional 23 http://www.leffingwell.com. analysis. Pharmacol. Biochem. Behav. 1981, 24 D. Whissell-Buechy, J. E. Amoore. Odour- 14, 787–798. blindness to musk: simple recessive in- 40 A. J. Stuper, P. C. Jurs. ADAPT: A computer heritance. Nature 1973, 242, 271–273. system for automated data analysis using 25 C. J. Wysocki, G. K. Beauchamp. Ability pattern recognition techniques. J. Chem. Inf. to smell androstenone is genetically deter- Comp. Sci. 1976, 16, 99–105. mined. Proc. Natl. Acad. Sci. USA 1984, 81, 41 http://zeus.chem.psu.edu/ADAPT.html 4899–4902. 28 1 Introduction to Olfaction: Perception, Anatomy, Physiology, and Molecular Biology

42 J. Cohen, P. Cohen. Applied multiple regres- 56 J. E. Cometto-Muniz, W. S. Cain. Sensory sion/correlation analysis for the behavioral sci- reactions of nasal pungency and odor to ences. Lawrence Erlbaum, Hillsdale NJ, 1983. volatile organic compounds: the alkylben- 43 V. A. Benignus, K. E. Muller, J. A. Graham, zenes. AIHAJ 1994, 55, 811–817. C. N. Barton. Toluene levels in blood and 57 H. Breer. Odor recognition and second brain of rats as a function of toluene level in messenger signaling in olfactory receptor inspired air. Environ. Res. 1984, 33, 39–46. neurons. Semin. Cell Biol. 1994, 5, 25–32. 44 J. A. Maruniak, W. L. Silver, D. G. Moulton. 58 A. D. Gibson, D. L. Garbers. Guanylyl Olfactory receptors respond to blood-borne cyclases as a family of putative odorant odorants. Brain Res. 1983, 265, 312-316. receptors. Ann. Rev. Neurosci. 2000, 23, 45 D. G. Moulton. Dynamics of cell populations 417–439. in the olfactory epithelium. Ann. N Y Acad. 59 F. Zufall, T. Leinders-Zufall. The cellular Sci. 1974, 237, 52–61. and molecular basis of odor adaptation. 46 D. Lancet. Olfaction. The strong scent of Chem. Senses 2000, 25, 473–481. success. Nature 1991, 351, 275–276. 60 W. C. Probst, L. A. Snyder, D. I. Schuster, 47 E. E. Morrison. Morphology and plasticiy of J. Brosius, S. C. Sealfon. Sequence the vertebrate olfactory epithelium. in alignment of the G-protein coupled receptor Science of olfaction (Ed.: M. J. Serby, superfamily. DNA Cell Biol. 1992, 11, 1–20. K. L. Chobor) Springer-Verlag, New York, 61 Y. Pilpel, D. Lancet. The variable and 1992, p. 31–50. conserved interfaces of modeled olfactory 48 K. Mori, G. M. Shepherd. Emerging receptor proteins. Protein Sci. 1999,8, principles of molecular signal processing 969–977. by mitral/tufted cells in the olfactory bulb. 62 D. Krautwurst, K. W. Yau, R. R. Reed. Semin. Cell Biol. 1994, 5, 65–74. Identification of ligands for olfactory recep- 49 K. J. Ressler, S. L. Sullivan, L. B. Buck. tors by functional expression of a receptor A zonal organization of odorant receptor library. Cell 1998, 95, 917–926. gene expression in the olfactory epithelium. 63 L. B. Buck. Information coding in the ver- Cell 1993, 73, 597–609. tebrate olfactory system. Annu. Rev. Neurosci. 50 K. Mori , H. Nagao, Y. Yoshihara. The 1996, 19, 517–544. olfactory bulb: coding and processing of 64 S. Rouqier, C. Friedman, C. Delettre, odor molecule information. Science 1999, G. van den Engh, A. Blancher, 286, 711–715. B. Crouau-Roy, B. J. Trask, D. Giorgi. 51 B. D. Rubin, L. C. Katz. Optical imaging of A gene recently inactivated in human odorant representations in the mammalian defines a new olfactory receptor family olfactory bulb. Neuron 1999, 23, 499–511. in mammals. Hum. Mol. Gen. 1998,7, 52 L. Belluscio, L. C. Katz. Symmetry, stereo- 1337–1345. typy, and topography of odorant represen- 65 S. Rouquier, A. Blancher, D. Giorgi. tations in mouse olfactory bulbs. J. Neurosci. The olfactory receptor gene repertoire in 2001, 21, 2113–2122. primates and mouse: evidence for reduction 53 B. Kettenmann, C. Hummel, H. Stefan, of the functional fraction in primates. Proc. G. Kobal. Multichannel magnetoencephalo- Natl. Acad. Sci. USA 2000, 97, 2870–2874. graphical recordings: separation of cortical 66 S. Zozulya, F. Echeverri, T. Nguyen. The responses to different chemical stimulation human olfactory receptor repertoire. Genome in man. Electroencephalogr. Clin. Neurophy- Biol. 2001, 2(6), research0018.1-0018.12, siol. Suppl. 1996, 46, 271–274. http://genomebiology.com/2001/2/6/ 54 D. H. Zald, J. V. Pardo. Emotion, olfaction research/0018. and the human amygdala: Amygdala 67 N. Ben-Arie N, D. Lancet, C. Taylor, activation during aversive olfactory M. Khen, N. Walker, D. H. Ledbetter, stimulation. Proc. Natl. Acad. Sci. USA 1997, R. Carrozzo, K. Patel, D. Sheer, H. Lehrach. 94, 4119–4124. Olfactory receptor gene cluster on human 55 J. E. Cometto-Muniz, W. S. Cain. Sensory chromosome 17: Possible duplication of an irritation. Relation to indoor air pollution. ancestral receptor repertoire. Hum. Mol. Ann. N. Y. Acad. Sci. 1992, 641, 137–151. Genet. 1994, 3, 229–235. 1.7 Final Comment 29

68 B. J. Trask, C. Friedman, A. Martin-Gallardo, Partitioning, And Energy Expenditure, L. Rowen, C. Akinbami, J. Blankenship, Pennington Center Nutrition Series, Volume 2 C. Collins, D. Giorgi, S. Iadonato, (Ed.: G. A. Bray, D. H. Ryan), Louisiana F. Johnson, W. L. Kuo, H. Massa, State University Press, Baton Rouge, 1992, T. Morrish, S. Naylor, O. T. Nguyen, p. 293–312. S. Rouquier, T. Smith, D. J. Wong, J. 80 T. A. Gilbertson, J. D. Boughter Jr, H. Zhang, Youngblom, G. van den Engh. Members of D. V. Smith. Distribution of gustatory the olfactory receptor gene family are con- sensitivities in rat taste cells: whole-cell tained in large blocks of DNA duplicated responses to apical chemical stimulation. polymorphically near the ends of human J. Neurosci. 2001, 21, 4931–4941. chromosomes. Hum. Mol. Genet. 1998,7, 81 S. S. Schiffman, E. Lockhead, F. W. Maes. 13–26. Amiloride reduces the taste intensity of 69 B. Malnic, J. Hirono, T. Sato, L. B. Buck. Naþ and Liþ salts and sweeteners. Proc. Natl. Combinatorial receptor codes for odors. Cell Acad. Sci. 1983, 80, 6136–6140. 1999, 96, 713–723. 82 M. Kim, C. M. Mistretta. 4-aminopyridine 70 J. P. Royet, C. Souchier, F. Jourdan, reduces chorda tympani nerve taste H. Ploye. Morphometric study of the glo- responses to potassium and alkali salts merular population in the mouse olfactory in rat. Brain Res. 1993, 612, 96–103. bulb: numerical density and size distribution 83 B. Lindemann. Chemoreception: tasting along the rostrocaudal axis. J. Comp Neurol. the sweet and the bitter. Curr. Biol. 1996,6, 1988, 270, 559–568. 1234–1237. 71 N. E. Rawson, J. Eberwine, R. Dotson, 84 M. A. Hoon, E. Adler, J. Lindemeier, J. Jackson, P. Ulrich, D. Restrepo. Expres- J. F. Battey, N. J. Ryba, C. S. Zuker. Putative sion of mRNAs encoding for two different mammalian taste receptors: a class of taste- olfactory receptors in a subset of olfactory specific GPCRs with distinct topographic receptor neurons. J. Neurochem. 2000, 75, selectivity. Cell 1999, 96, 541–551. 185–195. 85 E. Adler, M. A. Hoon, K. L. Mueller, 72 H. Zhao, L. Ivic, J. M. Otaki , M. Hashimoto, J. Chandrashekar, N. J. Ryba, C. S. Zuker. K. Mikoshiba, S. Firestein. Functional A novel family of mammalian taste recep- expression of a mammalian odorant receptors. Cell 2000, 100, 693–702. tor. Science 1998, 279, 237–242. 86 H. Matsunami, J. P. Montmayeur, 73 R. D. Mattes. The taste of fat elevates post- L. B. Buck. A family of candidate taste prandial triacylglycerol. Physiol. Behav. 2001, receptors in human and mouse. Nature 74, 343–348. 2000, 404, 601–604. 74 S. S. Schiffman, M. S. Suggs, A. L. Sostman, 87 J. Chandrashekar, K. L. Mueller, S. A. Simon. Chorda tympani and lingual M. A. Hoon, E. Adler, L. Feng, W. Guo, nerve responses to astringent compounds in C. S. Zuker, N. J. Ryba. T2Rs function rodents. Physiol. Behav. 1992, 51, 55–63. as bitter taste receptors. Cell 2000, 100, 75 S. S. Schiffman, B. G. Graham, E. A. Sattely- 703–711. Miller, Z. S. Warwick. Orosensory percep- 88 J. P. Montmayeur, S. D. Liberles, tion of dietary fat. Curr. Dir. Psychol. Sci. H. Matsunami, L. B. Buck. A candidate 1998, 7, 137–143. taste receptor gene near a sweet taste locus. 76 M. G. Tordoff. Calcium: Taste, intake, and Nat. Neurosci. 2001, 4, 492–498. appetite. Physiol. Rev. 2001, 81, 1567–1597. 89 M. Max, Y. G. Shanker, L. Huang, M. Rong, 77 S. S. Schiffman. Taste quality and neural Z. Liu, F. Campagne, H. Weinstein, coding: implications from psychophysics S. Damak , R. F. Margolskee. Tas1r3, and neurophysiology. Physiol. Behav. 2000, encoding a new candidate taste receptor, 69, 147–159. is allelic to the sweet responsiveness locus 78 T. A. Gilbertson. Gustatory mechanisms Sac. Nat. Genet. 2001, 28, 58–63. for the detection of fat. Curr. Opin. Neurobiol. 90 E. Sainz, J. N. Korley, J. F. Battey, S. L. 1998, 8, 447–452. Sullivan. Identification of a novel member 79 S. S. Schiffman, Z. S. Warwick. The biology of the T1R family of putative taste receptors. of taste and food intake. in The Science Of J. Neurochem. 2001, 77, 896–903. Food Regulation: Food Intake, Taste, Nutrient 30 1 Introduction to Olfaction: Perception, Anatomy, Physiology, and Molecular Biology

91 X. Li, M. Inoue, D. R. Reed, T. Huque, 93 G. Nelson, J. Chandrashekar, M. A. Hoon„ R. B. Puchalski, M. G. Tordoff, Y. Ninomiya , L. Feng, G Zhao, N. J. Ryba, C. S. Zuker. G. K. Beauchamp, A. A Bachmanov. High- An amino-acid receptor. Nature online resolution genetic mapping of the saccharin publication, 24 February 2002 (DOI preference locus (Sac) and the putative sweet 10.1038/nature726). taste receptor (T1R1) gene (Gpr70) to mouse 94 N. Chaudhari, A. M. Landin, S. D. Roper. distal Chromosome 4. Mamm. Genome 2001, A metabotropic glutamate receptor variant 12, 13–16. functions as a taste receptor. Nat. Neurosci. 92 G. Nelson, M. A. Hoon, J. Chandrashekar, 2000, 3, 113–119. Y. Zhang, N. J. Ryba, C. S. Zuker. Mammalian sweet taste receptors. Cell 2001, 106, 381–390.

Appendix The basic matrix equations used by Schiffman [38] to maximize the configurational similarity of the psychologically determined space in Fig. 1.8 with a space generated by weighted physicochemical parameters are:

P ¼ Pˆ þ E Pˆ ¼ DQ P ¼ DQ þ E

where P is an (n)(n1)/2 column vector whose elements pij represent all the inter- stimulus distances between stimulus i and stimulus j and where n is the total number of stimuli; Pˆ is an (n)(n1)/2 column vector representing the proximity measures based on weighted physicochemical parameters; D is an [(n)(n2)/2] by k scalar di- 2 stance matrix whose elements dðijÞk are the squared differences between stimulus i and stimulus j for each physicochemical parameter k; Q is a k element column vector of weights for the k physicochemical parameters; and E is an (n)(n1)/2 column vector representing the error between the subjective proximities and the proximities based on physicochemical measures.

The error to be minimized is

0 @E E=@Qk ¼ 0

leading to the least squares solution

Q ¼ðD 0DÞ1D 0P

The equations for canonical correlation used to relate the descriptors in Table 1.6 to the three dimensional arrangement in Fig. 1.9 are given below.

^yki ¼ ako þ ak1ðyi1Þþ::::: þ akr ðyir Þ

x^ki ¼ bko þ bk1ðxi1Þþ::::: þ bkr ðxir Þ 1.7 Final Comment 31 where xil, xi2, etc. are the values of stimulus i on dimensions 1 and 2 of the MDS space just as in multiple regression equations, and yi1, yi2, etc. are ratings of stimulus i on several physicochemical parameters. The intercepts and weights are solved to maximize the correlation between ^ykiand x^ki. 33

2 Chemical Sensing in Humans and Machines

J. Enrique Cometto-Mun˜ iz

Abstract Chemosensory detection of airborne chemicals by humans is accomplished principally through olfaction and mucosal chemesthesis. Odors are perceived via stimulation of the olfactory nerve (CN I) whereas nasal chemesthetic sensations (i.e., prickling, irritation, stinging, burning, freshness, piquancy, etc), grouped under the term nasal pungency, are mediated by the trigeminal nerve (CN V). Airborne compounds elicit odor sensations at concentrations orders of magnitude below those producing pungency. The physicochemical basis for odor and pungency potency of chemicals, either singly or in mixtures, is far from understood. The sensitivity of the sense of smell often outperforms that of the most sophisticated chemico-analytical methods like gas chromatography and mass spectrometry. The combined used of these techniques with human odor detection (olfactometry), however, has proved an invaluable tool for understanding the chemosensory properties of complex mixtures such as foods, flavors, and fragrances.

2.1 Human Chemosensory Perception of Airborne Chemicals

Humans detect the presence of volatile organic compounds (VOCs) in their surroundings principally through their senses of olfaction and “chemesthesis” [1, 2], the latter is also known as the “common chemical sense” [3, 4]. Activation of the olfactory nerve (CN I) produces odor sensations; Chapter 1 describes the biological basis of this chemosensory pathway. Activation of chemoreceptors on the trigeminal nerve (CN V) innervating the face mucosae produces chemesthetic responses (see, for example, [5]). These responses evoked in the nose include stinging, piquancy, burning, freshness, tingling, irritation, prickling, and the like. All these nasal sensations can be grouped under the term nasal pungency [6]. Chemesthetic responses to airborne VOCs can also be produced in the ocular, oral, and upper airway mucosae, where they are referred to as eye, mouth, and throat irritation. In the back of the mouth and the throat, other nerves, such as the glossophar-

Handbook of Machine Olfaction: Electronic Nose Technology. Edited by T.C. Pearce, S.S. Schiffman, H.T. Nagle, J.W. Gardner Copyright ª 2003 WILEY-VCH Verlag GmbH Co. KGaA, Weinheim ISBN: 3-527-30358-8 34 2 Chemical Sensing in Humans and Machines

yngeal (CN IX) and vagus (CN X), are also stimulated by airborne VOCs and contribute to perceived irritation. In this chapter we will focus on the human sense of smell and nasal chemesthesis. We will review psychophysical studies performed on both sensory modalities addressing the possible basis for the odor and irritation potency of VOCs. We will also summarize various techniques that combine the power of the human nose with that of chemical-analytical instruments, such as gas chromatography and mass spectrometry, to quantify the chemosensory activity of volatile chemicals and to help understand better the characteristics of human chemosensory perception.

2.2 Nasal Chemosensory Detection

Odor thresholds represent an important biological characteristic of airborne chemicals. Nevertheless, compilation of such values [7–9] show an extreme variability for any particular substance, even after attempting to standardize the values reported in different sources [10]. This scatter severely limits the practical application of the information available. An important block in our understanding of smell and nasal chemesthesis is the lack of information regarding what particular characteristics of chemicals govern the potency (i.e., threshold and suprathreshold) and type (i.e., quality) of olfactory and trigeminal sensations that they evoke. The situation stands in sharp contrast with the senses of vision and hearing where we have a precise knowledge of the range of electromagnetic and vibrational energy, respectively, to which our eyes and ears are tuned. From a few known, well-defined parameters of light and sound it is relatively straightforward to predict its visual and auditory perceptual properties. It is not easy to predict the odor or chemesthetic perceptual properties from the structural and physicochemical properties of a compound. Attempts to correlate odor with structural and physicochemical properties of odorants have focused, typically, on one or a small number of odor qualities (see reviews in [11, 12]), probably because broader generalizations have failed to lead to a productive outcome. As has been pointed out [13], an important drawback of many structure- activity relationships in olfaction [14–19] is the difficulty in interpreting the chemical features that are shown to correlate with odor activity. Regarding chemesthesis in the upper airways, a pioneer review paper [20] described the possible chemical mechanisms of sensory irritation. This study focused principally on “reactive” chemicals, that is, substances producing chemesthetic responses principally via direct chemical reaction with mucosal tissues. A more recent review of the topic [21] also addressed the mechanism by which relatively nonreactive compounds could produce pungency. In fact, relatively nonreactive VOCs are the prime candidates for the production of adverse chemosensory symptoms in cases of indoor air pollution such as the sick-building syndrome (cf. [22]). Among the various factors accounting for the large variability of measured odor thresholds, apart from true biological variability, are: method of vapor-stimulus control and/or delivery, psychophysical methodology, criteria to arrive at a threshold re- 2.2 Nasal Chemosensory Detection 35 sponse, number of subjects, and number of trials per subject [23, 24]. In the case of nasal pungency thresholds, a crucial additional factor is the use of a procedure that avoids odor biases, because almost all chemicals have both odor and pungency and the odor could be quite strong at the concentrations needed to produce barely perceptible nasal pungency. Additionally, in order to build a chemosensory structure-activity relationship, a chemical stimulus continuum of some sort can be very helpful. In a wide-ranging research program started more than 10 years ago [25], odor and nasal pungency thresholds were measured using a standardized procedure aimed at minimizing many of the variability sources mentioned above, and to produce a data set with robust internal consistency. Some of the procedural features employed included: 1) Delivery of vapors monorhinally (i.e. one nostril at a time) via “static” olfactometry [26] from plastic squeeze bottles [27]. 2) Short-term exposures (1–2 seconds). 3) Rigorous measurement and follow-up of presented vapor-phase concentrations by gas chromatography. 4) Use of a two-alternative, forced-choice procedure against a blank to minimize bias; presentation of chemicals in an ascending concentration series to minimize sensory adaptation; and the use of, at least, duplicate bottles containing identical concentrations to alternate sniff sampling and avoid depletion of stimulus in the headspace. 5) Use of a constant and fixed criterion for threshold, such as five correct choices in a row, across subjects, repetitions, chemosensory modality (i.e. odor and nasal pungency), and different studies. 6) Selection of subjects with no sense of smell (called anosmics) to measure nasal pungency thresholds thus avoiding odor biases, and of subjects with normal sense of smell (normosmics) to measure odor thresholds. Normosmics were matched to the anosmics by age, gender and smoking status, all demographic variables known to influence chemosensory perception (see review in [28]). 7) Selection of stimuli from homologous chemical series, where physicochemical properties change systematically and where carbon chain length provides a convenient “unit of change” (i.e., a continuum) against which to relate the sensory results.

2.2.1 Thresholds for Odor and Nasal Pungency

The systematic studies of odor and nasal pungency thresholds along homologous chemical series included testing of n-aliphatic alcohols [25], n-acetate esters [29], sec- and tert-alcohols and acetate esters [30], ketones [30], alkylbenzenes [31], and aliphatic aldehydes and carboxylic acids [6]. Figure 2.1 summarizes the results obtained with all these series. The outcome clearly shows how both chemosensory thresholds decline as carbon chain length increases; this means that sensory potency (both olfactory and trigeminal) increases along each homologous series. The rate at which odor thresholds decline, at 36 2 Chemical Sensing in Humans and Machines

Fig. 2.1 Thresholds for odor (empty squares) and nasal pungency (filled squares) along homologous chemical series of alcohols, acetate esters, ketones, alkylbenzenes, aliphatic aldehydes, and carboxylic acids. Only primary and unbranched homologs are joined by a line. The segment of dotted lines on nasal pungency shows those homologs for which pungency begins to “cut-off” (see text). Bars (sometimes hidden by the symbol) indicate standard deviation

least for the first few members of each series, tends to be higher than that for nasal pungency thresholds. In various instances, such as for acetate esters, ketones, and alkylbenzenes, odor thresholds seem to reach a plateau. Nasal pungency thresholds, in contrast, reach a “cut-off” effect [6]: beginning with a certain homolog member, nasal pungency fails to be consistently evoked, and this effect deepens for all ensuing members. In other words, the ability of that particular homolog and of all following homologs to produce nasal pungency fades away. The reduced biological response due to the cut-off effect, seen at some point in a chemical series, is a well-known pharmacological phenomenon in the field of anesthesia [32, 33]. At least two mechanisms can account for such cut-offs [33]: a physical mechanism whereby the maximum vapor-phase concentration of the stimulus molecule, at a certain temperature and pressure, falls below the threshold; and a biological mechanism whereby the 2.2 Nasal Chemosensory Detection 37 stimulus molecule lacks a crucial property to trigger transduction. For example, the molecule could be too large to fit into the binding pocket of a receptive macromolecule or to interact effectively with a target site.

2.2.2 Stimulus-Response (Psychometric) Functions for Odor and Nasal Pungency

Studies that aimed at measuring thresholds for olfaction and nasal chemesthesis with a uniform methodology, particularly in the context of testing homologous chemical series, proved to be useful tools in understanding how physicochemical properties govern sensory potency. The use of a standard testing procedure was instrumental in developing robust quantitative structure-activity relationships (QSARs) (see below). Nevertheless, measurement of a punctate chemosensory threshold according to a fixed criterion of performance has limitations [34]. A more comprehensive knowledge of the chemosensory processes involved can be gained by measurement of complete stimulus-response (called psychometric) functions (e.g., [23, 24]). These functions span the range from chance detection to virtually perfect detection and, thus cross the boundaries between perithreshold and suprathreshold sensations. Given a certain set of testing conditions, psychometric functions depict a continuous track of how the detectability of the chemical(s) grow with increasing concentration, rendering a dynamic picture of the process. Figure 2.2 presents psychometric functions for the odor and nasal pungency evoked by 1-butanol, 2-heptanone, butyl acetate, and toluene. All functions in Figure 2.2 show

Fig. 2.2 Psychometric function for the odor (empty symbols) and nasal pungency (filled symbols) detection of butyl acetate (diamonds), 2-heptanone (circles), toluene (triangles), and 1-butanol (squares) 38 2 Chemical Sensing in Humans and Machines

an ogival shape with a close-to-linear section in the middle of the range. As expected from previous studies on thresholds (see review in [5]), odor detection occurred orders of magnitude below nasal pungency detection. The gap between olfactory and chemesthetic detection (at halfway between chance and perfect detection) ranged between 3.4 and 6.4 orders of magnitude. The two chemosensory modalities also differed in the slope along the linear portion of the function. Odor functions for these four chemicals have slopes between 0.35 and 0.5 [34, 35] whereas nasal pungency functions have slopes between 0.7 and 1.0, except toluene which showed an even steeper slope in the range 1.7–2.9 [34, 36].

2.3 Olfactory and Nasal Chemesthetic Detection of Mixtures of Chemicals

In typical, everyday experiences, olfactory and chemesthetic sensations arise from exposures to mixtures of substances. Rarely are they the result of exposure to a single chemical. In addition, the study of the chemosensory detection of mixtures compared to the detection of the individual components has the potential to uncover basic principles of functioning of the senses of smell and chemesthesis. Studies on the olfactory detection of mixtures of airborne chemicals have relied, for the most part, on measurement of thresholds according to a fixed criterion of performance, and have typically expressed the results in terms of the stimulus (that is, concentration of the chemical). Their outcome suggests partial and simple stimulus agonism [37–39] with some indications of synergistic stimulus agonism as number of components increases [39–42]. To illustrate the meaning of these terms, let us take the example of a 3-component mixture whose constituents are present at sensory-equivalent concentrations, i.e. at the same multiple or submultiple of their respective individual thresholds. The terms simple, synergistic, and partial agonism indicate, respectively, that the mixture achieves threshold when each component is present at one third, less than one third, and more than one third (but less than one time) its individual threshold concentration. The term independence indicates that the mixture achieves threshold only when at least one of the components is present at its individual threshold. The term antagonism indicates that the mixture achieves threshold only when the components are present at a concentration even higher than their respective individual thresholds. A recent study looking at the olfactory (and trigeminal) detectability of binary mixtures of 1-butanol and 2-heptanone via measurement of psychometric functions lent support, as a first approximation, to an outcome of simple agonism [35]. Not surprisingly, studies on the trigeminal detection of mixtures are much fewer than those on olfaction. A comprehensive study, measuring trigeminal thresholds for single chemicals and for mixtures of up to nine components, revealed a trend for the degree of agonism to increase with the number of components and with the lipophilicity of such components [39]. A couple of recent investigations used psychometric function measurements to look in detail at the trigeminal detectability of binary mixtures compared to the detectability of the single components [35, 36]. The general out- 2.4 Physicochemical Determinants of Odor and Nasal Pungency 39 come again supported simple agonism with the suggested possibility, open to further scrutiny, that for chemesthetic responses, simple agonism might weaken to partial agonism as the detectability of the mixtures approach perfect detection [36], that is, as the mixtures leave the perithreshold region and enter into the suprathreshold region. If such weakening of agonism is confirmed and extended to olfactory responses, it would correlate well with the finding of partial agonism (called hypoadditivity) very commonly reported for mixtures of odorants at the suprathreshold range (e.g. [43]) even when the analysis considers “addition” of concentration (mass) and not simply addition of sensation [44]. It has been suggested that, within each chemosensory modality, compounds with similar slopes in psychometric functions will tend to show simple agonism in mixtures, whereas compounds with different slopes will tend to show a lesser degree of agonism, e.g. partial agonism [36]. At this stage, psychometric functions for additional substances tested in binary and higher order mixtures need to be measured to confirm the trend.

2.4 Physicochemical Determinants of Odor and Nasal Pungency

As mentioned above, the senses of olfaction and chemesthesis allow the detection of airborne chemicals. To gain a better understanding of how these sensory channels function it is important to know what particular features of chemicals govern their potency as odorants and irritants, including threshold and suprathreshold intensities. Regarding olfaction, a large number of such features have been suggested, including Wiswesser notation formulas [14], structural parameters directly derived from the chemical formula [45] or derived from gas chromatographic measurements [17, 19], steric and electronic descriptors [46], molecular vibration [47–49], partition coefficients (specifically, water-air and octanol-water) [50] and an electron-topological method [51]. Some of these investigations focused on one or just a few odor qualities (e.g. musk) whereas others studied a broader spectrum. Regarding chemesthesis, there have also been a number of chemical features reported to correlate with sensory irritation. Among them, normal boiling point [52], adjusted boiling point [53], saturated vapor pressure [54], Ostwald solubility coefficient (i.e., log L where L ¼ concentration in solvent/concentration in gas phase) [55], and partition coefficients, specifically water-air and octanol-water [56]. Interest- ingly, all these descriptors are physicochemical parameters and do not involve the precise chemical structure of the irritant.

2.4.1 The Linear Solvation Model

Many of the QSARs cited above for olfaction and chemesthesis are difficult to interpret either chemically or mechanistically [13]. A recently developed model has the advan- 40 2 Chemical Sensing in Humans and Machines

tage of not only providing a strong statistical fit to human psychophysical data, but also of conveying chemically and mechanistically meaningful information on both the stimulus (i.e. odorant or irritant) and the biophase where sensory reception initially takes place, e.g. for olfaction, the membrane covering the cilia of the olfactory receptor neuron, and, for nasal chemesthesis, the membrane of the free nerve endings of the trigeminal nerve. This model is based on a general solvation equation developed by Abra- ham [57, 58]: X X H H H 16 log SP ¼ c þ r R2 þ s p2 þ a a2 þ b b2 þ l log L ð1Þ

where SP is the dependent variable that, in the present context, represents a sensory property defined as the reciprocal of the odor detection threshold (1/ODT) or the reciprocal of the nasal pungency threshold (1/NPT). The reciprocals are used simply because the larger the quantity, the more potent is the odorant or irritant. There

are five independent variables: excess molar refractionP (R2), dipolarity/polarizability H H (p2 ), overall or effectiveP hydrogen-bond acidity ( a2 ), overall or effective hydro- H gen-bond basicity ( b2 ), and gas-liquid partition coefficient on hexadecane at 298 K (L16). The L16 descriptor is a combination of two properties of the odorant or irritant: 1) a general measure of size, and 2) the ability of the odorant or irritant to interact with a biophase through dispersion forces. The term c and the coefficient for each of the independent variables (r, s, a, b, and l) are obtained by multiple linear regression analysis. However, these are not simply fitted coefficients. They have chemical and mechanistic meaning since they reflect the complementary properties that the biophase must show in order to be receptive to the odorant or irritant. In other words, the independent variables provide a physicochemical characterization of the stimulus whereas the corresponding coefficients provide a characterization of the receptive biophase likely to interact with that stimulus. The r-coefficient measures the tendency of the biophase to interact with the odorant or irritant via polarizability-type interactions, mostly via p- and n-electron pairs. The s-coefficient reflects the biophase dipolarity/polarizability, since a dipolar odorant or irritant will interact with a dipolar biophase, and a polarizable odorant or irritant will interact with a polarizable biophase. The a-coefficient represents the complementary property to the odorant or irritant hydrogen-bond acidity, and thus is a measure of the biophase hydrogen-bond basicity, since an acidic odorant or irritant will interact with a basic biophase. Similarly, the b- coefficient is a measure of the biophase hydrogen-bond acidity, since a basic odorant or irritant will interact with an acidic biophase. Finally, the l-coefficient is a measure of the biophase lipophilicity [13].

2.4.2 Application of the Solvation Equation to Odor and Nasal Pungency Thresholds

The standardized procedure employed to measure the odor and nasal pungency thresholds depicted in Fig. 2.1 provided a firm basis to develop QSARs based on the solvation model described above. Under this model, the odorant or irritant is 2.4 Physicochemical Determinants of Odor and Nasal Pungency 41 seen as a solute that travels through a series of solvent phases (air, nasal mucus, nasal tissue) until it exerts its (sensory) action upon a receptive biophase. Thus, the model only applies to what can be called “transport” processes. These are processes where the fundamental step is either the distribution of the stimulus between biophases or the rate of transfer of the stimulus from one biophase to another. The model does not apply to stimuli acting through exact conformational or geometrical states since these sorts of molecular changes would barely affect the above mentioned physicochemical descriptors but, when relevant, could affect potency dramatically. In addition, the model does not apply to “reactive” compounds, that is, substances that produce nasal pungency via direct chemical reaction with nasal tissue [21]. The solvation equation would underestimate the potency of such chemically reactive stimuli [59, 60]. The original equation for odor thresholds [13] was recently improved [61] with the addition of two additional terms:

1) A parabolic term (D D2) where D is the maximum length of the odorant molecule obtained by computer-assisted molecular modeling and geometry optimization. 2) An indicator variable, H, chosen as 2.0 for carboxylic acids and aldehydes, and zero for all other odorants. The need to introduce H arises because carboxylic acids and aldehydes are more potent than predicted [61]. The odor equation looks as follows: X H H log ð1=ODTÞ¼7:445 þ 0:304R2 þ 1:652 p2 þ 2:104 a2 X H 16 2 ß þ 1:500 b2 þ 0:822 log L þ 0:369D0:016D þ1:000H ð2Þ with n ¼ 60, r2 ¼ 0:84, SD ¼ 0:601, where n is the number of odorants included, r is the correlation coefficient, and SD is the standard deviation in the dependent variable. All symbols are as described for Eq. (1). The solvation equation model has performed even better for the description and prediction of nasal pungency thresholds [6, 62–65] than for odor thresholds. Its success indicates that transport processes indeed constitute a key step in the production of nasal pungency by nonreactive airborne chemicals. The latest version of the nasal pungency equation looks as follows: X X H H H log ð1=NPTÞ¼8:519 þ 2:154p2 þ 3:522 a2 þ 1:397 b2

þ0:860 log L16 ð3Þ with n ¼ 43, r2 ¼ 0:955, SD ¼ 0:27, where all letters and symbols are as defined above. In this case, the term r:R2 from the general Eq. (1) did not achieve significance and was omitted. It must be pointed out that Eq. (3) does not account for the observed cut-off effect on nasal pungency that we have mentioned in Section 2.2.1. Future research should aim at optimizing the range of applicability of Eq. (3) by including a “size” factor capable of accounting for such molecular cut-offs in chemesthesis. This line of work is likely to gather critical knowledge not only on the molecular boundaries of airborne pungent stimuli but also on those of the putative nasal chemesthetic receptor as well. 42 2 Chemical Sensing in Humans and Machines

2.5 Human Chemical Sensing: Olfactometry

All studies exploring how humans detect and perceive airborne chemicals need to devise a strategy to generate and deliver the stimuli at predetermined concentration levels. Generation, delivery, and control of chemical stimuli entail more complexity than the equivalent processes for physical stimuli such as lights and sounds. In addition, there are practically no well-established, accepted, and widely used commercial devices to perform such tasks. In many cases, a one-of-a-kind olfactometer is built with much effort and time for one or a few studies, only to be left in disuse, replaced, or substantially modified for other studies. As a rule, no steps are taken in order to understand how results obtained with the “old” device compare with those obtained with the “new” one. In this section we will discuss three broad olfactometric techniques that, with variations, have been and are still being used in the study of human chemosensory perception [26].

2.5.1 Static Olfactometry

In general, olfactometric techniques can be classified into “static” or “dynamic” depending on whether the vapor stimulus is drawn from an enclosed container where the liquid and vapor phases of the tested chemical are in equilibrium, or the vapor flows continually in a carrier-gas stream, typically odorless air or nitrogen. Important aspects in the static approach include the type of container, the way in which the vapor is drawn to the nose, and the type of connection between the headspace of the container and the nose of the subject. Containers in static olfactometry are typically glass or (almost) odorless plastic. As a rule, a series of dilutions of the substance(s) of interest are prepared in individual vessels using an odorless solvent. Choice of the solvent is not straightforward. Dis- tilled and deionized water could serve in some cases but some chemicals are unstable in water. For example, esters tend to hydrolyze producing the alcohol and the carboxylic acid. Also, most odorants have little or extremely low water solubility. Alter- native solvents are lipophilic substances where odorants are more stable and soluble. These include, for example, mineral oil and propylene glycol. Nevertheless, these are not always completely odorless and might present a low odor background. Many of the olfactory and nasal chemesthetic studies mentioned above resorted to the use of squeeze bottles [66] (Fig. 2.3(a)). Their caps have pop-up spouts that fit into one or the other nostril allowing monorhinic testing, which in addition to their easy availability and simplicity of use has made them useful not only in research but also in the clinic [27]. A recent study, using three members each of homologous alcohols, acetates, and ketones, has shown that a newly developed glass vessel system possesses advantages over the plastic squeeze bottles, producing nasal pungency thresholds systematically lower by an average factor of 4.6 compared to those obtained via squeeze bottles [67] (Fig. 2.3(b). 2.5 Human Chemical Sensing: Olfactometry 43

Fig. 2.3 (a) Olfactory testing of a subject via plastic squeeze bottles and caps with pop-up spouts. (b) Olfactory testing of a subject via glass vessels with Teflon nosepieces

Subjects can sample the vapors in the headspace of the container actively by sniffing or they can receive them passively, for example, when the experimenter activates a valve that sends a fixed volume of headspace into the participant’s nostrils. The second method [68] makes stimulation independent of the sniffing pattern of the subjects but it can cause progressive drying of the nasal mucosa, leading to irritation with repetitive stimulation, and can also lead to confusion between air pressure and odor sensations [69]. In addition, more recent studies have shown that natural sniffing maximizes olfactory performance in humans [70]. The type of connection between the vapor container and the subject’s nostrils determines the effective concentration reaching the nose. The squeeze bottles, with their pop-up spouts that fit inside one nostril, represented an improvement over other containers that are simply open and placed under the subject’s nose, but still left room for dilution of the stimulus from surrounding air. The above mentioned glass vessels include Teflon made nosepieces that fit snugly into both nostrils of the subject, maximizing the efficiency of the stimulus delivery [67]. It is important to stress that in all these techniques of static olfactometry, the actual stimulus is the vapor above the solution in the container. In principle, the vapor concentration is proportional to the liquid concentration, but such proportionality varies with odorants, solvents, and, sometimes, among concentrations of the same odorant- solvent pair. For these reasons, actual measurement of the vapor-phase concentration in each container, and periodic follow-ups to ensure stability, become the only safe- guard against incorrect assumptions. Unfortunately, all too often olfactory investigations do not include such vapor measurements. Gas chromatography provides a relatively simple way to measure and calibrate vapor concentrations for use in static olfactometry. 44 2 Chemical Sensing in Humans and Machines

2.5.2 Dynamic Olfactometry

Under the principles of dynamic olfactometry, the chemical stimulus flows continuously in a carrier gas stream of either purified air or nitrogen. The various concentrations of the substance(s) tested are typically achieved by mixing, in different proportions, the carrier-gas line with the odorant line. A number of elements including tubing, capillaries, flowmeters, mass flow controllers, valves, saturating and mixing vessels, deodorizing and air conditioning (i.e., temperature and humidity) devices constitute the necessary equipment for the generation and control of odorants. As in the case of static olfactometry, the interface between the exit of the stimulus and the nose is an important feature regarding possible unwanted dilution of the targeted concentration. The complete assembly is referred to as an “olfactometer”. In a very detailed analysis of various olfactometers and of the many principles guid- ing their design, Dravnieks [71] has described devices used in both animal and human studies. Dravnieks himself proposed a Binary Dilution Olfactometer [71] (Fig. 2.4). This instrument combines portability and stability of concentrations with ease of use and maintenance. Its simplicity arises from the fact that it uses saturated vapor as the source of undiluted stimulus and employs a series of capillaries of various widths and lengths to achieve 7 fixed increasing dilutions of the odorant, all presented at a final flow rate of 160 mL/min. One of the suggested applications of this device was to use it with 1-butanol so as to express the odor intensity of any source in terms of an odor-equivalent concentration of 1-butanol (in ppm by volume) [72]. The technique became an ASTM (American Society for Testing and Materials) recommended procedure [73]. Dravnieks also developed a Dynamic Forced-Choice Triangle Olfact- ometer for measurement of thresholds [74, 75]. Both types of olfactometers found an important application in the measurement of environmental odors [76]). Chemical stimulation of the olfactory and trigeminal chemosensory systems in the nose gives rise to both peripheral electrical potentials [77, 78] and central evoked potentials [79]. In order to study such electrophysiological events, an olfactometer was needed that 1) delivered the stimulus without altering the mechanical or thermal conditions at the stimulated mucosa, and 2) produced a sharp, square-wave type, stimulus onset and offset. Such an instrument was pioneered by Kobal and collaborators [77, 79]. Their instrument achieved these goals by embedding pulses of odorant or irritant in a constantly flowing air stream under controlled temperature (36.5 8C) and humidity (80 % RH). An interesting development in the area of dynamic olfactometry emerged from the description of an olfactometer that also served to measure respiratory parameters [80– 83] (Fig. 2.5). The instrument evolved through the years and in its latest version presents the odorants and irritants to subjects through a mask, with a good seal monitored by pressure, covering the nose and mouth in a room-temperature warmed (25 8C) and humidified (35 % RH) airflow. The concentration of the stimulus on the line feeding the mask is continuously monitored by a photo-ionization detector (PID). 2.5 Human Chemical Sensing: Olfactometry 45

Fig. 2.4 (a) Drawing illustrating some of the principles in the Dravnieks Binary Dilution Olfactometer (from [71]). (b) A perspective drawing of the same olfactometer (from [14])

2.5.3 Environmental Chambers

Use of whole-body environmental chambers to explore human chemosensory responses provides a close approximation to a “natural” setting. In static and dynamic olfactometry, two crucial issues that must be controlled include the actual concentration of the stimulus, typically measured via detectors used in gas chromatography such as PID or flame ionization detector (FID), and the nosepiece/nose interface. A loose interface between the nostrils and the stimulus exit, whether under a static approach (e.g., squeeze bottles) or a dynamic approach (e.g., Dravnieks olfactometer) probably 46 2 Chemical Sensing in Humans and Machines

Fig. 2.5 Schematic representation of the test station for measurement of sensory responses and breathing parameters. (from [83].)

results in a dilution of the effective stimulus. Different sniffing styles among subjects may also contribute to variability. Investigation of the “typical” characteristics of human sniffing provide some interesting values: the “average” human sniff draws a volume of 200 mL, lasts a minimum of 0.4 sec and reaches an instantaneous flow rate of 30 L/min [70, 84, 85]. These studies also concluded that:

1) individuals vary in their sniffing techniques but are consistent with their patterns across odorants and tasks; 2) most of the odor information is obtained in the first sniff; 3) natural sniffing provides optimum odor perception.

Many of the above mentioned characteristics can not be easily achieved by static or dynamic olfactometry, hence the appeal of using environmental chambers. Neverthe- less, in a room-sized exposure chamber, build-up, control, and rapid change of stimulus concentration become complex and problematic as the large surfaces in the chamber (including the bodies and clothing of subjects) adsorb and desorb airborne chemicals. For these reasons, even when whole-body exposures constitute the gold standard, the pace of testing under this approach is much slower. This highlights the importance of understanding the rules of interconvertibility among sensory results obtained with the different approaches and, given the enormous number of odorants and irritants, the need to develop robust quantitative structure-activity relationships for 2.6 Instruments for Chemical Sensing: Gas Chromatography-Olfactometry 47 the prediction of chemosensory responses. Examples of these relationships have been provided above in Section 2.4. Chamber studies have been particularly useful when applied to the understanding of issues of indoor air quality and associated topics. Since exposures in chambers can last for hours, they possess a clear advantage over other strategies when studying the effect of time of stimulation on chemosensory perception. Studies performed in environmental chambers have explored, among others, sensory responses to environmental tobacco smoke [76, 86–89], body odor [90], volatile organic compounds [91–96], fragrance materials in air fresheners [97], and formaldehyde (a substance off-gassing from certain home-insulation materials) [98].

2.6 Instruments for Chemical Sensing: Gas Chromatography-Olfactometry

Gas chromatography (GC), one of the most widely used techniques in analytical chemistry, was first formalized in 1952 [99]. As described in a couple of recent reviews [100, 101], researchers interested in odors and aromas quickly took advantage of this powerful separation technique to identify the principal odorants of specific products, for example foods, beverages, fragrances, and perfumes [102]. This particular application of GC is now known as gas chromatography-olfactometry (GC-O). In brief, the method uses GC to separate the individual components of a mixture (e.g., a food product) which are then presented, as they elute, to a subject (called a sniffer) for sensory detection and/or characterization. Soon researchers found that direct sniffing from the GC effluent, at the exit of a non- destructive detector, had important drawbacks. Among them, the hot and dry gases dried the nasal mucosa, producing serious discomfort, and the odorous background emitted by hot plastic components interfered with the detection of the eluting odorants [100]. This prompted the design of substantial improvements in the system that eventually led to present day GC-O. An important step along the way was the addition of humidified air to the GC effluent, resulting in the delivery of a pulsed wave of odorant, similar to that eluting from the GC, but minimizing nasal dehydration and discomfort for the human sniffer [103]. Further improvements included a venturi system that eliminated background odors, was able to handle narrow-bore GC columns with minimum loss of resolution, and provided additional comfort to the subject [104]. As the techniques of GC and GC-mass spectroscopy (GC-MS) became widespread and more sophisticated, it was possible to separate and chemically identify the dozens or hundreds of individual substances present in food, flavor, and fragrance products. It has been argued [105] that this knowledge created the illusion that the flavor chemistry of these products was well understood. These powerful analytical techniques by themselves have no way of identifying and weighting which compounds are contributing significantly to flavor, and to what extent, hence the crucial importance of the GC-O approach that incorporates human sensory detection. In fact, there are indications that the performance of GC-O rivals, and can even outperform the most sensitive and selective chemico-analytical methods like GC-MS, particularly towards the most 48 2 Chemical Sensing in Humans and Machines

powerful odorants [106]. In addition, GC-O requires comparatively little sample preparation and no need for synthesis of labeled compounds. The usefulness of GC-O continues to grow and expand as it combines with the latest analytical tools such as solid-phase microextraction (SPME) [107, 108]. Many GC-O systems are designed to split the GC effluent, sending part to a chemical detector and part to the sniffing port. Typically, humans are more sensitive than most chemical detectors so it is common that less than 10 % of the effluent is directed to the sniffing port while more than 90 % is directed to the detectors [109]. However, the use of non-destructive detectors, such as a thermal conductivity detector, TCD, allows all the GC effluent to be sent to the sniffing port, maximizing sensitivity [101]. We have discussed issues that deal with the optimization of GC effluents for chemosensory evaluation by human subjects. There are also issues that deal with the overall strategy for presenting the stimulus (typically a complex mixture of odorants and non-odorants) to the subjects and, very importantly, the procedure used to gather and quantify sensory information from the subjects [109]. The application field were many of these methods were developed and investigated relates to food and flavor research. At present, there are at least three techniques commonly used in the study of the sensory properties of the chemical components of foods and flavors by GC-O. These are “charm analysis”, aroma extract dilution analysis (AEDA), and “osme” (from the Greek word meaning smell). We will briefly describe each of these methods.

2.6.1 Charm Analysis

This dilution technique was introduced in the middle 1980s [105]. On each run, the subject is exposed to the GC effluents from one of a series of increasing dilutions of the particular stimulus investigated, typically a complex mixture of chemicals. The participant strikes a key from a computer keyboard each time an odor begins to be detected and, again, when the odor is no longer detectable. During this interval, the subject is also required to report, for example with another key stroke, the quality of the perceived odor. The procedure renders a record of the time on the GC run where the odor occurred, its duration, and its quality. As the authors point out, a crucial part of the method calls for the use of chromatographic standards (e.g., n-paraffins) to transform the retention times at which odors appear into retention indexes, thus as- sociating the sensory response with a reproducible chemical property. A run as described above is made for each of the successive serial dilutions until no odor is detected. The responses are summarized as the “charm” value c, that is a simple function of the dilution factor d and the number of coincident responses n. The term “coincident responses” refers to the number of times that an odor is detected across successive dilutions for a particular retention index. In this way, the relationship is expressed as: c ¼ dn1. A charm response chromatogram is defined as a plot of c vs. retention index. Figure 2.6 illustrates how the charm plot is obtained. Results obtained by charm analysis compare well with those obtained by using traditional psychophysical procedures such as line-length (a visual analogue scale) and finger-span [110]. 2.6 Instruments for Chemical Sensing: Gas Chromatography-Olfactometry 49

Fig. 2.6 Example of a “charm” response chromatogram produced from the relationship c ¼ dn1, where d is the dilution constant and n is the number of coincident responses at any given retention index (from [105])

Charm analysis has been applied to study, among other products, apples [111], grapes [112, 113], orange juices [114] and the off-flavors form plastic packaging of food products [115].

2.6.2 Aroma Extract Dilution Analysis (AEDA)

AEDA is another dilution technique [116]. As in charm analysis, an extract from the product of interest is diluted in series and each dilution is analyzed by GC-O. In AEDA, results are expressed as flavor dilution (FD) factors. This factor is simply the ratio of the concentration of the odorant in the initial extract to its concentration at the highest dilution at which an odor is detected by GC-O [117, 118]. AEDA chromatograms plot the flavor dilution factor vs. retention index. Graphs obtained by charm analysis and by AEDA of the same flavor product are very similar [101] only that charm analysis produces areas for each relevant retention index (see Fig. 2.6) whereas AEDA produces heights, that is, a single number on the y-axis (equal to the FD) for each relevant retention index. In this way, AEDA focuses on the highest dilution at which a compound is detected whereas charm analysis also takes into account the time for which the odor is perceived [110]. AEDA has also been applied to the study of numerous food products, including olive oil, butter, Swiss cheese, meat, bread, beer, green tea, dill herb, and off-flavors [118], and wines [119]. 50 2 Chemical Sensing in Humans and Machines

2.6.3 Osme Method

The word “osme” given to this method [120] derives from Greek and means smell, hence the terms used above: “anosmia”, lack of sense of smell, and “normosmia”, normal sense of smell. In contrast to the two techniques described in Sec- tions 2.6.1 and 2.6.2, osme measures perceived odor intensity and is not based on dilutions to odor detection thresholds. The subject uses a time-intensity tracking procedure to rate the intensity of each eluting odorant from the GC and, at the same time, provides verbal descriptions of the odor-active regions of the chromatogram [121]. Similar to charm analysis and AEDA, retention times for the odor peaks are converted into standardized retention indices to confirm the chemical identity of the odorants. In some cases, further confirmation is achieved by GC-MS [121]. Variations on the specific procedure of time-intensity odor tracking, for example a PC mouse moved on a 60 cm scale vs. a rheostat apparatus that measured finger span, were shown to make no significant difference to the odor peaks obtained [110]. Osme has been applied to the analysis of wines [121] and hop oils and beers [122].

Acknowledgments Preparation of this article was supported by research grant number R01 DC 02741 from the National Institute on Deafness and Other Communication Disorders, Na- tional Institutes of Health, and by the Center for Indoor Air Research.

References

1 B. G. Green, J. R. Mason, M. R. Kare. American Society for Testing and In Chemical Senses. Vol. 2: Irritation (Ed.: Materials; 1978. B. G. Green, J. R. Mason, M. R. Kare), 8 Odor thresholds for chemicals with established Marcel Dekker, Inc., New York, 1990, occupational health standards. American v–vii. Industrial Hygiene Association. 1989. 2 B. G. Green, H. T. Lawless. In Smell 9 L. J. van Gemert. Compilations of odour and Taste in Health and Disease (Ed.: threshold values in air and water. TNO T. V. Getchell, R. L. Doty, L. M. Bartoshuk, Nutrition and Food Research Institute. J. B. Snow Jr.), Raven Press, New York, 1999. 1991, 235–253. 10 M. Devos, F. Patte, J. Rouault, P. Laffort, 3 G. H. Parker. J. Acad. Nat. Sci. Phila. 1912, L. J. van Gemert, eds. Standardized Human 15, 221–234. Olfactory Thresholds. Oxford:IRL Press; 4 C. A. Keele. Arch. Int. Pharmacodyn. Ther. 1990. 1962, 139, 547–557. 11 K. J. Rossiter. Chem. Rev. 1996, 96, 5 J. E. Cometto-Mun˜ iz. In Indoor Air Quality 3201–3240. Handbook (Ed.: J. D. Spengler, J. Samet, 12 M. Chastrette. SAR QSAR Environ. Res. J. F. McCarthy), McGraw-Hill, New York, 1997, 6, 215–254. 2001, 20.1–20.21. 13 M. H. Abraham. In Indoor Air and Human 6 J. E. Cometto-Mun˜ iz, W. S. Cain, Health. 2nd Edition (Ed.: R. B. Gammage, M. H. Abraham. Exp. Brain Res. 1998, 118, B. A. Berven), CRC Lewis Publishers, Boca 180–188. Raton, 1996, 67–91. 7 F. A. Fazzalari, ed. Compilation of odor and 14 A. Dravnieks. In Flavor Quality: Objective taste threshold value data. Baltimore: Measurement (ACS Symposium Series, 2.6 Instruments for Chemical Sensing: Gas Chromatography-Olfactometry 51

No. 51) (Ed.: R. A. Scanlan), American 37 D. G. Guadagni, R. G. Buttery, Chemical Society, 1977, 11–28. S. Okano, H. K. Burr. Nature 1963, 200, 15 P. Laffort, F. Patte. J Chromatogr. 1987, 406, 1288–1289. 51–74. 38 M. Q. Patterson, J. C. Stevens, W. S. Cain, 16 P. A. Edwards, P. C. Jurs. Chem. Senses J. E. Cometto-Mun˜ iz. Chem. Senses 1993, 18, 1989, 14, 281–291. 723–734. 17 L. S. Anker, P. C. Jurs, P. A. Edwards. Anal. 39 J. E. Cometto-Mun˜ iz, W. S. Cain, Chem. 1990, 62, 2676–2684. H. K. Hudnell. Percept. Psychophys. 1997, 18 P. A. Edwards, L. S. Anker, P. C. Jurs. Chem. 59, 665–674. Senses 1991, 16, 447–465. 40 A. A. Rosen, J. B. Peter, F. M. Middleton. 19 L. M. Egolf, P. C. Jurs. Anal. Chem. 1993, 65, J. Water Pollut. Control Fed. 1962, 34, 7–14. 3119–3126. 41 R. A. Baker. J. Water Pollut. Control Fed. 20 Y. Alarie. CRC Crit Rev Toxicol 1973,2, 1963, 35, 728–741. 299–366. 42 M. Laska, R. Hudson. Chem. Senses 1991, 21 G. D. Nielsen. CRC Crit. Rev. Toxicol. 1991, 16, 651–662. 21, 183–208. 43 B. Berglund, M. J. Olsson. Percept. 22 J. E. Cometto-Mun˜ iz, W. S. Cain. Ann. Psychophys. 1993, 53, 475–482. N. Y. Acad. Sci. 1992, 641, 137–151. 44 W. S. Cain, F. T. Schiet, M. J. Olsson, 23 J. C. Stevens, W. S. Cain, R. J. Burke. Chem. R. A. de Wijk. Chem. Senses 1995, 20, Senses 1988, 13, 643–653. 625–637. 24 W. S. Cain, J. F. Gent. J. Exp. Psychol.: Hum. 45 K.-O. Schnabel, H.-D. Belitz, Percep. & Perform. 1991, 17, 382–391. C. von Ranson. Z Lebensm Unters Forsch 25 J. E. Cometto-Mun˜ iz, W. S. Cain. Physiol. 1988, 187, 215–223. Behav. 1990, 48, 719–725. 46 M. Chastrette, J. Y. d. Saint Laumer. Eur. 26 W. S. Cain, J. E. Cometto-Mun˜ iz, J. Med. Chem. 1991, 26, 829–833. R. A. de Wijk. In Science of Olfaction (Ed.: 47 G. M. Dyson. Chem. Ind. 1938, 57, M. J. Serby, K. L. Chobor), Springer-Verlag, 647–651. New York, 1992, 279–308. 48 R. H. Wright. J. Theor. Biol. 1977, 64, 27 W. S. Cain. Ear Nose Throat J. 1989, 68, 473–502. 316–328. 49 L. Turin. Chem. Senses 1996, 21, 773–791. 28 J. E. Cometto-Mun˜ iz, W. S. Cain. In Smell 50 K. M. Hau, D. W. Connell. Indoor Air 1998, and Taste in Health and Disease (Ed.: 8, 23–33. T. V. Getchell, R. L. Doty, L. M. Bartoshuk, 51 N. M. Shvets, A. S. Dimoglo. Nahrung 1998, J. B. Snow Jr.), Raven Press, New York, 42, 364–370. 1991, 765–785. 52 J. Muller, G. Greff. Food Chem. Toxicol. 29 J. E. Cometto-Mun˜ iz, W. S. Cain. Phar- 1984, 22, 661–664. macol. Biochem. Behav. 1991, 39, 983–989. 53 D. W. Roberts. Chem. Biol. Interactions 30 J. E. Cometto-Mun˜ iz, W. S. Cain. Arch. 1986, 57, 325–345. Environ. Health 1993, 48, 309–314. 54 G. D. Nielsen, E. S. Thomsen, Y. Alarie. 31 J. E. Cometto-Mun˜ iz, W. S. Cain. Am. Ind. Acta Pharmacol. Nord. 1990, 1, 31–44. Hyg. Assoc. J. 1994, 55, 811–817. 55 G. D. Nielsen, L. F. Hansen, Y. Alarie. 32 N. P. Franks, W. R. Lieb. Nature 1985, In Chemical, microbiological, health and 316, 349–351. comfort aspects of indoor air quality – 33 N. P. Franks, W. R. Lieb. Environ. Health State of the art in SBS (Ed.: H. Kno¨ppel, Perspect. 1990, 87, 199–205. P. Wolkoff), Kluwer Academic Publishers, 34 J. E. Cometto-Mun˜ iz, W. S. Cain, Dordrecht, 1992, 99–114. M. H. Abraham, J. M. R. Gola. J. Appl. 56 K. M. Hau, D. W. Connell, B. J. Richardson. Toxicol. 2002, 22, 25–30. Toxicol. Sci. 1999, 47, 93–98. 35 J. E. Cometto-Mun˜ iz, W. S. Cain, 57 M. H. Abraham. Pure Appl. Chem. 1993, 65, M. H. Abraham, J. M. R. Gola. Physiol. 2503–2512. Behav. 1999, 67, 269–276. 58 M. H. Abraham. Chem. Soc. Rev. 1993, 22, 36 J. E. Cometto-Mun˜ iz, W. S. Cain, 73–83. M. H. Abraham, J. M. R. Gola. Toxicol. Sci. 59 M. H. Abraham, G. S. Whiting, Y. Alarie, 2001, 63, 233–244. J. J. Morris, P. J. Taylor, R. M. Doherty, 52 2 Chemical Sensing in Humans and Machines

R. W. Taft, G. D. Nielsen. Quant. Struct.- 80 J. C. Walker, D. B. Kurtz, F. M. Shore, Act. Relat. 1990, 9, 6–10. M. W. Ogden, J. H. I. Reynolds. Chem. 60 M. H. Abraham, G. D. Nielsen, Y. Alarie. Senses 1990, 15, 165–177. J. Pharm. Sci. 1994, 83, 680–688. 81 D. Warren, J. C. Walker, A. F. Drake, 61 M. H. Abraham, J. M. R. Gola, J. E. R. Lutz. Physiol. Behav. 1992, 51, 425–430. Cometto-Mun˜ iz, W. S. Cain. Chem. Senses 82 D. Warren, J. C. Walker, A. F. Drake, 2002, 27, 95–104. R. Lutz. Laryngoscope 1994, 104, 623–626. 62 M. H. Abraham, J. Andonian-Haftvan, 83 M. Kendal-Reed, J. C. Walker, J. E. Cometto-Mun˜ iz, W. S. Cain. Fundam. W. T. Morgan, M. LaMacchio, R. W. Lutz. Appl. Toxicol. 1996, 31, 71–76. Chem. Senses 1998, 23, 71–82. 63 M. H. Abraham, R. Kumarsingh, J. E. 84 D. G. Laing. Perception 1982, 11, 221–230. Cometto-Mun˜ iz, W. S. Cain. Arch. Toxicol. 85 D. G. Laing. Physiol. Behav. 1985, 34, 1998, 72, 227–232. 569–574. 64 M. H. Abraham, R. Kumarsingh, J. E. 86 W. S. Cain, B. P. Leaderer, R. Isseroff, Cometto-Mun˜ iz, W. S. Cain, M. Rose´s, L. G. Berglund, R. J. Huey, E. D. Lipsitt, E. Bosch, M. L. D´ıaz. J. Chem. Soc. Perkin D. Perlman. Atmos. Environ. 1983, 17, Trans. 2 1998, 2405–2411. 1183–1197. 65 J. E. Cometto-Mun˜ iz, W. S. Cain, 87 W. S. Cain, T. Tosun, L.-C. See, B. Leade- M. H. Abraham, R. Kumarsingh. Phar- rer. Atmos. Environ. 1987, 21, 347–353. macol. Biochem. Behav. 1998, 60, 765–770. 88 G. H. Clausen, P. O. Fanger, W. S. Cain, 66 J. E. Amoore, B. G. Ollman. Rhinology 1983, B. P. Leaderer. In Indoor Air. Volume 3. 21, 49–54. Sensory and Hyperreactivity Reactions to Sick 67 J. E. Cometto-Mun˜ iz, W. S. Cain, Buildings (Ed.: B. Berglund, T. Lindvall, T. Hiraishi, M. H. Abraham, J. M. R. Jola. J. Sundell), Swedish Council for Building Chem. Senses 2000, 25, 285–291. Research, Stockholm, 1984, 437–441. 68 C. A. Elsberg, I. Levy. Bull. Neurol. Inst. 89 J. C. Walker, P. R. Nelson, W. S. Cain, N. Y. 1935, 4, 5–19. M. J. Utell, M. B. Joyce, W. T. Morgan, 69 B. M. Wenzel. Psychol. Bull. 1948, 45, T. J. Steichen, W. S. Pritchard, 231–247. M. W. Stancill. Indoor Air 1997,7, 70 D. G. Laing. Perception 1983, 12, 99–117. 173–188. 71 A. Dravnieks. In Methods in Olfactory 90 G. H. Clausen, P. O. Fanger, W. S. Cain, Research (Ed.: D. G. Moulton, A. Turk, B. P. Leaderer. Environ. Int. 1986, 12, J. W. J. Johnston), Academic Press, 201–205. New York, 1975, 1–58. 91 H. K. Hudnell, D. A. Otto, D. E. House, 72 H. R. Moskowitz, A. Dravnieks, W. S. Cain, L. Mølhave. Arch. Environ. Health 1992, 47, A. Turk. Chem. Senses Flavor 1974,1, 31–38. 235–237. 92 S. K. Kjærgaard, L. Mølhave, O. F. Peder- 73 ASTM, E 544. Recommended practice for odor sen. Atmos. Environ. 1991, 25A, suprathreshold intensity referencing, Am. Soc. 1417–1426. Test. Materials, Philadelphia: 1975. 93 L. Mølhave, B. Bach, O. F. Pedersen. 74 A. Dravnieks, W. H. Prokop. J. Air Pollut. Environ. Int. 1986, 12, 167–175. Control Assoc. 1975, 25, 28–35. 94 L. Mølhave, J. Grønkjær Jensen, S. Larsen. 75 A. Dravnieks, W. H. Prokop, W. R. Atmos. Environ. 1991, 25A, 1283–1293. Boehme. J. Air Pollut. Control Assoc. 1978, 95 L. Mølhave. Ann. N. Y. Acad. Sci. 1992, 641, 28, 1124–1130. 46–55. 76 W. S. Cain, B. P. Leaderer. Environ. Int. 96 D. Otto, L. Mølhave, G. Rose, 1982, 8, 505–514. H. K. Hudnell, D. House. Neurotoxicol. 77 G. Kobal. Pain 1985, 22, 151–163. Teratol. 1990, 12, 649-652. 78 T. Hummel, M. Knecht, G. Kobal. Brain 97 F. T. Schiet, W. S. Cain. Perception 1990, 19, Res. 1996, 717, 160–164. 123–132. 79 G. Kobal, C. Hummel. Electroenceph. clin. 98 W. S. Cain, L. C. See, T. Tosun. In IAQ’86. Neurophysiol. 1988, 71, 241–250. Managing Indoor Air for Health and Energy Conservation (Ed.: American Society of Heating, Refrigerating and Air-Conditio- 2.6 Instruments for Chemical Sensing: Gas Chromatography-Olfactometry 53

ning Engineers, Inc., Atlanta, Georgia, 112 P. A. Braell, T. E. Acree, R. M. Butts, USA, 1986, 126–137. P. G. Zhou in Biogeneration of Aromas (Ed.: 99 A. T. James, A. J. P. Martin. Biochem. T. H. Parliment, R. Croteau), American J. 1952, 50, 679–690. Chemical Society, Washington, DC, 1986, 100 T. E. Acree. Anal. Chem. News & Features 75–84. 1997, 170A–175A. 113 T. E. Acree, E. H. Lavin, R. Nishida, 101 Y.-W. Feng, T. E. Acree. Foods Food Ingre- S. Watanabe in Flavour Science and Tech- dients J. Jpn. 1999, 179, 57–66. nology (Ed.: Y. Bessie`re, A. F. Thomas), 102 G. H. Fuller, R. Steltenkamp, G. A. Wiley & Sons, Geneva, 1990, 49–52. Tisserand. Ann. N.Y. Acad. Sci. 1964, 114 A. B. Marin, T. E. Acree, J. H. Hotchkiss, 116, 711–724. S. Nagy. J. Agric. Food Chem. 1992, 40, 103 A. Dravnieks, A. J. O’Donnell. J. Agric. Food 650–654. Chem. 1971, 19, 1049–1056. 115 A. Bravo, J. H. Hotchkiss, T. E. Acree. 104 T. E. Acree, R. M. Butts, R. R. Nelson, J. Agric. Food Chem. 1992, 40, 1881–1885. C. Y. Lee. Anal. Chem. 1976, 48, 1821– 116 F. Ulrich, W. Grosch. Z. Lebensm. Unters. 1822. Forsch. 1987, 184, 277–282. 105 T. E. Acree, J. Barnard, D. G. Cunningham. 117 W. Grosch. Trends Food Sci. Technol. 1993, Food Chem. 1984, 14, 273–286. 4, 68–73. 106 P. Pollien, L. B. Fay, M. Baumgartner, 118 W. Grosch. Flavour Fragr. J. 1994,9, A. Chaintreau. Anal. Chem. 1999, 71, 147–158. 5391–5397. 119 Y. Kotseridis, A. Razungles, A. Bertrand, 107 K. D. Deibler, T. E. Acree, E. H. Lavin. R. Baumes. J. Agric. Food Chem. 2000, 48, J. Agric. Food Chem. 1999, 47, 1616–1618. 5383–5388. 108 R. T. Marsili, N. Miller. J. Chromatogr. Sci. 120 M. R. McDaniel, R. Miranda-Lo´pez, 2000, 38, 307–314. B. T. Watson, N. J. Micheals, L. M. Libbey. 109 T. E. Acree, J. Barnard. In Trends in Flavour In Flavor and Off-flavors (Proceedings of the Research (Ed.: H. Maarse, D. G. van 6th International Flavor Conference) (Ed.: der Heij), Elsevier, Amsterdam, 1994, G. Charalambous), Elsevier Science, 211–220. Amsterdam, 1990, 23. 110 H. Guichard, E. Guichard, D. Langlois, 121 R. Miranda-Lo´pez, L. M. Libbey, S. Issanchou, N. Abbott. Z. Lebensm. Unters. B. T. Watson, M. R. McDaniel. J. Food Sci. Forsch. 1995, 201, 344–350. 1992, 57, 985–993, 1019. 111 D. G. Cunningham, T. E. Acree, J. Barnard, 122 N. Sanchez, C. L. Lederer, G. Nickerson, R. M. Butts, P. A. Braell. Food Chem. 1986, L. M. Libbey, M. R. McDaniel. In Food 19, 137–147. Science and Human Nutrition (Ed.: G. Charalambous), Elsevier Science, Amsterdam, 1992. 55

3 Odor Handling and Delivery Systems

Takamichi Nakamoto

Abstract A handling and delivery system significantly contributes to the capability and reliability in an odor sensing system. Various techniques of the sample flow, static, and preconcentrator systems are described in the present chapter. The sample flow system is convenient because the measurement cycle is short and easy to handle. The static system is the basic one used to measure the steady-state sensor response. A preconcentrator is often used to enhance the sensitivity, and can be also used to autonomously enhance the selectivity of a sensor array. Direct exposure of the sensor to the vapor is sometimes used in field measurement. The analysis of the transient sensor response, a homogeneous sensor array for an olfactory video camera, and the sensor responses in the plume-tracing robot are briefly introduced. Due to the variety of methods available, the most appropriate odor handling and delivery system should be selected for the project.

3.1 Introduction

There are two main types of odor handling and delivery, the sample flow system and the static system. In the sample flow system the sensors are placed in the vapor flow, which allows the rapid exchange of vapor and hence many samples can be measured within a short time. In the static system there is no vapor flow around the sensor, and measurements are usually made on the steady-state responses of the sensors exposed to vapor at a constant concentration. The sample flow and static systems are closed units. In a third method the direct exposure to the vapor is described, which is an open system having no sensor chamber, hence rapid concentration change around the sensors is measured. Three examples are given.

3.2 Physics of Evaporation

Most of the samples tested in an odor sensing system are liquids from which odorants are evaporated. It is therefore important to know the physicochemical behavior of evaporation when you design an odor handling and delivery system. One of the most important points is that saturated vapor pressure is dependent on temperature. The vapor concentration should be kept below the maximum corresponding to the saturated vapor pressure, otherwise the excess of the vapor pressure above the saturated point leads to its condensation into liquid drops. The relationship between saturated vapor pressure P and temperature T is:

lnðPÞ¼c=T þ d ð1Þ

where c and d are constants. Vapor pressures at several temperatures are summarized

in the literature [1]. Let P3 and P4 be the saturated vapor pressures at T3 and T4, respectively. The constants are different for different vapors and can be determined from Eqs. (2) and (3).

lnðP ÞlnðP Þ c ¼ 3 4 ð2Þ 1 1 T3 T4

and

T lnðP ÞT lnðP Þ d ¼ 4 4 3 3 : ð3Þ T4 T3

The saturated vapor can therefore be obtained at arbitrary temperatures. The pressure of a compound with high odor intensity is typically small whereas highly volatile compounds have high saturated vapor pressures. When there is a mixture of compounds, the phenomenon of the vapor-liquid equilibrium state becomes a little complicated. In ideal solutions, Raoult’s law expressed as:

0 PA ¼ NAPA ð4Þ

is valid. PA is partial pressure of compound A, NA the molar ratio of that compound in 0 the solution, PA the vapor pressure of the pure compound. Equation (4) indicates that the partial pressure of the ideal solution is equal to the product of its molar ratio and the vapor pressure of the pure compound. In the ideal solution, the superposition theorem for the plural compounds is valid. Most compounds, however, are non-ideal solutions. In the non-ideal solution, Eq. (4) is replaced by:

0 PA ¼ cANAPA ð5Þ 3.3 Sample Flow System 57 where cA is an activity coefficient dependent upon NA. Since interaction between the components occurs in the non-ideal solution, the superposition theorem for the compound mixture is not valid. In that case, the equation derived by Wilson is more suitable [2].

3.3 Sample Flow System

The sample flow system is the most popular odor handling and delivery system. Sev- eral sample flow systems exist such as headspace sampling, diffusion, permeation, and bubbler, and sampling bag methods are described.

3.3.1 Headspace Sampling

Figure 3.1 shows a schematic diagram of a headspace sampling method. The headspace is the space just above the liquid sample in a bottle. The carrier gas such as dry air is supplied at the inlet and the vapor evaporated at the liquid surface carried by the carrier gas is supplied to the sensors. Solenoid valves alternately switch the pure carrier gas and the headspace sample vapor, and the difference in the sensor output is recorded. The frequency shift from that in air to that in the sample vapor is regarded as the sensor response in the case of a quartz crystal microbalance (QCM) gas sensor. A semiconductor gas sensor response of a ratio of the resistance in the sample vapor to that in the air. The distance between the liquid surface and the tips of the syringe needles should be kept constant since the vapor in the headspace is often unsaturated and its concentration varies according to its distance. The headspace sampling method is an easy method to use as described in the literature [3–7]. Although many samples can be measured within a short time, the supplied vapor concentration is not known and varies during the vapor supply. The vapor concentration at the outlet of the bottle gradually changes until it reaches the liquid- vapor equilibrium as is illustrated in Fig. 3.2(a). The vapor-concentration profile sometimes influences the waveform of the sensor response, which is a convolution of its profile and the sensor impulse response. When a sufficiently narrow vapor pulse, as shown in Fig. 3.2(b), is supplied to the sensor, the sensor response is not influenced by

Fig. 3.1 Headspace sampling 58 3 Odor Handling and Delivery Systems

Fig. 3.2 Concentration at outlet of headspace sampler, (a) long duration of flowing carrier gas and (b) vapor pulse

the concentration profile. The pulse width of the vapor should be sufficiently smaller than the sensor response/recovery time. The pulse vapor supply method can therefore be used to ignore the influence of the concentration variation during vapor supply [7]. The example of the headspace sampling method is shown in detail in Fig. 3.3 [8]. Dry air is supplied to the sample bottle through a mass flow controller (MFC). MFC is used to precisely control the flow rate independently of its pressure load. The sample bottle is a vial of volume 22 mL. Liquid samples, such as perfumes and flavors, with the typically constant volume of 4 mL are poured into the vials using a micropippete. The vials are placed in a temperature-controlled bath to avoid temperature induced variations of vapor pressure. The syringe needles are driven by an autosampling stage that moves from vial to vial allowing many samples to be measured automatically. Care should be taken that the arm of the stage is not deformed when the syringe needles pierce the rubber seal (septum) of the vial, because the distance between the liquid surface and the tips of the needles must be kept constant. The dry air and the sample vapor are alternately switched by miniature solenoid valves controlled by a computer and supplied to a sensor cell a sensor housing. A solenoid valve with a small internal volume is recommended, and is often driven by DC voltage of 12 V or 24 V. An internal solenoid wall made of Teflon prevents odorant adsorption; additionally it is preferable to repeatedly and quickly switch the solenoid valve for a few minutes after each measurement to ensure it is kept clean. The drive circuit simply consists of a discrete transistor. Zenor diodes or surge absorbers are often used to suppress the surge generated when the valve is switched. The input of the circuit can be connected to the printer port of a computer. The sensor cell is one of the most important parts in an odor-handling system, its structure determining the response time. The sensor response is sometimes influenced by its position within the cell, especially when dense vapor with a high boiling 3.3 Sample Flow System 59

Fig. 3.3 Headspace measurement system using autosampling stage point is supplied. The cell should therefore have an internal volume as small as possible to minimize any effect due to the sensor location. The sensor response however, is sometimes slow enough to ignore that effect. There are several types of flow cells as shown in Figs. 3.4(a) and (b). The structure illustrated in Fig. 3.4 (a) is easy to fabricate because the flow rate at each sensor is the same. The sensor response is independent of location when placed in parallel as is illustrated in Fig. 3.4(b) [9]. Additionally the same flow rate at each sensor can be guaranteed by having a symmetrical structure. A more sophisticated structure with a tiny internal volume is shown in reference [7]. The material of the

Fig. 3.4 Structures of sensor cells for a sensor array. (a) series type and (b) parallel type 60 3 Odor Handling and Delivery Systems

sensor cell is typically stainless. A type of rubber having low odor adsorption should be used to prevent gas leaking when the sensor is attached to the sensor cell. If the sensor responses are likely to be influenced by temperature, the sensor cell can be kept at constant temperature by use of a peltier device, or immersion in a thermo bath. In the example in Fig. 3.3, the water from the temperature-controlled bath used for the sample bottles is circulated around the sensor cell.

3.3.2 Diffusion Method

In this method diffusion of vapor from a tube of accurately known dimensions is measured [10]. Low concentrations are usually measured, because it is difficult to obtain vapor with concentration more than a few percent above saturation using this method. An example apparatus is shown in Fig. 3.5. The liquid in the reservoir is allowed to evaporate and the vapor slowly diffuses from a reservoir through the diffusion tube into a flowing gas stream at a constant rate. The resultant mixture concentration is determined by the ratio of the diffusion rate to that of the flowing gas stream. The reservoir filled with liquid is kept at constant temperature since the diffusion coefficient of the vapor depends upon the temperature. The diffusion rate is given by:

DMPA P S ¼ ln ð6Þ RTL P p

where S is the rate of diffusion of vapor out of the capillary tube (g/ml), M is the relative molecular mass of the vapor (g/mol), P the pressure in the diffusion cell at the open end of the capillary (atm), A the cross-sectional area of the tube (cm2), D the diffusion coefficient (cm2/s), R the molar gas constant (mL atm mol1 K1), T temperature (K), L

Fig. 3.5 Apparatus for diffusion method 3.3 Sample Flow System 61

Fig. 3.6 Permeation tube the length of the capillary tube (cm), and p is the partial pressure of the sample vapor (atm). The actual concentration sometimes deviates from Eq. (6) when the vapor above the liquid is not saturated. An alternative method is to precisely measure the mass change of the liquid reservoir during the constant period using a balance. The reduction in the amount of liquid over a certain time indicates the diffusion rate. It is a reliable method in spite of the time taken. Several researchers in the EN field use the standard gas generator based upon the diffusion method because it is commercially available. Examples of sensor systems including diffusion methods are described in the references [11, 12].

3.3.3 Permeation Method

The permeation method is similar to the diffusion method, using similar equipment except that a permeation tube is used (Fig. 3.6). Liquefied gas, when enclosed in an inert plastic tube, may escape by dissolving in and permeating though the walls of the tube. The permeation rate is proportional to the length of the tube and varies logarith- mically with 1/T, hence temperature should be kept constant. Permeation tubes of several kinds of vapors are commercially available.

3.3.4 Bubbler

A bubbler is a bottle in which a vapor is generated by bubbling, as illustrated in Fig. 3.7. A carrier gas such as air is passed through the liquid in the bottle, and takes away the generated vapor. Although it is easy to obtain the vapor by this method, several points should be taken into account. The headspace over the liquid sample sometimes does not saturate. Glass particles are sometimes put in the liquid so that the area of contact between the liquid and the carrier gas can be increased. Moreover, tiny liquid particles, not vaporized ones, are sometimes carried to the sensors due to heavy bubbling at a fast flow rate. Examples of the bubblers are given in references [13, 14]. 62 3 Odor Handling and Delivery Systems

Fig. 3.7 Bubbler

3.3.5 Method using a Sampling Bag

Vapor is generated after a liquid sample is injected to the large-air-volume sampling bag by syringe, and then evaporated. The vapor in the bag is then sucked out using a pump and introduced into a sensor cell, as illustrated in Fig. 3.8(a). The vapor concentration is determined by the combined volume of the injected liquid and that of the sampling bag. MFCs are used if a vapor blender is constructed. The concentration of each vapor is determined by the ratio of the corresponding flow rate to the total flow rate. The sample flow system in Fig. 3.8(a) is a simplified one, the actual system has solenoid valves to switch the vapor abruptly [15]. It takes a little time for the MFC flow rate to settle to the set point value. The baseline of the sensor response depends on the flow rate, which should therefore be kept constant. A more sophisticated system is illustrated in Fig. 3.8(b). The valves V1 and V1,V2 and V2 are complementarily switched. The flow rates of both the vapor from the blender and the air are always the same and constant. Each flow path to either the sensor cell or the bypass is abruptly switched by the solenoid valves without changing the flow rate at the sensor cell. The air in the vapor blender is used to keep the flow rate constant at the outlet of the vapor blender. The material of the sampling bag should be carefully selected to avoid water and other molecules permeating through. The generation of vapors can also occur in certain types of plastic bag. Adsorption of the sampled vapor inside the bag cannot be ignored in case of low concentration. A fluorine-containing resin bag is the best one due to low permeability and low adsorption capability. A glass vessel is often used to sample the atmosphere in environmental testing, and so requires careful handling so that it cannot be broken – it is also expensive. Systems using several MFCs are reported because vapor with an arbitrary concentration is automatically and rapidly generated [16]. They are convenient in spite of the fact that they are expensive. 3.3 Sample Flow System 63

Fig. 3.8 Vapor supply method using a sampling bag (a) simplified method and (b) actual method 64 3 Odor Handling and Delivery Systems

3.4 Static System

The fundamental static system measures the steady-state response of a sensor to a vapor at constant concentration and at a constant temperature. In the case of a QCM sensor, the most basic characteristic such as a partition coefficient of a sensing film, defined as the ratio of the concentration in the film to that in the vapor, can be obtained in this system. The principle is illustrated in Fig. 3.9. The tiny volume, typically a few microliters, of liquid sample is injected into a chamber having a volume of a few liters, and is evaporated. The sensor response is measured after equilibrium is reached [17–24]. The chamber is typically made of Teflon or glass to avoid vapor adsorption onto the internal wall. The whole chamber can be immersed in a temperature-controlled bath, thus the whole system can be kept at the same temperature; in the sample flow system the temperature at the sensor sometimes does not agree with that of the vapor. Manual injection of the sample liquid by the syringe is the basic method, however it is possible to automate this procedure [25]. Because the volume of the plumbing tube cannot be ignored, a technique similar to FIA (Flow Injection Analysis) is used to sample a few microliters of the liquid precisely. The automated system consists of a sample selector, a sample injector, and the measurement system. It selects the samples among several candidates, injects the sample liquid and measures the sensor responses after equilibrium. Since it takes time to measure the steady-state response due to the slow evaporation of the sample liquid, the automation is quite indispensable if many data need to be systematically measured.

Fig. 3.9 Principle of the static measurement system 3.5 Preconcentrator 65

3.5 Preconcentrator

3.5.2 Sensitivity Enhancement

A preconcentrator tube is often used to enhance the sensitivity of the sensor [26, 27]. After it accumulates the vapor, a heat pulse is applied to the tube to desorb the concentrated vapor, and the limit of detection is thus improved. Although it originates from the technique of gas chromatography, it is often used for sensors. A simplified preconcentrator system is illustrated in Fig. 3.10. A preconcentrator stainless tube with a length of a few cm is packed with adsorbent such as Tenax- TA. The adsorbent of a few tens of milligrams is held in place with glass wool. The tube is heated using a coil of insulated nichrome wire around it in case of thermal desorption. A temperature controller is used to adjust the power supplied to the heater so that the temperature given by a computer can be maintained. The typical temperature during heating is around 200 8C, hence the use of metal connectors for plumbing the preconcentrator is recommended. However, a certain heatproof flexible tube is available for plumbing it if the heat capacity of the connectors cannot be ignored. The temperature characteristic varies in different preconcentrators since it is difficult to reproducibly wind up the heating coil. The gap between the coil and the stainless tube is critical. A flexible heater stuck to the preconcentrator tube is preferable.

Fig. 3.10 Simplified system of preconcentrator 66 3 Odor Handling and Delivery Systems

Another way is to heat the preconcentrator tube directly by flowing the current through the surface of the tube. The tube diameter should be minimized to increase the heater resistance for direct heating. There is a tradeoff between the heater resistance and the flow rate. There are several adsorbents for the preconcentrator. They are Tenax TA, Tenax GR, Carbopack B, Carbotrap, Carboxen 569, Carbosieve SIII etc [28]. Some of them are polar and some are nonpolar. The selection of the adsorbent should be according to the purpose. Some kinds of adsorbents can be used as coating films of QCM gas sensors because they are dissolved into an organic solvent. However, its characteristic seems to be different from that of particles. The sensor response is slow but cannot accumulate the vapor in the film.

3.5.2 Removal of Humidity

Hydrophobic adsorbents do not capture water. Water just goes though the preconcentrator tube whereas other vapor molecules are accumulated. It is possible to desorb the vapors without water at the heating stage after passing the sample to the tube. Since many sensor responses are affected by humidity change, it is convenient to remove the humidity before the actual measurement. Many samples such as juice, soup, and coffee include water. The removal of the water influence is indispensable for reliable measurement from the practical point of view. The removal of alcohol is sometimes required for alcoholic beverages such as beer, whiskey, liquor, wine etc. The influence of alcohol is critical especially for semiconductor gas sensors since the contributions of other components are masked by the alcohol. It is better to keep the preconcentrator temperature a little higher than room temperature even during the adsorption process so that the removal of water can be completely performed [29]. Moreover, a slightly complicated sequence is sometimes used to avoid the exposure of the sensor to the unwanted vapor (water and/or ethanol) in cases when it would take much time to recover from the response to that vapor. How- ever, the most simple and basic system of the preconcentrator is that shown in Fig. 3.10.

3.5.3 Selectivity Enhancement by Varying Temperature

3.5.3.1 Selectivity Enhancement using a Preconcentrator In addition to sensitivity increase, it is possible to enhance the selectivity of samples by using a preconcentrator. There are two ways to enhance the selectivity. First is to utilize chromatographic behavior when the gases pass through the preconcentrator tube. The second is to separate samples by varying the desorption temperature since that temperature changes from compound to compound. In the first method, the chromatographic behavior is observed [30]. The samples interact with the adsorbent and the degree of the interaction depends upon the sample 3.5 Preconcentrator 67 type. When the interaction is strong, it takes time for the sample to elute at the exit of the preconcentrator tube, in the same manner as in gas chromatography, whereas it does not take much time for sample elution in the case of low interaction. Although the separation of the samples at the exit of the preconcentrator is not sufficient, the retention time is just within a few tens of seconds and sample discrimination can be achieved by a sensor array and pattern recognition techniques. The eluted samples from the preconcentrator tube is detected by a sensor array made up of multiple sensors with different characteristics, and with its output pattern recognized by a neural network or multivariate analysis. It can be regarded as a kind of higher-order sensing technique [31] since the information is included in both transient waveform of each sensor and output pattern from the sensor array. The second method of selectivity enhancement using the preconcentrator tube is to vary the temperature for the vapor desorption [32, 33]. When the vapor accumulated in the preconentrator tube is desorbed by raising its temperature, each vapor will desorb at its own particular desorption temperature. Although it depends upon its boiling point, other factors such as polarity seem to influence it. When the temperature of the preconentrator tube is changed several times, the vapor with low desorption temperature appears due to the small temperature increase, whereas the one with high desorption temperature comes to the sensors due to the large temperature increase. Thus, the adequate sequence of the preconcentrator heating improves the selectivity. The abrupt temperature change is preferable to a ramp shape of the temperature profile because a sharp sensor response is obtained at the point of the abrupt temperature change.

3.5.3.2 Autonomous System with Plasticity It is possible to autonomously obtain the heating sequence according to the samples desorption temperatures. The pattern separation among the samples is improved after the optimization of the heating sequence. Since the number of odor types is huge and there are too many parameters to be optimized, such autonomous behavior is helpful to achieve good capability of discrimination for a short time. It is a kind of active sensing system [34] since the sensing system itself enhances its capability autonomously through interaction with the targets. The flexible and accurate system can be realized based upon that concept, compared with the conventional passive sensing system. This concept was first realized in the semiconductor gas sensing system and is called a characteristic of plasticity [11]. Plasticity is the biological capability, e.g. synaptic modification, to organize in such a way as to adapt to an environment. The characteristic of plasticity can be realized in an odor sensing system only when three fundamental technologies are available. They are the gas sensor device with its characteristic easily changed by a controllable parameter, the evaluation index of the adaptation, and the algorithm for changing the parameters. In the preconcentrator system, the gas sensor device mentioned above is the preconcentrator tube with variable desorption temperature in addition to a sensor. Its characteristic is easily controlled by the voltage applied to the preconcentrator heater. 68 3 Odor Handling and Delivery Systems

The second fundamental technology in the preconcenetrator system is the index of the pattern separation among the target samples. The sensor-array output pattern can be regarded as a vector with each component corresponding to each sensor response. Several pattern separation indices for the output pattern from a sensor array such as the Euclidean distance in that vector space, the vector angles, Maharanobis distance, Wilks Lambda [35] are available. The simple indices such as the Euclidean distance and the vector angle are preferable when the number of the data is insufficient. The third fundamental technology in the preconcentrator system is the algorithm often called an optimization algorithm [36]. This is the algorithm used to obtain the appropriate parameter values by evaluating the index. The parameters are repeatedly modified in the optimization process until a good evaluation index is obtained. There are several optimization algorithms such as the simplex method and method of stee- pest descent [15, 37]. In the preconcentrator system, the optimization process is illustrated in Fig. 3.11. Since several heat pulses with different temperatures are repeatedly applied to the preconcentrator tube at every measurement, the pattern separation index is expressed as a function of those temperatures. The shape of the curved surface of the pattern separation index is not a priori known. The temperature profile is iteratively modified to find the point with the maximum index. The exploration task is achieved using the optimization algorithm. ðiþ1Þ In the case of the gradient method, the jth peak temperature at the i þ 1 step Tj is determined by:

ðiþ1Þ ð0Þ @I Tj ¼ Tj þ e ð7Þ Tj

Fig. 3.11 Principle of realizing plasticity 3.5 Preconcentrator 69 where I is the pattern separation index, e the empirically determined constant. The point 0 is that with the maximum index before the i þ 1 measurement. @I around the @Tj point 0 can be approximately obtained using the data around that point [38]. The derivation of @I is shown in the appendix. @Tj

3.5.3.3 Experiment on Plasticity An example of autonomous enhancement of the selectivity is described here. The preconcentrator tube was packed with 30 mg Tenax-TA. Two samples were pure propyl acetate and the mixture of propyl acetate and hexyl acetate (ratio 1:1 v/v). The headspace vapors were supplied to seven QCM (20 MHz, AT-CUT) sensors at the flow rate of 200 ml/min and the time for the vapor accumulation was 13 s. The sensor coatings were Tenax-TA, UCON-90000, DEGS (Diethylene Glycol Succinate), squalane, sphin- gomyelin, ethylcellulose and PolyEthyleneGlycol (PEG) 1000. The heat pulses were applied three times and the final temperature peak was fixed to 230 8C. The temperature profile can be expressed by the two parameters of the first and second peak temperatures. Since three heat pulses were applied during one cycle and the number of the sensors was seven, the dimension of the sensor response pattern was 21. The pattern separation index used was a vector angle between two samples. It was found that the index was successfully improved after the temperature profile was modified five times. Two typical sensor responses extracted among seven sensors before and after temperature-profile modification are shown in Fig. 3.12 (a) to (d). The sensors were QCMs coated with polar film DEGS and nonpolar film squalane. The first peaks were the responses to the vapors not accumulated at the preconcentrator tube during the vapor supply. The peaks at 110 s and 160 s were the responses caused by the second and third heat pulses. Although the first heat pulse was applied at 50 s, the responses at that point were small. It was found from the figures that the difference of the response pattern between the two samples became larger after the exploration of the temperature profile. The difference appeared at the peak due to the third heat pulse. The two samples were easily discriminated after the optimization when the responses at the third peak were taken into account. The plastic characteristic can be realized using preconcentrator tube with variable temperature and optimization algorithm. 70 3 Odor Handling and Delivery Systems

Fig. 3.12 Sensor responses before and after modification (a) mixture of propyl acetate and hexyl acetate before optimization, (b) propyl acetate before modification, (c) mixture of propyl acetate and hexyl acetate after modification, (d) propyl acetate after modification

3.6 Measurement of Sensor Directly Exposed to Ambient Vapor

3.6.1 Analysis of Transient Sensor Response using an Optical Tracer

The flow-type and the static systems mentioned above are closed systems. It is also possible to directly expose a sensor to a gas. Direct exposure is often performed when the rapid concentration change in an open system should be captured. How- ever, the sensor response does not correspond to instantaneous gas concentration due to its response delay, even if it is open to the ambient atmosphere. The sensor dynamics can be analyzed when both sensor response and gas concentration change are simultaneously obtained [39, 40]. Thereafter, it is possible to model the sensor dynamics. 3.6 Measurement of Sensor Directly Exposed to Ambient Vapor 71

The instantaneous gas concentration is approximately obtained using an optical tracer because a response speed of an optical sensor is much faster than that of the gas sensor. One of the smart methods is to utilize the optical tracer accompanied with smell. The use of white smoke from joss sticks is a good candidate for that purpose. The gas concentration is measured as the brightness of the CCD camera image, and the transient response of the sensor to the smoke measured simultaneously. The light sheet is illuminated by a xenon lamp through a slit just above the gas sensor. When the smoke of the joss sticks flows, the gas sensor responds to the smoke. Si- multaneously, the light scattered by the smoke particles is captured by the camera. The optical data is sampled at the video rate and the brightness integration over the area of the gas sensor can be regarded as optical sensor response. The transient response of a semiconductor gas sensor was observed in a wind tunnel [40]. In the case of the semiconductor gas sensor, the rise time of the response is different from the recovery time. Thus, the transient response of the gas sensor cannot be modeled by a simple linear time-invariant system, and hence the two-phase model was proposed. In this model, the time-series data is divided into the response phase and the recovery one. It is assumed that the gas concentration in each phase can be expressed by a second- order differential equation:

d2sðtÞ dsðtÞ þ a þ b sðtÞ¼g l ðtÞð8Þ dt2 i dt i i s where s(t) is gas sensor response Rgas=Rair at time t, ai, bi, gi are constants (i ¼ 1: response phase, i ¼ 2: recovery phase). lsðtÞ is the steady-state sensor response calculated from a calibration curve. Equation (8) is transformed into the discrete-time equation.

sðk þ 1Þ¼pisðkÞþqisðk 1ÞþrilsðkÞð9Þ where s(k) and lsðkÞ are gas sensor response and transformed steady-state sensor response corresponding to the brightness at time kDt. pi, qi, ri are the constants; i ¼ 1: response phase, i ¼ 2: recovery phase. Moreover, the following constraint is required since sðk þ lÞ¼sðkÞ¼sðk 1Þ¼lsðkÞ in the steady state.

pi þ qi þ ri ¼ 1 ði ¼ 1; 2Þ: ð10Þ

The scheme for dividing a time-series data into the two phases is as follows. If the gas concentration increases rapidly and lsðkÞ becomes less than s(k), the gas sensor response begins to decrease toward lsðkÞ. Thus, the data at that moment can be regarded as the response-phase data. If lðkÞ > sðkÞ, the data at that moment can be regarded as the recovery phase data in the same way. The parameters, pi, qi and ri are estimated for the response phase and the recovery phase respectively using the least-squares method. The response of the semiconductor gas sensor (TGS800, Figaro) is compared with the calculated value based upon Eqs. (8) to (10), as shown in Fig. 3.13. The gas sensor 72 3 Odor Handling and Delivery Systems

Fig. 3.13 Comparison of gas sensor response calculated from the optical data with measured data

response calculated from the optical data was in good agreement with the experimental data. Since the accuracy depends upon the gap between the sensor and the light sheet, that gap should be minimized to obtain the high accuracy of the data. This is one of the methods used to model gas sensor behaviors. If the flow-type system is used, the actual concentration profile at the sensor is sometimes different from that at the vapor source. The exact waveform of the gas concentration is required to model the behavior of the gas sensor response. Since the gas concentration just above the sensor is obtained in this method, it is beneficial for sensor-behavior modeling. When the gas sensor response is very fast, the modulation of the gas concentration by moving gas outlets was effective to estimate its time constant [41].

3.6.2 Homogenous Sensor Array for Visualizing Gas/Odor Flow

Another example of direct exposure of the sensor to ambient vapor is the visualization system of the gas-concentration distribution. A two-dimensional homogeneous sensor array can capture the dynamic scene of the gas flow as illustrated in Fig. 3.14(a). It is called an olfactory video camera since the dynamic distribution of the gas concentration can be stored in a computer and can be played back in the same way as that of a conventional video. Knowing the direction of the gas flow from the dynamic image and simultaneous measurement of the gas concentration at many points enhances the reliability of the gas-flow direction estimation, because the influence of the wind turbulence is large on the measurement data. After the initial experiment on the pulse drive semiconductor gas sensor array [42], the 5 5 QCM gas sensor array was fabricated [43]. The recovery time of a QCM gas sensor, which is quite essential in the gas flow visualization, is typically less 3.6 Measurement of Sensor Directly Exposed to Ambient Vapor 73

Fig. 3.14 Olfactory video camera (a) concept and (b) binary image from olfactory video camera than 1 second. A miniaturized QCM (27.8 MHz, AT-CUT, SMD type) with internally installed oscillator (8 4 mm) was used. It was coated with phosphatidylcholine, which is relatively sensitive to triethyl amine used as a target vapor. It is one of the typical bad smells and the system is expected to be applied to environmental monitoring. The compact 25-channel frequency counter implemented into a FPGA (Field Programmable Gate Array) was used to measure their sensor responses every second. The 25-channel frequency data were transferred to a computer via RS232C interface. The experiment was conducted in a wind tunnel with wind speed less than 5 cm/s. The headspace vapor of triethyl amine was spouted from a nozzle at the rate of 75 ml/ min. The dynamic scene of triethyl amine behavior was captured by the olfactory video camera as is shown in Fig. 3.14 (b). The images are for three successive seconds and are displayed as binary images to enhance the contrast. The flow direction was almost always grasped using this system since many fragments of the odor cloud generated by the turbulence come to the array and the vapor distribution is not uniform. The clear image was obtained here because of quick response/recovery time of QCM gas sensors. Furthermore, the direction estimation was successfully performed using the image processing algorithm [44] even if the instantaneous wind direction is not constant. Some gas sensors can be directly exposed to the vapor to obtain the instantaneous concentration in the field with wind turbulence and the example of QCM gas sensor has been described in this subsection. 74 3 Odor Handling and Delivery Systems

Fig. 3.15 Gas sensor responses mounted on a mobile robot searching for target vapor source

3.6.3 Response of Sensor Mounted on an Odor-Source Localization System

The final example of the direct exposure of the sensor to the vapor is an odor-source localization system [45–47]. It is a plume-tracing robot that can locate the source of the odor using the gas sensors and the anemometric sensors. The gas sensors mounted on that robot were directly exposed to the vapor. Tin-oxide gas sensors (TGS822, Figaro) are mounted on the robot to determine the gas concentration gradient. The robot moves to find the plume if it is situated outside the plume, whereas it moves along the wind direction if it is inside the plume. The fast response/recovery time as well as sensitivity is required. The sensor responses to the vapor during the process of the target-source exploration are shown in Fig. 3.15. The ethanol vapor was spouted in the clean room where the wind field was relatively constant. The responses of the four gas sensors mounted on the same robot are shown here. The four sensor responses were calibrated in advance. The starting point of the robot was 1.3 m away from the target source and the speed of 1 the robot was 3 cm s . Since the sensor response is expressed as Rgas=Rair, the response value is small when the concentration is high. It is seen in the figure that the robot was approaching the target source because the ethanol concentration was increasing. However, the speed of the robot was limited by the response/recovery time of the gas sensor. When the robot moved too fast, it wandered around the same place and could not escape from it. Especially, the recovery time is important since the sensor with fast response time does not always have fast recovery time. A sensor with faster recovery time is required for the robot application.

3.7 Summary

The most convenient odor handling and delivery system is a sample-flow system, because it is easy to handle and the measurement cycle is short. On the other hand, the static system is suitable for studying the fundamental behavior of the sen- 3.7 Summary 75 sor. The direct exposure of the sensor to the vapor is sometimes performed in case of the field measurement. The appropriate method should be selected according to the purpose.

Appendix: Optimization Algorithm for Realizing Plasticity

@I Let us briefly see how @Tj in Eq. (7) is determined. The curved surface mentioned above is defined as

I ¼ f ðT1; T2; :::; TmÞðA:1Þ where m is the number of the temperature peaks. The function above can be expanded around point 0 in the following manner. @f ð0Þ @f ð0Þ I I0 ¼ ð0Þ ðT1 T1 Þþ ðT2 T2 Þ T1¼T @T1 1 @T2 T ¼T ð0Þ 2 2 @f ð0Þ þ::: þ ðTm Tm ÞðA:2Þ @T ð0Þ m Tm ¼Tm

where I0 is the index at point 0. Using n point data around point 0, the following equations are obtained. @f ð0Þ @f ð0Þ I1 I0 ¼ ð0Þ ðT11 T1 Þþ ðT21 T2 Þ T1¼T @T1 1 @T2 T ¼Tð0Þ 2 2 @f ð0Þ þ::: þ ðTm1 Tm Þ @T ð0Þ m Tm ¼Tm @f ð0Þ @f ð0Þ I2 I0 ¼ ð0Þ ðT12 T1 Þþ ðT22 T2 Þ T1¼T @T1 1 @T2 T ¼Tð0Þ 2 2 @f ð0Þ þ::: þ ðTm2 Tm Þ @T ð0Þ m Tm ¼Tm . . @f ð0Þ @f ð0Þ In I0 ¼ ð0Þ ðT1n T1 Þþ ðT2n T2 Þ T1¼T @T1 1 @T2 T ¼Tð0Þ 2 2 @f ð0Þ þ::: þ ðTmn Tm Þ @T ð0Þ m Tm ¼Tm

Tlk is the lth peak temperature of the kth point around the point 0. Note that n should be larger than m.IfD~I, DT, @f , i.e., the approximate gradient vector at the point 0, are @~T defined as 2 3 I I 6 1 0 7 6 I2 I0 7 D~I ¼ 6 . 7; ðA:4Þ 4 . 5

In I0 76 3 Odor Handling and Delivery Systems 2 3 T T ð0Þ T Tð0Þ T T ð0Þ ::: T Tð0Þ 6 11 1 21 2 31 3 m1 1 7 6 ð0Þ ð0Þ ð0Þ ::: ð0Þ 7 6 T12 T1 T22 T2 T32 T3 Tm2 T1 7 6 . . . . . 7 DT ¼ 6 . . . . . 7 ðA:5Þ 6 . . . . . 7 4 . . . . . 5 . . . . . ð0Þ ð0Þ ð0Þ ð0Þ T1n Tn T2n T2 T3n T3 ::: Tmn T1

and 2 3 @f 6 7 6 @T1 7 6 7 6 @ 7 6 f 7 @f 6 7 ¼ 6 @T2 7; ðA:6Þ ~ 6 7 @T 6 . 7 6 . 7 4 5 @f

@Tm

then, Eq. (A.7) is obtained. @f D~I ¼½DT : ðA:7Þ @~T

Note that DT is generally non-square matrix. e2, the sum of the squares of errors in n measurement points is @f T @f e2 ¼ D~I ½DT DI ½DT : ðA:8Þ @~T @~T

@e2 @e2 is replaced by the variable ai. is @Ti @~a 2 3 @e2 6 7 6 @ 7 6 a1 7 6 7 6 @e2 7 @e2 6 7 6 @a 7 ¼ 6 2 7: ðA:9Þ @~a 6 . 7 6 . 7 6 7 4 @e2 5

@am

It becomes

@e2 ¼ 2ð½DTT ½DT~a ½DTT D~IÞ: ðA:10Þ @~a 3.7 Summary 77

@e2 Since @~a is zero at the point with the least squares of errors, the gradient direction~a is given by

~a ¼ð½DTT ½DTÞ1ð½DTT D~IÞ: ðA:11Þ

ð½DTT ½DTÞ is a symmetrical matrix. When the determinant of that matrix is close to zero, the inverse matrix is unstable and is not reliable. In that case, the eigenvalue analysis technique called the singular value decomposition technique is used to suppress the contribution of the negligibly small eigenvalues. The pseudo-inverse matrix is obtained using SVD technique [38].

References

1 D. R. Lide (Ed.). Handbook of Chemistry and 16 J. W. Grate, D. S. Ballantine, H. Wohltjen. Physics, 76th Edition, CRC Press, 6–77 Sensors and Actuators B, 11 (1987) 173. (1995). 17 W. P. Carey, K. R. Beebe, B. R. Kowalski. 2 G. M. Wislon. J. Am. Chem. Soc., 86 (1964) Anal. Chem. 59 (1987) 1529. 127. 18 J. V. Hatfield, P. Neaves, P. J. Hicks, K. 3 T. Nakamoto, A. Fukuda, T. Moriizumi, Persaud, P. Travers. Sensors and Actuators Y. Asakura. Sensors and Actuators B, 3 B, 18–19 (1994) 221. (1991) 221. 19 J. Ide, T. Nakamoto, T. Moriizumi. Sensors 4 T. Nakamoto, A. Fukuda, T. Moriizumi. and Actuators A, 49 (1995) 73. Sensors and Actuators B 10 (1993) 85. 20 H. Abe, T. Yoshimura, S. Kanaya, 5 J. W.Gardner, T. C.Pearce, S. Friel, Y. Takahashi, Y. Miyashita, S. Sasaki. P. N. Bartlett, N. Blair. Sensors and Analytica, Chimca Acta, 194 (1987)1. Actuators B, 18-19 (1994) 240. 21 K. Yokoyama, F. Ebisawa. Anal. Chem. 55 6 M. A. Craven, J. W. Gardner. Trans Inst (1993) 677. MC 20 (1998) 67. 22 T. C. Pearce, J. W. Gardner, S. Friel, 7 A. Iguchi, T. Nakamoto, T. Moriizumi. P. N. Bartlett, N. Blair, Analyst, 118 (1993) Sensors and Actuators B (2000) 155. 371. 8 J. Ide, T. Nakamoto, T. Moriizumi. Sensors 23 H. Muramatsu, E. Tamiya, I. Karube. Anal. and Actuators B, 13–14 (1993) 351–354. Chim. Acta, 225 (1989) 399. 9 H. Sundgren, F. Winquist, I. Lundstrom. 24 Y. Okahata, O. Shimizu, H. Ebato. Bull. Technical digest of Transducers 91 (1991) Chem. Soc. Jpn., 63 (1990) 3082. 574. 25 K. Nakamura, T. Nakamoto, T. Moriizumi. 10 R. S. Barratt. The analyst, 106 (1981) 817. Sensors and Actuators B, 61 (1999)6. 11 T. Nakamoto, T. Fukuda, T. Moriizumi. 26 J. W. Grate, S. L. Rose-Pehrsson, Sensors and Actuators B, 3 (1991)1. D. L. Venezky, M. Klusty, H. Wohltjen. 12 M. Nakamura, I. Sugimoto, H. Kuwano, Anal. Chem. 65 (1993) 123. R. Lemos. Technical Digest of 27 Q. Cai, J. Park D. Heldsinger, M. Hsieh, Transducers 93, 1993, p.434. E. T. Zeller. Sensors and Actuators B 62 13 M. Ohnishi, T. Ishibashi, Y. Kijima, (2000) 121. C. Ishimoto, J. Seto. Sensors and Materials, 28 W. A. Groves, E. T. Zellers, G. C. Frye. 1(1992) 53. Analytica Chimca Acta 371 (1998) 131. 14 S. J. Martin, A. J. Ricco, D. S. Ginley, 29 J. Kita, Y. Aoyama, M. Kinoshita, H. Nakano, T. E. Zipperian. IEEE Trans. On UFFC, 2, H. Akamatsu. Technical Digest of Sensor UFFC-34 (1987) 142. Symposium, IEEJ 2000, p.301. 15 T. Nakamoto, S. Utsumi, N. Yamashita, 30 R. E. Shaffer, S. L. Rose-Pehrsson, T. Moriizumi. Sensors and Actuators B, 20 R. A. McGill. Field Analytical Chemistry (1994) 131. and Technology, 2 (1998) 179. 78 3 Odor Handling and Delivery Systems

31 K. S. Booksh, B. R. Kowalski. Anal. Chem. 66 40 T. Yamanaka, H. Ishida, T. Nakamoto, (1994) 782A. T. Moriizumi. Sensors and Actuators A, 32 Y. Isaka, T. Nakamoto, T. Moriizumi. 69 (1998) 77. Technical Digest of Transducers 99 (1999) 41 P. Tobias, P. Martensson, A. Goras, 3P3.3. I. Lundstrom. Sensors and Actuators B, 33 T. Nakamoto, Y. Isaka, T. Ishige, 58 (1999) 389. T. Moriizumi. Sensors and Actuators B, 42 H. Ishida, T. Nakamoto, T. Moriizumi. 69 (2000) 58. Meeting abstract of electrochemical society, 34 T. Nakamoto, H. Ishida, T. Moriizumi. Proc. 1999, p. 1078. IEEE International Symposium on Indust- 43 T. Nakamoto, T. Tokuhiro, H. Ishida, rial Electronics (1997) SS128. T. Moriizumi. Latenews of Transducers 99, 35 W. R. Dillon, M. Goldstein. Multivariate 1999, LN9. Analysis, Wiley, 1984, p. 163–422. 44 H. Ishida, T. Yamanka, N. Kushida, 36 B. S. Gottfried, J. Weisman. Introduction to T. Nakamoto, T. Moriizumi. Sensors and optimization theory, Prentice-Hall, Engle- Actuators B, 65 (2000) 14. wood Cliffs, 1973, p. 84. 45 T. Nakamoto, H. Ishida, T. Moriizumi. Anal. 37 M. A. Sharaf, D. L. Illman, B. R. Kowalski. Chem. 4 (1999) 531A. Chemometrics, Wiley, 1986, 164–310. 46 R. A. Russell, R. A. Thiel, D. Deveza, 38 T. Nakamoto, H. Matsushita, N. Okazaki. A. Mackay-Sim. IEEE Int. Conf. Robotics Sensors and Actuators A, 50 (1995) 191. and Automation, (1995) 556. 39 J. W. Gardner, E. Llobet, E. L. Hines. Proc. 47 Y. Kuwana, I. Shimoyama, H. Miura. Int. IEE Circuits, Devices and Systems, 146 Conf. Intelligent Robots and Systems, 1995, (1999) 101. 530. 79

4 Introduction to Chemosensors

H. Nanto, J. R. Stetter

4.1 Introduction

We believe that the 21st century can be the aroma age. The culture of aroma developed with human civilization and the good smell of foodstuffs gives great comfort to the human heart. The sense of smell is, therefore, one of the most interesting of the five human senses and yet is understood the least. The human nose is widely used as an analytical sensing tool to assess the quality of such as drinks, foodstuffs, perfumes, and many other household products in our daytime activities, and of many products in the food, cosmetic, and chemical industries. However, practical use of the human nose is severely limited by the fact that the human sense of smell is subjective, often affected by physical and mental conditions, and tires easily. Consequently, there is considerable need for a device that could mimic the human sense of smell and could provide an objective, quantitative estimation of smell or odor. Recently, there has been increasing interest in the development of such a device, the so-called ‘electronic nose (e-nose)’. This is an electronic instrument that is capable of detecting and recognizing many gases and odors, and comprises a sensor array using several chemosensors and a computer. The different types of chemosensors, especially odor sensors, that have been employed within an e-nose are described in this chapter.

4.2 Survey and Classification of Chemosensors

A chemosensor is a device that is capable of converting a chemical quantity into an electrical signal and respondate the concentration of specific particles such as atoms, molecules, or ions in gases or liquids by providing an electrical signal. Chemosensors are very different from physical sensors. Although approximately 100 physical measurands can be detected using physical sensors, in the case of chemosensors this number is higher by several orders of magnitude. The types of chemosensors that can be

used in an e-nose need to respond to odorous molecules in the gas phase, which are typically volatile organic molecules with different relative molar masses. Expensive Restricted availability of light sources Size MEMs fabrication, interface electronics? Interface electronics? Operates at high temperature Very sensitive to humidity Very sensitive to humidity Needs Pd, Pt, Au, Ir (expensive) Odorant reactionproduct must penetrate gate noise immunity Low cost noRh interference noise immunity Differential devices can be quite sensitive Well understood technology Operates at room temperature, microfabricated CMOS-based chemosensor Applicable to CMOS-based chemosensor Inexpensive, microfabricated Integrated, Applicable to CMOS-based chemosensor Commercial ppb-ppm Research High electrical Commercial, several types, 1.0 ng mass change Commercial, several types, 1.0 ng mass change Research Low cost Slow response Commercial, many types, 0.1–100 ppm Research Applicable to Commercial, many types, 5–500 ppm special order only, ppm Dip-coating ResearchElectrodes High electrical Screen printing, Dip-coating, Spin coating Screen printing, Dip- coating, Spin coating Screen printing, Dip-coating, Spin coating Microfabricated, Ceramic fab. Microfabricated, Electroplating, Plasma CVD, Screen printing, Spin coating Spin coating Sputtering Fluorescence, chemoluminescence SPR Microfabricated, SAW Microfabricated, QCM Microfabricated, Thermister (pyroelectric) PellistorThermocouple Microfabricated Microfabricated Research Research Low cost Low cost Slow response Slow response Conducting polymer Fiber-optic chemosensor chemosensor chemosensor chemosensor Intensity/spectrum I-V/C-V Chemotransistor MOSFET Microfabricated Commercial, Classification of chemosensors that have been exploited so far. Metal oxide semiconductor, MOS; MOS field effect Amperometry current Toxic Gas Sensor Electrocatalyst Composite Optical Refractive index Resonant-type Gravimetric Piezoelectricity Mass-sensitive Capacitive CapacitancePotentiometric Chemocapacitor Voltage/e.m.f. Chemdiode Polymer Schottky Diode Microfabricated, Microfabricated Research Integrated, Tab. 4.1 transistor, MOSFET; quartz crystal microbalance, QCM; surface acoustic wave, SAW;Principle surface plasmon resonance, SPR. Conductometric Conductance Measurand Chemoresistor MOS Sensor type Microfabricated, Fabrication methods Availability/sensitivity Advantages Disadvantages Calorimetric Temperature Themal 4.3 Chemoresistors 81

Chemosensors as listed in Table 4.1 have been exploited and some already manufactured. Principles such as electrical, thermal, optical, and mass can be used to organize these chemosensors according to their device class. The chemosensors using metal oxide semiconductors (MOS), organic conducting polymers (CP), chemocapacitors, MOS field-effect transistors (MOSFET), quartz crystal microbalance (QCM), surface acoustic wave (SAW), surface plasmon resonance (SPR), fluorescence, and others that can be easily used as the sensor for an e-nose are included in the following discussion. Details about the types of chemosensors discussed here and others can be found in the literature [1–6].

4.3 Chemoresistors

Chemoresistors based on the conductivity change of MOS or organic CPs by chemical reaction with gaseous molecules are the simplest of gas sensors, and are widely used to make arrays for gas and odor measurements.

4.3.1 MOS

Metal oxides such as SnO2, ZnO, Fe2O3, and WO3 are intrisically n-type semiconductors. At temperatures of 200–500 8C, these respond to reducible gases such as H2, CH4, CO, C2H5,orH2S and increase their conductivity. The conductivity r and the resistivity q is given by

r ¼ 1=q ¼ enl ð1Þ where e is the charge on the electron (1:6022 1019 C), n the carrier (electron or hole) concentration (cm3) and l the carrier mobility (cm2 V1s1). In the atmosphere, some oxygen atoms areadsorbed on thesurface of n-type semiconductorsto trap free electrons from the semiconductor, and consequently a highly resistive layer is produced in the vicinity of the semiconductor surface. The adsorption of oxygen atoms on the semiconductor surface and at grain boundaries of polycrystalline semiconductors creates an electrical-double layer that acts as the scattering center for conducting electrons. The consequent increase in free electrons and decrease in scattering centers results in an increase in conductivity. The mechanism is similar for p-type semiconductors but is of opposite sign [101]. The mechanism of the increase in carrier concentration by reacting with the reducible gases as described above can be understood from the following reactions:

1 e þ O ! OðsÞ ð2Þ 2 2

RðgÞþOðsÞ ! ROðgÞþe ð3Þ 82 4 Introduction to Chemosensors

where e is an electron from the conduction band of the oxide semiconductors, R(g) is the reducible gas, and s and g imply surface and gas, respectively. Equation (2) implies that oxygen is physico-chemically adsorbed onto lattice vacancies in the oxide semiconductor, and consequently the conductivity of the oxide semiconductor becomes lower than that in the case of no adsorbed oxygen. An electron is, however, generated by the reaction with reducible gases R(g) through Eq. (3). Consequently, the conductivity is increased following Eq. (3) as a result of the increase in carrier concentration. In contrast, p-type semiconductors such as CuO, NiO, and CoO respond to oxidizable

gases such as O2,NO2, and Cl2 [101]. The schematic diagram in Fig. 4.1 explains the conductivity increase due to the car-

rier mobility for SnO2 gas sensors. In clean air, oxygen atoms that trap free electrons in the bulk SnO2, is adsorbed onto the SnO2 particle surface, forming a potential barrier in the grain boundaries as shown in Fig. 4.1a. This potential barrier restricts the flow of electrons, causing the electrical conductivity to decrease, because the potential barrier acts as the scattering center for electron conduction. When the sensor is exposed to an atmosphere containing reducible gases, e.g. combustible gases, CO, and other similar

vapors, the SnO2 surface adsorbs these gas molecules and causes oxidation. This low- ers the potential barrier, allowing electrons to flow more easily, thereby increasing the electrical conductivity as shown in Fig. 4.1b. The reaction between gases and surface oxygen will vary depending on the operating temperature of the sensor and the activity of sensor materials. The increasing sensitivity and selectivity of the sensors for exposure to gases can be realized by incorporation of a small amount of impurities and catalytic metal additives such as palladium (Pd) or platinum (Pt). The impurities act as extrinsic donors (or acceptors) and, consequently, controlling the doped amount of impurities can change the conductivity of the sensors. Doping of the catalytic metal to the sensor or coating with thin catalytic

Fig. 4.1 Schematic diagram explaining the conductivity increases

caused by the carrier mobility increase in SnO2 gas sensors. (a) Oxygen

is adsorbed onto the SnO2 particle surface, forming a potential barrier in the grain boundaries. (b) The potential barrier is lowered by means of reaction of the oxygen atoms with reducing gas, allowing electrons to flow more easily, thereby increasing the electrical conductivity 4.3 Chemoresistors 83

Tab. 4.2 Commercially available metal oxide semiconductor chemosensors.

Manufacturer Applications Model Typical detection range and features

FGARO ENG Combustible gas TGS813 For detection of various combustible gases TGS816 500–10 000 (ppm)

TGS842 Improved sensitivity to CH4 500–10 000 (ppm)

TGS821 High selectivity and sensitivity to H2 500–10 000 (ppm) Toxic gas TGS203 High selectivity and sensitivity to CO 50–1000 (ppm)

TGS825 High sensitivity to H2S 5–100 (ppm)

TGS826 High sensitivity to NH3 and amine compounds 30–300 (ppm) Solvent vapor TGS822 High sensitivity to alcohol and organic compounds such as toluene and xylene TGS823 Halocarbon gas TGS830 High sensitivity to various CFCs, HCFCs TGS831 100–3000 (ppm) TGS832 Air quality control TGS800 High sensitivity to gaseous air contaminants (such as cigarette smoke and gasoline exhaust) 1–10 (ppm) Cooking control TGS880 Vaporized gases and water vapor form food in the cooking process 10–1000 (ppm) TGS882 Alcohol vapor from food in the cooking process 50–5000 (ppm) TGS883 Water vapor from food in cooking process 1–150 (g m3)

NEW Combustible gas CH-H High sensitivity and selectivity to H2, COSMOS ELEC.CO, 50–1000 (ppm)

LTD. CH-M High sensitivity to VOCs such as CH4 and i-

C4H10 1000–10 000 (ppm) CH-CO High sensitivity to CO 100–1000 (ppm) CH-E2 High sensitivity to alcohol CH-E3 1–1000 (ppm) CH-L High sensitivity to LPgas

Toxic gas CH-N High sensitivity and selectivity to NH3

AET-S High sensitivity and selectivity to H2S Thin film type 84 4 Introduction to Chemosensors

metal film of the sensor surface changes the selectivity of the sensor. As described above, the crystallographic structure of the semiconductors used as the sensor material is commonly polycrystalline, and thus includes many grain boundaries. These grain boundaries act as the scattering centers for conducting electrons to produce the change of carrier mobility, and therefore consequently the extent of crystallinity affects the sensitivity of the sensors.

The most widely used semiconducting material as a gas sensor is SnO2 doped with small amounts of impurities and catalytic metal additives. By changing the choice of impurity and catalyst and operating conditions such as temperature, many types of gas

sensors using SnO2 have been developed. The gas sensors using metal oxide semiconductors exhibit relatively poor selectivity for gases and remain responsive to a many kinds of combustible gases. Table 4.2 lists some of the commercially available

gas sensors of SnO2 and ZnO that are manufactured by New Cosmos Electric Co., Ltd and Figaro Engineering Inc. (Japan). Figure 4.2 shows schematically the basic construction of the sintering-type and thin- film-type of gas sensors. The type of sensor materials and operating temperatures of typical gas sensors using MOSs that have been reported so far are listed in Table 4.3.

Tab. 4.3 Type of conduction and operating temperatures of typical gas sensors using metal oxide semiconductors.

Materials (Dopants) n-type or p-type Top ( 8C) Detecting gases Ref.

ZnO(Al) n 200 H2 30

ZnO(Al) n 350 NH3 31 ZnO(Al,In,Ga) n 400 TMA 32 ZnO n 280–470 CO 33

ZnO n 450 CCl2F4, CHClF2,34

WO3(Pt) n 250–400 N2H4,NH3,H2S, 35

WO3 n 500 CO, CH4,SO2 36

TiO2(Ru) n 560 TMA 37

a-Fe2O3 n 400 H2,CH4 38

c-Fe2O3 n 420 H2,CH4,C3H8C4H10,C2H5OH 39

CdIn2O3 n 300 CO 40

CuTa2O6 n 400 H2,CO 41

CuO/ZnO p/n 250 H2,CO 42

Co3O4 p 200–500 CO, H2,NOX 101

Cr2O3(Ti) n 420 TMA 43

In2O3 (Mg or Zn) n 420 TMA 43

BaSnO3 n 300–500 H2, CO, CH4,H2S, SO2 44

Bi2Sn2O7 p 500 H2, CO, C2H4,NH3 44

Bi6Fe2Nb6O30 n/p 500 C3H8,Cl2,NO2,SO2,H2S45

4.3.2 Organic CPs

Chemoresistors made from organic CPs also exhibit a change in conductance when they are exposed to reducible or oxidizable gases. Organic CPs show reversible 4.3 Chemoresistors 85

Fig. 4.2 The basic construction of the sintering-type (a) and thin-film- type (b) of the gas sensors that are commercially available

changes in conductivity when chemical substances (e.g. methanol, ethanol, and ethyl acetate) adsorb and desorb from the polymer. The mechanism by which the conductivity is changed by this adsorption is not clear at present. There are a large number of different electronically conducting polymers. Polypyr- role was first prepared electrochemically in 1968 [23] and has been most extensively studied so far. Electroconducting conjugated polymers (ECP) can exhibit intrinsic electronic conductivity. Their structure contains a one-dimensional organic backbone with alternating single and double bonds, which enables a super-orbital to be formed for electronic conduction. The most commonly applied polymers for gas-sensing applications have been polypyrrole, polyaniline, polythiophene, and polyacetylene, which are 86 4 Introduction to Chemosensors

based on pyrrole, aniline or thiophene monomers [24]. Because of their properties they have remarkable transduction matrices that are sensitive to gases and vapors, resulting in a straightforward conductance change via the modulation of their doping level. The early studies [25, 26] of the gas-sensing application of organic CPs concentrated on the response to reactive gases such as ammonia and hydrogen sulfide. Gustafsson et al. [27] have reported that gas sensors using polypyrrole films exhibit a high sensitivity for ammonia gas. Subsequent work [28–30] also showed that gas sensors using organic CPs such as polypyrrole respond to a wide range of organic vapors such as methanol. More recently, studies have been carried out on preparation of thin-film CPs for gas sensing applications [25, 31]. Thin films of heteroaromatic monomers such as pyrroles, thiophenes, indoles, and furans were grown electrochemically on interdigitated electrodes to produce gas-sensitive chemoresistors [25]. Chemoresistors using organic CPs respond to a wide range of polar molecules at temperatures as low as room temperature (RT) and more recent reports suggest that a high sensitivity down to 0.1 ppm is possible. This result indicates that organic CP is a potentially useful material for applications in odor-sensing and e-nose applications. The use of organic CPs as odor sensor materials is very attractive for the following reasons:

1) a wide range of materials can be simply prepared; 2) they are relatively low cost materials; 3) they have a high sensitivity to many kinds of organic vapors; 4) gas sensors using organic CPs operate at low temperatures.

Comparison between the properties of the organic CP odor sensor and the MOS odor sensor is shown in Table 4.4.

Tab. 4.4 Comparison of the properties of the conducting polymer odor sensor and the metal oxide odor sensor (thick-film and thin-film types).

Properties Conducting polymer SnO2 (thick film) SnO2 (thin film)

Key measurand Conductance Conductance Conductance Fabrication Electrochemical paste Sputtering, Sol-gel growth, plasma CVD Choice of materials Wide Limited Limited Operating temperature 10–110 8C 250–600 8C 250–600 8C Molecular Receptive range Wide range Combustible vapors Combustible vapors Detection Range less than 20 ppm 10–1000 ppm 1–100 ppm Response time 60 s 20 s 20 s Size Less than 1 mm2 1 3 mm Less than 1 mm2 Power Consumption Less than 10 mW 800 mW 80 mW Integrated array Yes No Yes Stability Moderate Relatively poor Poor

Interferences Acidic gases, water SO2,Cl2,H2OSO2,Cl2,H2O 4.3 Chemoresistors 87

Another way to use CPs is to make non-conducting materials, e.g. silicone [32] and polystyrene [33], conductive by inclusion of carbon-black metal powder. These sensors are used in e-noses and can exhibit high sensitivity [34].

4.4 Chemocapacitors (CAP)

The principle of chemocapacitors using polymers is schematically shown in Fig. 4.3. There are two steady states for the sensitive layer during operation. In the first state as shown in Fig. 4.3a, no gaseous analyte molecules are present in the sampling environment and consequently only air is incorporated into the polymer. As a result, a certain capacitance C of the sensitive polymer layer is measured and constitutes the baseline. In the second state, gaseous analyte molecules are present in the sampling environment as shown in Fig. 4.3b. When the polymer absorbs the gaseous analyte, the sensitive polymer layer changes its electrical (e.g. dielectric constant e) and physical properties (e.g. volume V) to produce deviations (De, DV) from the first state (reference state). The changes in electrical and physical properties of polymers are the result of reversible incorporation of gaseous analyte molecules into the polymer matrix. The CMOS-based chemical sensor using chemocapacitive microsensors for detecting volatile organic compounds (VOCs) was built with two interdigitated electrodes spin-coated or spray-coated with polymers such as (poly)etherurethane (PEUT) by Koll et al. [35].

Fig. 4.3 Chemocapacitor based on capacitance measurement of sensitive layers. There are two steady states for the sensitive layer during operation; (a) no analyte molecules are present in the sampling environment, and (b) analyte molecules are present in the sampling environment 88 4 Introduction to Chemosensors

4.5 Potentiometric Odor Sensors

Gas sensors utilizing the electrical characteristics of Schottky diodes and the MOSFET have also been investigated. Those using the Schottky diode are based on a change in the work function because of the presence of chemical species on their surface. Ex- amples are catalytic metals (inorganic Schottky diodes) such as Pd and Pt, and organic CPs (organic Schottky diodes) such as polypyrrole. Gas sensors using a MOSFET are based on metal-insulator-semiconductor structures in which the metal gate is a catalyst for gas sensing. In this section, mainly potentiometric odor sensors using MOSFETs are included and discussed.

4.5.1 MOSFET

The microchemosensor using the structure of a MOSFET in which the gate is made of a gas-sensitive metal such as Pd was first proposed by Lundstrom in 1975 [36]. This sensor exhibited a threshold voltage shift depending upon the gas concentration and was particularly sensitive to hydrogen down to the ppm level with maximum threshold

Fig. 4.4 Basic structures of n-channel MISFET and MISCAP, which operate on the same basic principle but differ in measurands 4.6 Gravimetric Odor Sensors 89

Tab. 4.5 Materials used in the different odor sensors. MOSFET – metal oxide semiconductor field effect transistor.

Chemosensor Structure Examples of sensor Examples of type materials used detecting gases

MOSFET type Metal-gate MOSFET Pd(Pt)-gate FET H2, CO, H2S, NH3

(SiO2,SnO2-Si, SiC)

Schottky type Metal/Semiconductor Pd-TiO2 (ZnO) H2, CO, CH3SH

p/n Nb2O3-Bi2O3 p/n ZnO-CuO

Metal/polymer Al/poly(3-octythiophene) NH3, NOx

Chemoresistors n-type semiconductors SnO2, ZnO, a-Fe2O3,TiO2,In2O3, H2, CO, alcohols,

V2O3,SnO2 þ Pd, ZnO þ Pt, hydrocarbons,

SnO2þ ThO2 þ Pd, O2,NO2,Cl2

p-type semiconductors CoO, Co3O4, CuO, H2,O2, CO, alcohols

Sm0.5Sr0.5CoO3,Co0.3Mg0.7O,

La0.35Sr0.65Co0.7Fe0.3O3-x

Conducting polymers Anthracene, phthalocyanine, NO, NO2,O2,SO2,

polypyrrol, polyacrylonitorile, CO, NH3, alcohols polyphenylacetylene voltage shift of about 0.5 V. The use of other metal gate materials such as Pt and Ir and operating the sensors at different temperatures has led to reasonable selectivity to gases such as NH3,H2S, and ethanol [37]. There are two basic structures such as MISFET (metal-insulator-semiconductor FET) and MISCAP (MIS CAPacitor ). The basic structures of n-channel MISFET and MISCAP that operate on the same basic principle but differ in measurands are shown in Fig. 4.4. In the MISFET, the drain current iD flowing through the semiconductor is controlled by the surface potential due to the applied gate voltage VG, and in the MISCAP the capacitance of the MIS structure is determined by the surface potential. These devices can respond to exposure to any gas that changes the surface potential or the work function of the gate metal. The materials used in MOSFET-type odor sensors as well as the Schottky- type odor sensors are listed in Table 4.5 in comparison to those of MOS-type and CP-type odor sensors.

4.6 Gravimetric Odor Sensors

Recently, gravimetric odor sensors using acoustic wave devices that operate by detecting the effect of sorbed molecules on the propagation of acoustic waves have been investigated for application to an e-nose. Two main types utilizing QCM (or bulk acoustic wave, BAW) and SAW devices have been used as the odor sensors, although other types of device have been investigated. In both types, the basic device consists of a piezoelectric substrate, such as quartz, lithium niobate and ZnO, coated with a suitable sorbent membrane [38]. Sorption of vapor molecules into the sorbent membrane coated on the substrate can then be detected by their effect on the propagation of the 90 4 Introduction to Chemosensors

acoustic wave causing changes in the resonant frequency and the wave velocity. The acoustic waves used are at ultrasonic frequencies ranging typically from 1 to 500 MHz. Both types are discussed in this section.

4.6.1 QCM

QCM or thickness shear mode (TSM) devices using BAWs in piezoelectric materials are probably the simplest type of odor sensor using a piezoelectric device. Rock crystal such as single crystal quartz has an interesting property in that it is distorted by applied electric voltage and conversely an electric field is generated by applied pressure. This phenomenon is called the piezoelectric effect. Because of this effect, upon excitation by application of a suitable a.c. voltage across the quartz crystal, the crystal can be made to oscillate at a characteristic resonant frequency. A QCM odor sensor comprises of a slice of a single crystal of quartz, typically around 1 cm in diameter, with thin-film gold electrodes that are evaporated onto both surfaces of the sliced crystal. The quartz crystal oscillates in such manner that particle displacements on the QCM sensor surface are normal to the direction of wave propagation. The thickness of the quartz crystal determines the wavelength of the fundamental harmonics of oscillation. The resonant frequency of the QCM sensor is related to the change of the mass of QCM loading by the Sauerbrey equation [39]:

2 1=2 Df ¼2f0 mf =AðqqlqÞ ð4Þ

where Df is the change in resonant frequency, f0 is the resonant frequency, mf is the mass change due to adsorption of gases, A is the electrode area, qq is the density of quartz and lq is the shear modulus. For typical AT-cut quartz crystal operating at 10 MHz, a mass change of the order of 1 ng produces a frequency change of about 1 Hz. Thus small changes in mass can be measured using a QCM coated with a molecular recognition membrane on which odorant molecules are adsorbed, as shown in Fig. 4.5. The selectivity of the QCM sensor is determined by the coating membrane deposited on the surface of the crystal. The functional design of the polymer-film-coated QCM odor sensor, based on the solubility parameter of the sensing membrane and detecting gases, was carried out in

Fig. 4.5 Schematic diagram of the structure of a QCM chemosensor. The sensor consists of polymer membrane that recognizes analyte molecules and odors, and a QCM as a transducer 4.6 Gravimetric Odor Sensors 91 order to develop a sensor with excellent selectivity and high sensitivity for harmful gases such as toluene, xylene, ammonia, and acetaldehyde by Nanto et al [40, 41]. The polymer films such as propylene-butyl, polycarbonate, and acrylic resin of which the solubility parameters almost coincide with those of toluene, acetaldehyde, and ammonia gas, respectively, are chosen as the sensing membrane material coated on the QCM surface. They found that propylene-butyl-coated sensor exhibited a high sensitivity and excellent selectivity for toluene and xylene gases, as expected from the functional design based on solubility parameters. They also found that the polycarbonate-coated and acrylic-resin-coated sensors exhibited high sensitivity and excellent selectivity for acetaldehyde and ammonia gases, respectively, also as expected. The result strongly suggests that the solubility parameter is effective in

Tab. 4.6 Research on e-noses using different types of chemosensors, including: quartz crystal microbalance, QCM; surface acoustic wave, SAW; metal oxide semiconductor, MOS; MOS field effect transistor, MOSFET. Pattern recognition types: multi-layer perception, MLP; principal component analysis, PCA; fuzzy learning vector quantization, FLVQ; cluster analysis, CA; Kohonen network, KOH; linear regression, LR; feature weighting, FW; least square, LS; discriminant function analysis, DFA; and fuzzy reasoning, FUZ.

Chemosensor type Number of sensors Applications Pattern recognition Ref.

QCM 8 Spirits, perfumes, odors MLP, PCA, FLVQ 46–51 4 Odors 52 8 Odors PCA, CA 53 6 Odors 54, 55 3 Harmful gases PCA 18, 19 SAW 6 Perfumes 56 4 Odors 57 12 58–60 MOSFET 10 Meat MLP, KOH 61 324 Odors 62 MOS 3 Odors 63 3 Odors, tobacco LR, FW 64 12 Odors, coffee LS 65, 66 12 Odors, beverages PCA, CA 67 12 Odors, beers MLP 68, 69 3 Odors CA, LS 70 12 Wines MLP, KOH 71 6 Odors LS 72 8 Odors 73 8 Odors MLP, LS 74 6 Odors LS 75, 76 7–8 Odors LR, PCA, CA 77, 78 6 Spirits, coffee CA, PCA, DFA 79, 80 3 Odors MLP 81 6 Odors KOH 82 3 Fish 83 3 Odors FUZ 84 AGS 4 Grain KNN, NN 106 92 4 Introduction to Chemosensors

the functional design of the sensing membrane of QCM odor sensors. The research on e-nose applications using QCM odor sensors as well as those using other type of chemosensors such as SAW, MOSFET, and MOS are listed in Table 4.6. Recently, studies on QCM odor sensors with plasma-polymerized organic film as the molecular recognition membrane [42–45] and odor sensors using fundamental and overtone modes of QCM with high frequency [46, 47] have been reported.

4.6.2 SAW

The SAW device is made of a relatively thick plate of piezoelectric materials (ZnO and lithium niobate) with interdigitated electrodes to excite the oscillation of the surface wave [87–89]. The SAW is stimulated by applying an a.c. voltage to the fingers of an interdigitated electrode to lead to a deformation of the piezoelectric crystal surface. The SAW devices are usually operated in one of two configurations such as a delay line and a resonator. In both cases, the propagation of the SAW is affected by changes in the properties of the piezoelectric crystal surface. In common gas sensors using a SAW device with a dual delay line structure, one arm of the delay line is coated with the sorbent membrane, the other acts as a reference to reduce the change of environmental conditions such as temperature drift and other effects. The change in frequency of

the SAW with sorption of vapor, Df V, is given by

DfV ¼ DfpcVKp=qp ð5Þ

for a simple mass loading effect, where Dfp is the change in frequency caused by polymer membrane itself, cV is the vapor concentration, Kp is the partition coefficient and qp is the density of the polymer membrane used. Considerable work [87] has been reported on the measurement of inorganic gases

such as NO2,H2,H2S, and SO2, and organic gases and vapors such as CH4,C6H6, and C2H5OH. This type of sensor using polymer materials as a sensing membrane can be chemically modified to obtain a higher degree of specificity, because the choice of chemically sensitive membrane determines the selectivity of the sensor. The SAW odor sensors generally work at much higher frequencies of the order of GHz than that of the BAW odor sensor (10 MHz). The main problems with SAW odor sensor are a relatively poor long-term stability and a high sensitivity to humidity. A good review of acoustic sensors is available [6]. 4.7 Optical Odor Sensors 93

4.7 Optical Odor Sensors

4.7.1 SPR

SPR is an optical phenomenon in which incident light excites a charge-density wave at the interface between a highly conductive metal and a dielectric material. The conditions for excitation are determined by the permittivities of the metal and the dielectric material. The SPR transduction principle is widely used as an analytical tool for measuring small changes in the refractive index of a thin region adjacent to the metal surface. The optical excitation of surface plasmon on a thin metallic film has, therefore, been recognized as a promising technique for sensitive detection of chemical species such as odor, vapor and liquid [90]. Several methods have been employed to monitor the excitation of SPR by measuring the light reflected from the sensor interface. These include analysis of angle modulation [91], wavelength modulation [92], intensity modulation [93], and phase modulation [94]. Optical SPR sensors are sensitive to the change in the refractive index of a sample surface. Recently, it has been reported that toxic gases such as ammonia, toluene, xylene, ethylacetate, 4-methyl-2-pentanone, and propionic acid can be detected by measuring the SPR using angle modulation [95]. The SPR was measured using the Kretschmann configration, illustrated in Fig. 4.6, with a prism and a thin, highly conductive gold metal layer deposited on the prism base. The LED emitting 660 nm light was used as the light source to excite the SPR. The SPR reflection spectrum (reflected light intensity versus angle of incidence with respect to the normal of the metal/dielectric interface) was measured by coupling transverse magnetically polarized mono- chromatic light into the prism and measuring the reflected light intensity of the ray exiting the prism versus the angle incidence. In order to utilize this system as a gas sensor, a very thin film of methyl methacrylate, polyester resin, or propylene ether as the sensing membrane was deposited on gold metal thin film using a spin-coating method. The reflected light was measured using a CCD camera attached to a personal computer. The angle at which the minimum reflection intensity occurs is the resonance angle at which coupling of energy occurs between the incident light and the

Fig. 4.6 Kretschmann configuration of SPR apparatus used in toxic gas detection [29] 94 4 Introduction to Chemosensors

Fig. 4.7 Schematic configuration of the SPR sensor

surface plasmon waves. Four channel images of reflected light were observed by using the CCD camera. The schematic configuration of the SPR sensor is shown in Fig. 4.7. The SPR sensor with synthetic polymer thin film on the gold metal film as a sensing membrane exhibited high sensitivity for toxic gases such as ammonia, toluene, xylene, ethylacetate, 4-methyl-2-pentanone, and propionic acid.

4.7.2 Fluorescent Odor Sensors

Recently, a new sensing device has been developed that consists of an array of optically based chemosensors providing input to a pattern recognition system. This type of chemosensor consists of optical fibers deposited with fluorescent indicator Nile Red dye in polymer matrices of varying polarity, hydrophobicity, pore size, elasticity, and swelling tendency to create unique sensing regions that interact differently with vapor molecules [96]. Fiber-optic sensors most often consist of an analyte-sensing element deposited at the end of an optical fiber. Individual optical fibers with a diameters as small as 2 lm and imaging bundles with a diameter of 500 lm are available, enabling easy miniaturization, and are free from electrical interference. In a fiber-optic chemosensing system, the optical sensing element is typically composed of a reagent phase immobilized at the fiber tip by either physical entrapment or chemical binding. This reagent phase usually contains a chemical indicator that experiences some change in optical properties, such as intensity change, spectrum change, lifetime change, and wave- 4.7 Optical Odor Sensors 95

Fig. 4.8 (a) The most common configuration of an optical fiber chemosensor utilizing fluorescence, and (b) an example of the response length shift in fluorescence, upon interaction with analyte gases or vapors. The responses depend upon the nature of the organic vapor and the strength of its interaction with the different polymer systems used. The most common configuration of optical fiber chemosensor utilizing fluorescence and example of the response are shown in Fig. 4.8. The authors then analyzed the transient responses of the sensor array to distinguish different organic vapors such as odor samples a, b, and c. At present, the sensitivity of some types of optical chemosensor is not high (detection limits of several 1000 ppm) and there is little information about the lifetime, reproducibility or stability of the sensor system. Nevertheless, this is an interesting approach and one worthy of future work.

4.7.3 Other Optical Approaches

The use of a colorimeter coupled to optical fibers makes an inherently simple sensor [97], can be found in many forms, and was one of the earliest of the optical chemical sensor approaches. Color changes, or more generally, changes in absorption or emission of radiation, and polymer swelling by changes in refractive index of fiber coatings can be monitored optically. More recent approaches make e-noses from arrays of mi- crobeads on the end of a fiber [96, 98]. These systems can be made exquisitely sensitive with the appropriate chemistry on the fiber tip. The future of optical arrays within the e-nose are very promising. 96 4 Introduction to Chemosensors

4.8 Thermal (Calorimetric) Sensors

There are two sensor classes that are based on thermal technology. Those using pyroelectric [38] or thermopile sensors with coatings that absorb the analyte of interest. The underlying thermal sensor records the heat of solution of the analyte in the coating. They are quantitative because the more analyte that is absorbed, the more heat is generated. The theory and analytical performance of these sensors is similar to the coated SAW or chemiresistor polymer sensors, except that the underlying transducer is a heat sensor. The second class of thermal sensor is the Pellister, catalytic bead, or combustible gas sensor [99]. The catalytic sensor is typically a tiny bead of catalyst a millimeter or less in diameter that surrounds a coil of thin, 0.025 mm, Pt wire that acts as a Pt resistance thermometer. When resistively heated to about 500 8C, any contact with a hydrocarbon causes catalytic oxidation of the hydrocarbon with commensurate liberation of the heat of combustion. This heat is at the surface of the catalyst bead and some is lost to the surroundings while some is transferred to the tiny catalyst sensor bead. The heat transferred to the bead raises the temperature of the sensor, and it is this temperature change that is sensed as a change in resistance by the thin Pt wire. The sensor is typically placed in a Wheatstone Bridge circuit to measure the tiny changes in resistance of the Pt wire. The larger the resistance change, the higher the concentration of hydrocarbon. These sensors are typically used for combustible gases and were used in very early e-noses [100]. There are many formulations of the catalyst material and these sensors are operated at constant temperature or at constant voltage to serve different applications.

4.9 Amperometric Sensors

The amperometric gas sensor, or AGS, was one of the first sensors to be used in an e- nose format [100, 101, 103] and has been included in a heterogeneous sensor array- based instrument [132]. Amperometry is an old electroanalytical technique that encompasses coulometry, voltammetry, and constant potential techniques, and is widely used to identify and quantify electroactive species in liquid and gas phases. For liquid phase analytes, the electrodes and analytes are immersed in a common electrolyte and these have resulted in electronic tongues [102]. In contrast, application of amperometry to gas-phase analytes involves a unique gas-liquid/solid interfacial transport process. The AGS is a class of electrochemical gas sensors sometimes called voltammetric, micro-fuel cell, polarographic, amperostatic, or other names [103, 104]. The common characteristic of all AGSs is that measurements are made by recording the current in the electrochemical cell between the working and counter electrodes as a function of the analyte concentration. Figure 4.9 illustrates an amperometric sensor consisting of working, counter, and reference electrodes dipped in an electrolyte. The analyte is reacted electrochemically, i.e. oxidized or reduced, and this process, 4.8 Thermal (Calorimetric) Sensors 97

Fig. 4.9 An amperometric gas sensor

governed by Faraday’s Law, either produces or consumes electrons at the working electrode. The amperometric class of electrochemical sensor complements the other two classes of electrochemical sensors, i.e. potentiometric sensors that measure the Nernst potential at zero current, and conductometric sensors that measure changes in impedance [130]. The AGS, Figure 4.9, is controlled by a potentiostatic circuit and produces its current or signal when exposed to a gas/vapor containing an electroactive analyte. The analyte diffuses into the electrochemical cell and to the working electrode surface and where it participates in a redox reaction. The cell current is directly related to the rate of reaction taking place at the electrode surface and is described by application of Faraday’s Law, relating the mass, W, of a substance of molecular mass M (grams mol1) as:

QM W ¼ ð6Þ Fn where Q is the charge per unit electrode area, F is Faraday’s constant in coulombs/ equivalent, and n is the number of electron equivalents per mole of the reacting analyte. Assuming there are no other reacting species in the solution, the observed current, dQ/dt (t ¼ time) or i, is directly proportional to the amount of analyte, W, that is supplied to the working electrode and, this in turn can be related to the gaseous analyte concentration (see Eq. 7). The potentiostat allows control of the working electrode thermodynamic potential while the reaction occurs. The AGS is made reactive towards a variety of analytes by choosing different potentials, working electrode catalysts, electrolytes, porous membranes, and different electroanalytical methods. The working electrode reaction that produces current in the example of a CO sensor in Fig. 4.9 is usually taken as:

þ CO½gþH2O ¼ CO2 þ 2H ½aqþ2e :

The CO diffuses or is pumped to the region of the working electrode, dissolves in the electrolyte, diffuses to the working electrode surface where it undergoes reaction with subsequent desorption of the CO2 product and conduction of the 2e away through the metal electrode. The more CO that is present, the larger the current. Typical currents are in the micro- or pico-ampere level for ppm level reactants. Response times, mea- 98 4 Introduction to Chemosensors

sured as time to 90% of signal, have ranged from milliseconds for some oxygen sensors to several minutes for other analytes. It is usually preferable that a sensor works in the limiting current region in which the magnitude of the sensor signal is practically independent of the electrode potential. In theory, the limiting current region can be achieved in any case when the rate-limiting step is a step prior to electron transfer. The rate of electrode reaction may be limited by the rate of diffusion through a membrane or a capillary that is placed somewhere between the gas stream containing the analyte and the catalyst layer of the electrode. In

such cases, the limiting current, ilim, can be written:

ilim ¼ k½COgas ð7Þ

where the constant k is the proportionality constant relating the gaseous concentration to the current in some convenient units like lA (ppmv)1 (parts per million by volume). The amperometric gas sensor is one of the most widely used sensors for toxic

gas detection, i.e. CO, NO, NO2,H2S, SO2,O2, and so on. The AGS was used in the e- nose [105] for one of the earliest determinations of bacterial contamination [106] and identification of discrete analytes [107]. The AGS has been microfabricated [99, 108] but such versions are not yet commercially available. The main advantages of the amperometric approach are high sensitivity, a good deal of control over selectivity accompanied by relatively low cost, small size, and long stable lifetimes.

4.10 Summary of Chemical Sensors

Commercially available-nose instruments listed in Table 4.7 are concentrated on two main types of chemosensors, such as MOS-type and CP-type. More recent work is beginning to exploit other sensors for application to the food and drink industries as listed in Table 4.8. There are a number of books and references in other sections of this Handbook that point the user towards the myriad of e-noses that have been constructed as well as the various classes and types of sensors. New sensors, including micro instruments, will also contribute to the growing number of e-noses that will inevitably lead to an improvement in analytical capability. More and more is being demanded of sensors as time goes on. Quantitative and qualitative analytical results are not enough and we are requested to answer more pertinent and complex questions such as: Where is the contamination? Is this hazardous? Is this pure or the same as something else? These questions are often complex chemically. Sensors provide critical data for the e-nose and other analytical instruments that can address such complicated analytical tasks. Without good performance we have no chance for good data or good answers to these types of questions. Sensors and sensory data must therefore continue to be improved. Tab. 4.7 Commercially available e-nose instruments. Abbreviations: metal oxide Pattern recognition: artificial neural network, ANN; distance classifiers, DC; semiconductor, MOS; organic conducting polymer, CP; quartz crystal microbalance, principal component analysis, PCA; statistical pattern recognition, SPR; discriminant QCM; surface acoustic wave, SAW; gas chromatography, GC; quadrupole mass function analysis, DFA; cluster analysis, CA; and principal components regression, spectrometry, QMS; infrared, IR; and MOS field effect transistor, MOSFET. PCR.

Manufacturer Chemosensor type Number of sensors Size of Instrument (Cost US$) Pattern recognition Comments

Airsens analysis GmbH MOS 10 Laptop (20 000 –43 000) ANN, DC, PCA, SPR Small, fast & robust (Germany) Alpha MOS-Multi CP, MOS, QCM, SAW 6–24 Desktop (20 000 –100 000) ANN, DFA, PCA Autosampler and air Organoleptic Systems (France) conditioning unit available AromaScan PLC (UK) CP 32 Desktop (20 000 –75 000) ANN Autosampler and air conditioning unit available Array Tech QCM 8 Bloodhound Sensors CP 14 Laptop ANN, CA, PCA Small company, instrument Ltd. (UK) based on research at Lees University Cyrano Science Inc. (USA) CP 32 Palmtop (5000) PCA EEV Ltd. Chemical Sensor CP, MOS, QCM, SAW 8–28 Desktop ANN,DFA, PCA System (UK) Electronic Sensor GC, SAW 1 Desktop (19 500 –25 000) SPR

Technology Inc. (USA) Sensors Chemical of Summary 4.10 Hewlett-Pakard Co. (USA) QMS – Desktop (79 900) Standard chemometrix HKR-Sensorsysteme GmbH QCM 6 Desktop ANN, CA, DFA, PCA Small company. Based on re- (Germany) search at University of Munich Lennartz Electronic GmbH MOS, QCM 16 –40 Desktop (55 000) ANN, PCA, PCR MOSES II (Germany) Mastiff Electronic CP 16 Sniffed palms for personal Systems Ltd. identification Nordic Senser IR, MOS, MOSFET, QCM 22 Laptop (40 000 –60 000) ANN, CPA Identification of purity, Technologies AB (Sweden) origin. RST Rostock Raum-fahrt MOS, QCM, SAW 6–10 Desktop (50 000) ANN, PCA und Umweltschatz GmbH (Germany) 99 Neotronics Science Ltd. (UK) CP 12 — — Medium size d company. Shimadzu Co. (Japan) MOS 6 Desktop (70 000) PCA Large company. Sawtek Inc. SAW 2 Palmtop (5000) Proprietary 100 4 Introduction to Chemosensors

Tab. 4.8 Chemosensors used in recent e-nose studies for application to food and drink industries.

Food or Drink Test Chemosensor type Number of Ref. sensors

Alcohols Identification MOS (SnO2)1267

Fish (cod, haddock) Freshness MOS (SnO2)6 83

Fish (squid) Freshness MOS (MgO-In2O3)9 85

Coffee Discrimination MOS (SnO2)1266

Fish Freshness MOS (Ru-In2O3) 1 43, 86

Soup Quality control MOS (Ru-WO3)4 87 Sea foods (squid, oyster, Freshness MOS (Al-ZnO) 1 88–90 sea bream, sardine)

Alcohol Freshness MOS (ZnO-SnO2)1 91 Ground pork/Beef Discrimination and Mixed 15 61 effect of ageing

Wine Varieties and vintages MOS (SnO2,WO3)4 92 of same wine

Beef Freshness MOS (WO3-ZnO) 1 93 Fish (trout) Freshness MOS 8 94 Wheats Grade quality MOS, AGS 4 Â 4 95, [106] Wheats and cheese Discrimination and CP 20 96 ageing Cheeses Maturity of cheddars CP 20 97 Coffees Discrimination CP 12 98 between varieties Beers Diacetyl taint in CP 12 99 synthetic beer Beers Discrimination CP 12 100 between lager and ales Liqors Discrimination CP 5 101 between brandy, gin and whisky Boar Taints in meat MOS 14 102 Sausage meats Discrimination MOS 6 103 Water Taints in drinking MOS 4 104 water Colas Discrimination MOS 6 103 between diet and normal colas Coffees Discriminate MOS 6 80, 105 C. arabica and C. robusta Food flavors Flavor identification QCM 8 46 (orange, strawberry, apple, grape, peach) Tomatoes Effect of irradiation Mixed 7 106 and stress Whiskies Discrimination QCM 8 51 of Japanese whiskies 4.10 Summary of Chemical Sensors 101

References

1 P. Hauptmann. Sensors-Principles and Ap- 19 M. Miyayama. Tech. Digest of The 15th Sensor plications, (Carl Hanser Verlag & Prentice Symp. (Japan), 1997, 229. Hall), 1993, 115–153. 20 Y. Takao, Y. Miya, Y. Tachiyama, 2 J. W. Gardner. Microsensors-Principles and Y. Shimizu, M. Egashira. Denki Kagaku, Application,s (John Wiley & Sons), 1994, 1990, 58, 1162. 224–246. 21 P. T. Moseley, A. M. Stoneham, 3 J. W. Gardner, P. N. Bartlett. E-noses- D. E. Williams. Techniques and Mechanisms Principles and Applications, (Oxford), 1999, in Gas Sensing Eds. P. T. Mosely, 67–100. J. O. W. Norries and D. E. Williams, 4 K. Toko. Biomimetic Sensor Technology, (Adams Higer, Bristol), 1991, 248–267. (Cambridge University Press), 2000, 22 P. T. Moseley. Sens. Actuators B, 1991,3, 92–111. 167. 5 Sensors: A Comprehensive Survey, Eds. 23 A. Dall’Olio, G. Dascola, V. Varacca, W. Go¨pel, J. Hesse and J. N. Zemel. (VCH V. Bocchi. C. R. Se´ances Acad. Sci. Paris Ser. Verlagsgesellschaft GmbH, Weinheim, C, 1968, 267, 433–435. Germany), 1991. Book series.and Sensors 24 G. Bidan. Sens. Actuators B, 1992, 6, 45–56. Update book volume series, baltes, et al., 25 J. J. Miasik, A. Hooper, B. C. Tofield. eds. VCH Verlagsgesellschaft GmbH, J. Chem. Soc., Faraday Trans., 1986, I82, Weinheim, Germany. 1997-present. 1117-1126. 6 D. W. Ballantine, R. M. White, S. J. Martin, 26 G. Gustafsson, I. Lundstrom. Synth. Met., A. J. Ricco, G. C. Frye, E. T. Zellers, 1987, 2721, 203–208. H.Wohltjen. Acoustic Wave Sensors, Theory 27 G. Gustafsson, I. Lundstrom, B. Liedberg, Design and Physico-Chemical Applications, C. R. Wu, O. Inganas. Syth. Met., 1989, 31, (Academic Press, NY) 1997. 163–179. 7 S. Pezzini, A. Gaiambitto, A. Riva, 28 P. N. Batlett, P. B. M. Archer, S. K. Ling- J. L. Gurnani, C. M. Mari. In: High Tech. Chung. Sens. Actuators, 1989, 19, 125–140. Ceramics, Ed. P.Y.Vinecenzini, (Elsevier, 29 P. N. Bartlett, S. K. Ling-Chung. Sens. Amsterdam), 1987, 2155. Actuators, 1989, 19, 141–150. 8 H. Nanto, T. Minami, S. Takata. J. Appl. 30 P. N. Bartlett, S. K. Ling-Chung. Sens. Phys., 1986, 61, 482. Actuators, 1989, 20, 287–292. 9 H. Nanto, S. Tsubakino, T. Kawai, 31 J. W. Gardner, P. N. Bartlett. Nano- M. Ikeda, S. Kitagawa, M. Habara. J. Mater. technology, 1991, 2, 19–33. Sci., 1994, 29, 6529. 32 (a) G. J. Maclay, C. Yue, M. W. Findlay, 10 B. Bott, T. A. Jones, B. Mann. Sens. J. R. Stetter. Appl. Occupational and Actuators, 1984, 5, 65. Env. Hygiene, 1991, 6(8), 677–682. 11 M. Shiratori, M. Katsura, H. Okuma. Proc. (b) U.S. Patent 4,847,594; July 11, 1989. 1st Sensor Symp. (Japan), 1981, 69. Sensor for Detecting the Exhaustion 12 P. J. Shaver. Appl. Phys. Lett., 1967, 11, 255. of an Adsorbent Bed. 13 V. Lantto, P. Romppainen, S. Leppavuori. 33 J. R. Stetter, S. Zaromb, M. W. Findlay Jr. Sens. Actuators, 1988, 15, 347. Sens. Actuators A, 1984, 6, 269–288. 14 M. Egashira, Y. Shimizu, Y. Takao. Sens. 34 M. C. Burl, B. J. Doleman, A. Schaffer, Actuators B, 1990, 1, 108. N. S. Lewis. Sens. Actuators B, 2001, 72(2): 15 Y. Nakatani, M. Sakai, M. Matsuoka. Jpn. 149–159. J. Appl. Phys., 1983, 22, 912. 35 A. Koll, S. Kawahito, F. Mayer, 16 Y. Nakatani, M. Matsuoka, Y. Iida. IEEE C. Hagleitner, D. Scheiwiller, O. Brand, Trans. Components Hybrid Manufact. H. Baltes. Proc. SPIE, 1998, 3328, 223–232. Technol., 1982, CHMT-5, 522. 36 I. Lundstrom, S. Shivaraman, C. Svensson, 17 Z. Szklarki, B. Zakrzewski, M. Rekas. L. Lundkuist. Appl. Phys. Lett., 1975, 26, Thin Solid Films, 1989, 174, 269. 55–57. 18 P. T. Mosely, D. E. Williams, L. 37 I. Lundstrom, E. Hedborg, A. Spetz, O. W. Norries. Sens. Actuators, 1988, 14, H. Sundgren, F. Winquist. Sensors and 79. Sensory Systems for an E-nose, Eds. 102 4 Introduction to Chemosensors

J. W. Gardner and P. N. Bartlett, NATOASI 57 Y. Okahata, G. En-na, H. Ebata. Anal. Series (Kluwer, Dordrecht), 1992, 212, Chem., 1989, 62, 1431–1438. 303–319. 58 M. Ohnishi, T. Ishibashi, Y. Kijima, 38 J. N. Zemel. Sens. Actuators A, 1996,56 C. Ishimoto, J. Seto. Sens. Mater., 1992,1, (1–2) 57–62. 53–60. 39 G. Z.Sauerbrey. Z, Phys., 1959, 155, 59 S. M. Chang, E. Tamiya, I. Karube, M. Sato, 206–222. Y. Masuda. Sens. Actuators B, 1991,5, 40 H. Nanto, N. Dougami, T. Mukai, 53–58. M. Habara, E. Kusano, A. Kinbara, 60 D. S. Ballantine, S. L. Rose-Pehrsson, T. Ogawa, T. Oyabu. Sens. Actuators B, J. W. Grate, H. Wohltjen. Anal. Chem., 2000, 66, 16–18. 1986, 58, 3058–3066. 41 H. Nanto, Y. Yokoi, T. Mukai, J. Fijioka, 61 S. L. Rose-Pehrsson, J. W. Grate, E. Kusano, A. Kinbara, Y. Douguchi. Mater. D. S. Ballantine, P. C. Jurs. Anal. Chem., Sci. Eng., 2000, 12, 43–48. 1988, 60, 2801–2811. 42 S. Kurosawa, N. Kamo, D. Matsui, 62 S. L. Rose-Pehrsson, J. W. Grate. SPIE Y. Kobatake. Anal. Chem., 1990, 62, 353. Proc., 1993, 299–311. 43 D. B.Radloff, S. Kurosawa, K. Hirayama, 63 F. Winquist, E. G. Hornsten, H. Sundgren, T. Arimura, K. Otake, A. Sekiya, I. Lundstrom. Meas. Sci. Technol., 1993,4, N. Minoura, M. Rapp, Hans-J. Ache. Mol. 1493–1500. Cryst. Liq. Cryst., 1997, 295, 141. 64 I. Lundstrom, R. Erlandsson, U. Frykman, 44 H. Nanto, Y. Yokoi, Y. Hamaguchi, E. Hedborg, A .Setz, H. Sundgren. Nature, S. Kurosawa, T. Oyabu, E. Kusano, 1991, 352, 47–50. A. Kinbara. Technical Report of IEICE, 2000, 65 K. C. Persaud, G. H.Dodd. Nature, 1982, OME 2000-95, 39. 299, 352–355. 45 T. Matsumoto, K. Tanabe, S. Kurosawa, 66 H. V. Shurmer, J. W. Gardner, H. T. Chan. T. Mukai, H. Nanto. Chem. Software, 2000, Sens. Actuators, 1989, 18, 361–371. 22, 85. 67 H. V. Shurmer, J. W. Gardner, P. Corcoran. 46 S. Kurosawa, S. Higashi, H. Aizawa, Sens. Actuators B, 1990, 1, 256–260. Dae-Sang.Han. M. Yoshimoto. Chem. 68 J. W. Gardner, H. V. Shurmer, T. T. Tan. Sensors, 2000, 16, 37. (in Japanese) Sens. Actuators B, 1992, 6, 71. 47 S. Kurosawa, D. Tachiyuki, Das-Sang.Han, 69 J. W. Gardner. Sens. Actuators B, 1991,4, H. Aizawa, M. Yoshimoto. Chem. Sensors, 109–115. 2000, 16, 103. (in Japanese) 70 J. W. Gardner, E. L. Hines, M. Wilkinson. 48 T. Nakamoto, A. Fukuda and T. Moriizumi. Meas. Sci. Tech., 1990, 1, 446–451. Sens. Actuators B, 1993, 10, 85–91. 71 J. W. Gardner, E. L. Hines, H. C. Tang. 49 Y. Sakuraba, T. Nakamoto, T. Moriizumi. Sens. Actuators B, 1992, 9, 9–15. Trans. Inst. Electron. Comm. Eng., 1990, 72 A. D. Walmsley, S. J. Haswell, E. Metcalfe. J73D-II, 1863–1871. Anal. Chem., 1991, 250, 257–264. 50 J. Ede, T. Nakamoto, T. Moriizumi. Sens. 73 P. Corocoran, P. Lowery. Proc. of the 4th Actuators B, 1993, 13–14, 351–354. Inter. Conf. On Artificial Neural Networks, 51 K. Ema, M. Yokoyama, T. Nakamoto, 1995, 415–420. T. Moriizumi. Sens. Actuators, 1989, 18, 74 B. S. Hoffheins, R. J. Lauf. Sensor Expo 291–296. Proceedings, 1988, 205, 1–7. 52 T. Nakamoto, A. Fukuda, T. Moriizumi, 75 X. Wang, S. Yee, P. Carey. Sens. Actuators B, Y. Asakura. Sens. Actuators B, 1991,3, 1993, 13-14, 458–461. 221–226. 76 X. Wang, J. Fang, P. Carey, S. Yee. Sens. 53 T. Nakamoto, A. Fukuda, T. Moriizumi. Actuators B, 1993, 13–14, 455–477. Sens. Actuators B, 1990, 1, 473–476. 77 A. Ikegami, M. Kaneyasu. Proc. of Inter. 54 H. Muramatsu, E. Tamiya, I. Karube. Anal. Conf. on Solid State Sensors and Actuators, Chem., 1990, 63, 399–408. 1985, 136–139. 55 K. Yokoyama, F. Ebisawa. Anal. Chem., 78 M. Kaneyasu, A. Ikegami, H. Arima, 1993, 65, 673–677. S. Iwanaga. IEEE Comp., 1987, CHMT-10, 56 Y. Okahata, O. Shimizu. Langmuir, 1987,3, 267–273. 1171–1172. 4.10 Summary of Chemical Sensors 103

79 H. Abe, T. Yoshimura, S. Kanaya, 99 W. J. Buttner, J. R. Stetter, G. J. Maclay. Y. Takahashi, Y. Miyashita, S. I. Sasaki. Sens. Mater., 1990, 2, 99–106. Anal. Chem., 1987, 194, 1–9. 100 (a) J. R. Stetter, S. Zaromb, M. W. Findlay. 80 H. Abe, S. Kanaya, Y. Takahashi, U.S. Patent 5055266, 1991. (b) J.R.Stetter, S. I. Sasaki. Anal. Chem., 1988, 215, S.Zaromb, W.R.Penrose, U.S. Patent 155–168. 4670405, 1987. (c) J.R.Stetter, Chemical 81 T. Aishima. J. Agr. Food, 1991, 39, Sensor Arrays: Practical Insights and 752–758. Examples, in Sensors and Sensory Systems 82 T. Aishima. Anal. Chem., 1991, 243, for an E-nose, Eds. J.Gardner and 293–300. P.N.Bartlett, (Kluwer Academic Publis- 83 T. Nakamoto, H. Takagi, S. Usami, hers). 1992, 273–301. T. Moriizumi. Sens. Actuators B, 1992,8, 101 J. R. Stetter. J. Colloid Int. Sci., 1978, 65(3), 181–186. 432–443. 84 F. A. M. Davide, C. Di.Natale, A. D’Amico. 102 F. P. Winquist, P. Wide, I. Lundstro¨m. Sens. Actuators B, 1994, 18-19, 244–258. Anal. Chim. Acta., 1997, 357, 21–31. 85 R. Olafsson, E. Martindotti, G. Olafsdotti, 103 S. C. Chang, J. R. Stetter, C. S. Cha. O. I. Sigfusson, J. W. Gardner. Sensors and Talanta, 1993, 40(4), 461–467. Sensory Systems for an E-nose, NATO 104 Z. Cao, W. J. Buttner, J. R. Stetter. ASI Series E, Ed. J. W. Gardner and Electroanalysis, 1992, 4, 253–266. P. N. Bartlett, (Kluwer Academic 105 Artificial Chemical Sensing: Proceedings of the Publishers, Dordrecht), 1992, 257–272. Eighth International Symposium on Olfaction 86 Yea, R. Konishi, T. Osaki, K. Sugahara. and the E-nose (ISOEN 2001), March 26-28, Sens. Actuators A, 1994, 45, 159–165. 2001, Washington DC., Eds. J. R. Stetter, 87 C. G. Fox, J. F. Alder. Techniques and W. R. Penrose, (The Electrochemical mechanisms in gas sensing, Eds. P.T.Mosely, Society, Pennington, NJ), 2001. I.O.W.Norries and D.E.Williams, (Adam 106 J. R. Stetter, M. W. Findlar, K. M. Hilger, Bristol), 1991, 324–346. Schroeder, C. Yue, W. R. Penrose. Anal. 88 J. W. Grate, S. J. Martin, R. M. White. Anal. Chim. Acta., 1993, 284, 1. Chem., 1993, 65, 940–948. 107 J. R. Stetter, P. C. Jurs, S. L. Rose. Anal. 89 J. W. Grate, S. J. Martin, R. M. White. Anal. Chem., 1986, 58, 860–866. Chem., 1993, 65, 987–996. 108 Buttner, J. William, G. J. Maclay, 90 B. Liedberg, C. Nylander, I. Lundstrom. J. R. Stetter. Sens. Actuators B, 1990,1, Sens. Actuators, 1983, 4, 299–302. 303–307. 91 E. Kretschmann. Z. Phys., 1971, 241, 313. 109 Y. Takao, Y. Shimizu, M. Egashira. Sens. 92 K. S. Johnston, S. R. Karlson. C. Jung, Mater., 1992, 3, 249. S. S. Yee. Mater. Chem. Phys., 1995, 42, 242. 110 Y. Shimizu, Y. Takao, M. Egashira. 93 B. Chadwick, M. Gal. Jpn. J. Appl. Phys., J. Electrochem. Soc., 1988, 135, 2539. 1993, 32, 2716. 111 M. Egashira, Y. Shimizu, Y. Takao. Sens. 94 S. Nelson, K. S. Johnston, S. S. Yee. Sens. Actuators B, 1993, 13–14, 443. Actuators B, 1996, 35/36, 187. 112 H. Nanto, H. Sokooshi, T. Usuda. Sens. 95 H. Nanto, M. Habara, N. Dougami, Actuators B, 1993, 10, 79. T. Mukai, H. Sugiyama, E. Kusano, 113 H. Nanto, H. Sokooshi, T. Kawai, T. Usuda. A. Kinbara, Y. Douguchi. Tech. Digest J. Mater. Sci. Lett., 1992, 11, 235. of the 7th Inter. Meeting on Chemical Sensors, 114 H. Nanto, H. Sokooshi, T. Kawai. Sens. 1998, 695–697. Actuators B, 1993, 13–14, 175. 96 J. White, J. S. Kauer, T. A. Dikkinson, 115 H. Nanto, T. Morita, M. Habara, K. Kondo, D. R. Walt. Anal. Chem., 1996, 2191–2202. Y. Douguchi, T. Minami. Sens. Actuators B, 97 D. S. Ballantine Jr., D. Callahan, 1996, 35–35, 384. G. J. Maclay, J. R. Stetter. Talanta, 1992, 116 C. D. Natale et al.. Sens. Actuators B, 1995, 39(12), 1657–1667. 24–25, 801. 98 K. J. Albert, D. R. Walt, D. S. Gill, 117 H. Miura et al.. IEE of Japan, 1996, E117, T. C. Pearce. Anal. Chem., 2001, 73(11), 306. 2501–2508. 118 M. S. Berberich, S. Vaihinger, W. Go¨pel. Sens. Actuators B, 1994, 18–19, 282. 104 4 Introduction to Chemosensors

119 A. Pisanelli, A. A. Qutob, P. Travers, 127 A. A. Fekada, E.L. Hines, J. W. Gardner. S. Szyszko, K. C.Persaud. Life Chem. In: Artificial Neural Networks and Genetic Reports, 1994, 11, 303. Algorithms, Eds. R.A.Albrecht, C.R.Reeves 120 K. C. Persaud, P. J. Travers. Handbook of and N.C.Steele, (Springer-Verlag, biosensors and e-noses, Ed. E. K. Rogers, New York), 1993, 691–698. (CRC Press Inc., Ohio), 1997, 52–59. 128 T. Aishima. ASIC 14th Colloque 121 J. W. Gardner. P. N. Bartlett. Proc. of (San Francisco), 1991. Olfaction and Taste XI, (Springer Verlag), 129 F. Winquist et al.. Proc. of 8th Inter. Conf. 1994. on Solid State Sensors and Actuators, 1995. 122 J. W. Gardner, T. C. Pearce, S. Friel, 130 N. Barsan, M. Schweizer-Berberich, P. N. Bartlett, N. Blair. Sens. Actuators B, W. Go¨pel. Fresenius J. Anal. Chem., 1999, 1994, 18, 240. 365(4), 287–304. 123 T. C. Pearce, J. W. Gardner, S. Friel, 131 J. B.Miller. IEEE Sensors, 2001, 1, 88. P. N. Bartlett, N. Blair. Analyst, 1993, 118, 132 J. R. Stetter, S. Zaromb, W. R. Penrose, 371. M. W. Findlay Jr., T. Otagawa, A. J. Sincali. 124 J. M. Slater, J. Paynter, E. J. Watt. Analyst, Portable device for detecting and identifying 1993, 118, 371. hazardous vapors, in: Proc. 1984 Hazardous 125 B. Bourrounet, T. Talou, A. Gaset. Sens. Material Spills Conference, April 9–12, Actuators B, 1995, 26–27, 250. Nashville, TN, 1984, 183–194. 126 T. Tan, Q. Lucas, L. Moy, J. W.Gardner, P. N. Bartlett. LC-GC International, 1995,8, 218. 105

5 Signal Conditioning and Preprocessing

R. Gutierrez-Osuna, H. Troy Nagle, B. Kermani, Susan S. Schiffman

5.1 Introduction

The topics covered in this chapter establish the connection between gas sensors and pattern recognition, the two fundamental modules of an odor-sensing instrument that are covered in Chapters 4 and 6, respectively. A number of electronic circuits are involved in integrating pattern analysis algorithms with the underlying chemical transduction mechanisms, as shown in Fig. 5.1. First, the response of the odor sensors (e.g., a resistance change) needs to be measured and converted into an electrical signal (e.g., a voltage). This operation is performed by means of interface circuits. Second, the electrical signal undergoes analog conditioning (e.g., filtering) to enhance its information content. Third, the analog signal is sampled, digitized and stored in computer memory (not covered in this chapter due to space constraints). Finally, the sampled signal is digitally preprocessed (e.g., autoscaling) in order to make it suitable for pattern analysis. This chapter is organized in three basic parts: interface circuits, signal conditioning, and preprocessing. Section 5.2 presents the fundamental interface circuits for the three primary odor sensor types: resistive, piezoelectric, and field-effect. Section 5.3 reviews the primary functions performed by analog signal conditioning circuits. Sec- tion 5.4 covers data preprocessing – the first stage of digital signal processing. The issue of sensor and instrumentation noise, one of the most important factors determining electronic-nose performance, is also reviewed in Section 5.5. The chapter con-

Fig. 5.1 Organization of this chapter

cludes with a review of current instrumentation trends aimed at increasing the selectivity of odor sensor systems.

5.2 Interface Circuits

Sensor interface circuits constitute the first stage of electronic instrumentation. The purpose of these circuits is to generate an electrical signal that reflects changes in the sensors. Since interface circuits are tightly coupled to the underlying sensing technology, we will focus our presentation on three widely used odor sensors: conductivity (metal-oxide and conductive-polymer chemoresistors), piezo-electric (surface acoustic wave and quartz crystal microbalance), and field effects (metal-oxide field-effect transistors). In addition, this section reviews the issue of operating temperature control, essential for the operation of metal-oxide transducers.

5.2.1 Chemoresistors

In chemoresistive sensors the presence of volatile compounds changes the conductance (or resistance) of the sensing membrane. Interface circuits for these sensors are, therefore, relatively simple since they only involve measuring resistance changes. Two types of resistance measurement circuits are commonly used: voltage dividers and Wheatstone bridges. These circuits are presented and analyzed in the following subsections. Linear versions of these circuits that involve operational amplifiers are presented in section 5.3.5 as a special type of analog signal conditioning. Finally, AC impedance measurement techniques for chemoresistors are briefly reviewed at the end of this section.

5.2.1.1 Voltage Dividers The standard method for measuring large resistance changes is a voltage divider, as shown in Fig. 5.2a. This instrumentation circuit is very popular due to its simplicity.

The resistive sensor is placed in series with a load resistor RL and connected to a voltage reference VCC. The current through the sensitive element and load resistance becomes:

VCC IS ¼ ð1Þ RS þ RL

Changes in sensor resistance are then measured as voltage changes across the sensor

(VS) or the load resistor (VL). For convenience, we will use the voltage across the load resistor since it is a single-ended measurement and the subsequent derivation becomes simpler. Using Ohm’s Law (V ¼ IR), the resulting output voltage becomes: 5.2 Interface Circuits 107

VCC VL ¼ ISRL ¼ RL ð2Þ RS þ RL

The value of the load resistor should be selected to maximize the sensitivity of the circuit, that is, the slope of the VL RS curve, which can be calculated as: @V @ R R S ¼ L ¼ L V ¼ V L ð3Þ @ @ CC CC 2 RS RS RS þ RL ðRS þ RLÞ

The maximum of the selectivity is finally determined by finding the zeros of its partial derivative with respect to RL: ! dS @ R ¼ L V ¼ 0 ð4Þ @ 2 CC dRL RL ðRS þ RLÞ

It can be shown that the optimal load resistor is RL ¼ RS, this is the sensor resistance at the operating point, typically defined by a reference gas (e.g., clean air). The voltage divider is the circuit recommended by several metal-oxide sensor manufacturers [1, 2] but it has several shortcomings. First, the relationship between the sensor resistance

RS and the output voltage VL is nonlinear since the current IS through the sensor depends not only on the load resistor but also on the sensor resistance (refer to section 5.3.5.1 for linearization circuits). Second, and more importantly, the circuit is only appropriate for measuring large resistance changes, such as those typical of metal- oxide sensors. Conducting polymer chemoresistors have sensitivities one order of magnitude lower [3] and require the use of Wheatstone bridges.

Fig. 5.2 (a) Voltage divider and (b) Wheatstone bridge for resistive sensors. (c–d) Sensitivity improvements with a gain stage 108 5 Signal Conditioning and Preprocessing

5.2.1.2 The Wheatstone Bridge When the resistance changes to be measured are small relative to the baseline resistance of the sensor, the information in the output voltage will consist of small fluctuations superimposed on a large offset voltage. Although the sensitivity can be boosted with a gain stage, the problem remains since a large portion of the dynamic range of the ADC will be ‘wasted’ in measuring the offset voltage. One solution for measuring small resistance changes is to subtract the offset voltage with a second voltage divider, as shown in Fig. 5.2b. The differential voltage in the bridge is: VCC VCC RL R2 VOut ¼ RLIS R2I2 ¼ RL R2 ¼ VCC ð5Þ RS þ RL R1 þ R2 RS þ RL R1 þ R2

As in the voltage divider of Fig. 5.2a, sometimes called a half-bridge circuit, the ma-

ximum sensitivity for the Wheatstone bridge is obtained by choosing resistors R1, R2and RL equal to the sensor baseline resistance. This measurement approach is known as a deflection method, because the sensor response is measured as a differential voltage when the bridge becomes unbalanced. An alternative approach, known

as the null method, consists of adjusting the resistors R1, and R2 to cancel the differential voltage VOUT . The sensor resistance is then obtained from the balance condition:

R1 RS R1 VOUT ¼ 0 $ ¼ ! RS ¼ RL ð6Þ R2 RL R2

By comparing Eqs. (5) and (6) it can be inferred that, unlike deflection measurements, the null method is insensitive to fluctuations in the supply voltage. The deflection method, on the other hand, is easier to implement and yields faster responses, making it more appealing for dynamic measurements. It must be noted that the Wheatstone bridge (deflection-method) has the same sensitivity as a voltage divider. Notice that the only difference between Eqs. (2) and (5) is

the offset voltage provided by the R1 R2 arm, which does not depend on the sensor resistance. The main advantage of the Wheatstone bridge is that it affords higher amplification gains since the offset voltage has already been removed. To illustrate this point, assume a gas sensor that has a resistance that decreases in the presence

of an odor, RS ¼ R0ð1 aÞ. Figure 5.2c shows the response of both circuits for Ø= a 1=3, R1 ¼ R2 ¼ RL ¼ R0, and VCC ¼ 10V. If this signal is to be captured with a data acquisition system that has a dynamic range of 0 V to 10 V, the maximum gain that can be applied to the voltage divider is only 5/3. Although the Wheatstone bridge has the same initial sensitivity (slope), removal of the baseline offset allows a maximum gain of 10, as shown in Fig. 5.2d. The figure also illustrates the nonlinearity introduced by the deflection measurements. It is important to mention that voltage dividers and Wheatstone bridges can be used

to remove common-mode effects by replacing the load resistor RL with a reference sensor that is shielded from the variable being sensed by the primary sensor but un- shielded from environmental conditions. This approach is widely employed in strain gages to compensate for temperature interference, and in pellistors for both tempera- 5.2 Interface Circuits 109 ture and humidity compensation [4]. The linearized voltage dividers covered in section 5.3.5.1 are also commonly used for compensation purposes. These types of measurements, based on the ratio between a primary sensor and a reference sensor, are known as ratiometric techniques [5].

5.2.1.3 AC Impedance Spectroscopy Impedance spectroscopic techniques are commonly used to determine the contribution of the different structures in a device (e.g., surface, bulk, grain, and contacts). Impedance spectroscopy is performed by applying a small-amplitude AC voltage to the sensor and measuring the resulting current. By sweeping the frequency of the AC signal and measuring the impedance at multiple frequencies, an equivalent electrical model can be derived that reveals the contributions of each structure for different gases [6, 7]. Im- pedance spectroscopy requires specialized (and expensive) test and measurement equipment such as impedance analyzers or frequency response analyzers. Several studies have proposed the use of impedance spectroscopy to improve the selectivity of chemoresistors. Weimar and Go¨pel [8] have employed two-point measurements at frequencies between 1 Hz and 1 MHz to extract the complex impedance of a custom tin-oxide sensor with interdigitated electrodes. Figure 5.3a shows the Cole-

Fig. 5.3 (a) Cole-Cole impedance plot and equi-

valent circuit for an interdigitated SnO2 sensor [8].

(b) CO and NO2 sensitivity versus frequency of a

SnO2 sensor [9]. (c) Dissipation factor versus frequency response of a conducting polymer sensor [10] 110 5 Signal Conditioning and Preprocessing

Cole impedance plot of a sensor exposed to pure carrier gas, before and after the addi-

tion of 10 000 ppm H2. The parameters of the equivalent electrical circuit shown in the upper right corner of the figure were obtained by fitting the impedance model

R1 þ R2kC2 þ R3kQ3 (solid line) to the experimental data (dotted). The resistance R1 models contributions from the bulk and the surface of the tin oxide. Contribution from the SnO2/Pt contacts are modeled by only one parallel component (R2, C2) since the two-point setup cannot separate the impedance of the two electrodes. These contact contributions are responsible for the large semicircle in the figure. The third contri-

bution (R3, Q3), caused by migration of surface species along the grain boundaries at low frequencies, is responsible for the small semicircle in the impedance plot. This

contribution becomes inductive in the presence of H2 (notice that the small semicircle is mirrored with respect to the one for synthetic air). This study indicates that sensi-

tivity to CO, NO2, and H2 can be improved by measuring the AC impedance of the sensor at DC, 3 kHz, and 20 kHz, respectively. Qualitatively similar conclusions, shown in Fig. 5.3b have been reported [9]. Amrani et al. [10] have performed impedance spectroscopy at higher frequencies (100–1000 MHz) to characterize conducting polymer sensors. Their results, summarized in Fig. 5.3c, indicate that methanol, ethyl acetate, and acetone (with dipole moments of 1.69 lD, 1.78 lD and 2.88 lD, respectively) induce peaks in the dissipation factor (the ratio of resistance to reac-

tance, R/XC) at different frequencies, with the peak amplitude being a monotonically increasing function of the vapor concentration.

5.2.2 Acoustic Wave Sensors

Instrumentation electronics for acoustic wave gas sensors are more complex than those employed for chemoresistors, as they involve AC signals of high frequency (e.g., MHz range). According to the number of piezo-electric transducers used in the device, acoustic wave sensors can be classified into one-port and two-port devices:

* One-port devices consist of a single transducer that is used both as an input and as an output. The port is used to generate an acoustic signal, which is combined with the charges induced in the device to produce a measurable impedance change, or a shift in resonance frequency if using an oscillator circuit. A representative sensor for this type of device is the QMB, also known as a thickness-shear mode sensor. * Two-port devices, as the name indicates, have separate inputs and outputs. An input interdigitated transducer (IDT) is used to induce an acoustic signal, which propagates across the surface of the device. When the acoustic wave reaches the output transducer, an electrical signal is regenerated, and its phase and/or amplitude changes with respect to the input signal are used as measurement variables. A representative two-port device is the SAW delay line sensora.

a) One-port or resonant SAW sensor configurati- substrate to reflect the acoustic waves back ons are also employed. A single IDT is placed in to the IDT, creating a ‘resonant cavity’ in the the center of the device and mechanical ‘groo- center of the device [12]. ves’ are micro-fabricated on the edges of the 5.2 Interface Circuits 111

Fig. 5.4 Instrumentation configurations for acoustic wave sensors: (a) oscillator circuit, (b) impedance meter, and (c) network analyzer. (d) Dual delay SAW structure for temperature compensation [3, 11, 12]. (e) QMB sensor interface circuit [15]

Three instrumentation configurations, illustrated in Fig. 5.4, are commonly employed for acoustic wave sensors: oscillator circuits, vector voltmeters, and network analyzers. Oscillator circuits can be used for one-port (not shown in the figure) and two-port devices (Fig. 5.4a). The sensor is used as the resonant element in the feedback loop of an RF-amplifier circuit. Mass changes in the sensitive layer induce shifts in the resonance frequency, which are measured with a frequency counter. Oscillator circuits have several advantages, including low cost, relative simplicity, and excellent frequency stability [11]. However, these circuits generally provide information about wave velocity, and not amplitude, which may be necessary to monitor wave attenua- tions. A second configuration, shown in Fig. 5.4b, overcomes this limitation, providing both wave velocity and amplitude measurements in two-port devices. A signal generator is used to supply an RF voltage to the input transducer, and a vector voltmeter measures phase and amplitude changes at the output IDT relative to the input signal. Vector voltmeters are, however, relatively expensive pieces of laboratory equipment, and their phase measurements are 10–100 times less sensitive than frequency measurements with oscillator circuits. A third alternative, shown in Fig. 5.4c, is to use a network analyzer to perform a complete characterization of the device at multiple frequencies [11, 12]. To compensate for interferents (e.g., temperature, pressure, drift), SAW sensors are typically used in the dual configuration illustrated in Fig. 5.4d. One delay line is coated 112 5 Signal Conditioning and Preprocessing

with a sensing film that responds strongly to odors, and the second line is used as a reference to capture only interferent effects. Subtraction of the two signals yields a measurement that is, theoretically, independent of the common-mode interferents [13]. Fig. 5.4e shows a compact, low-power circuit for a QMB sensor [14, 15]. A 10 MHz sensor crystal is connected to an integrated oscillator whose output frequency decreases when odor molecules are absorbed into the crystal coating. The output of the sensor oscillator is compared to a reference oscillator with an uncoated 10 MHz crystal by means of a D flip-flop, which generates the difference frequency.

5.2.3 Field-Effect Gas Sensors

As described in Chapter 4, two configurations can be used in metal-insulator-semiconductor field-effect gas sensors: capacitor (MISCAP) and transistor (MISFET). The two structures depicted in Fig. 4.4 of Chapter 4 yield similar information, the differences being in the required measurement circuitsb. In the case of MISCAP sensors, changes in the voltage-capacitance curve can be measured with a small AC-voltage (e.g.,

1 MHz) superimposed on a DC-potential [16]. Changes in the ID VG curve of MIS- FET sensors, on the other hand, may be measured with constant-voltage [17] or constant-current circuits [18]. Figure 5.5 shows a conventional two-terminal arrangement

Fig. 5.5 MISFET gas sensors: (a) two-terminal configuration and (b) possible constant-current interface circuit [18, 19]

for an n-channel MISFET with a common gate-drain configuration, and a possible

constant-current interface circuit. The shift in the VGDS ID curve upon exposure to volatile organic compunds is the change in the threshold voltage, which is in turn related to the shift in work function, surface states, and charge. A current source

is used to inject a constant current into the drain, and the resulting voltage VGDS is buffered (see Section 5.3.2) and sampled to create a time-resolved signal. Field-effect sensors operate at high temperatures (100–200 8C for Si substrates, up to 700 8C for

b) MISCAPs have a simpler structure and are, therefore, often used for exploratory work [16] 5.2 Interface Circuits 113

SiC) and, like metal-oxide chemoresistors, require temperature control circuits. Field- effect sensors also suffer from baseline drift, which can be compensated for by using differential configurations having an active gate FET and a passive reference FET [16].

5.2.4 Temperature Control

Metal-oxide gas sensors are commonly operated in the so-called isothermal mode, in which the temperature of the sensor is kept constant during exposure to odorsc. The simplest and most widely used method for pseudo-isothermal operation consists of applying a constant voltage across the terminals of the resistive heater RH, as shown in Fig. 5.6a. Temperature stability is achieved by using heater materials with a positive temperature coefficientd so that the thermoresistive effect serves as negative feedback [20]. This simple constant-voltage operation may be used when temperature stability is not critical.

Fig. 5.6 (a) Constant heater voltage and (b) constant heater resistance circuits [20]

Improved stability (e.g., to gas-flow cooling effects) may be achieved by controlling the heater resistance rather than the heater voltage [21]. In constant-resistance operation, the sensor heater is placed in a Wheatstone bridge and compared against a reference potentiometer that determines a set-point resistance, as shown in Fig. 5.6b. Deviations from the set-point resistance result in a differential voltage across the bridge, which is used to control a current or voltage source. Capteur Ltd. implements constant-resistance control by using a FET operating as a voltage-controlled current source [22]. Constant resistance, however, requires heater materials with a reasonably high thermoresistive coefficient. A third alternative consists of embedding a temperature sensor in the substrate [8], or using the heater as a temperature sensor [24, 25]. The latter method, however, also

c) If the sensor is normally operated at low d) The heater resistance RH is a function of tem- temperature, it is then necessary to shift to a perature T: RH ¼ R0ð1 þ aTÞ, where R0 is the high temperature to burn off excess organic baseline resistance at zero degrees and a is the contaminants from the sensor surface [28]. temperature coefficient. For positive a,the heater resistance increases with temperature. 114 5 Signal Conditioning and Preprocessing

requires a large positive thermoresistive coefficient, which is not the case for certain commercial metal-oxide sensors [26]. Sensor surface temperatures can also be measured with infrared thermometers, but these measurements have been shown to be rather inaccurate [26]. Additional temperature control strategies may be found in the literature [27].

5.3 Signal Conditioning

The electrical signals generated by sensor interface circuits are often not adequate for acquisition into a computer, and must be further processed by a number of analog signal conditioning circuits. The four basic roles of these circuits: buffering, amplification, filtering, and special functions, are surveyed in the following subsections along with a brief review of operational amplifiers.

5.3.1 Operational Amplifiers

Operational amplifiers (op-amps) are analog integrated circuits widely used to implement a variety of instrumentation circuits. Although a thorough coverage of op-amps is beyond the scope of this chapter, we provide a brief review that will allow the reader to analyze the circuits presented in the remaining sections of this chapter. An op-amp, shown in Fig. 5.7a, is essentially a high-gain amplifier that generates an output voltage

V0 ¼ GOLVd proportional to the difference voltage Vd between a noninverting (þ) and an inverting input (). The power necessary to perform the signal amplification 4 6 (GOL ﬃ 10 10 ) is derived from the supply voltages (ÆVS) and, therefore, the output voltage V0 is constrained by VS V0 þVS. Op-amp circuits in this open-loop configuration are not practical since very small difference voltages Vd will drive the output voltage to saturation. In addition, the open-loop gain GOL has a limited bandwidth (GOL decays significantly with frequency), and is very sensitive to temperature and power supply fluctuations. For these reasons, op-amps circuits typically contain a feedback loop to control the gain, as shown in Fig. 5.7b. A large number of these op-amp feedback circuits can be analyzed by assuming ideal op-amp characteristics, primarily (1) infinite open-loop gain and bandwidth

GOLðf Þ¼1, (2) infinite input impedance ZIN ¼1, and (3) zero output impedance ZOUT ¼ 0. The latter simply implies that loading effects are negligible, that is, V0 ¼ VOUT in the equivalent op-amp circuit of Fig. 5.7a. These ideal characteristics lead to two ‘golden rules’ that are sufficient for analyzing many practical op-amp feedback circuits [23, 29]:

* Rule 1: Inputs stick together. Since the gain is infinite and VOUT must be bounded, the feedback network will enforce an output VOUT that cancels the differential voltage Vd ¼ 0. * Rule 2: Inputs draw no current. This follows from the assumption that ZIN ¼1. 5.3 Signal Conditioning 115

Fig. 5.7 (a) Op-amp simplified internal model and (b) analysis of feedback circuits. Amplifier circuits: (c) buffer, (d) inverting amplifier, (e) difference amplifier, and (f) instrumentation amplifier

To illustrate the use of these rules, we derive the transfer function of the circuit shown in Fig. 5.7b. From Rule 1 we can establish that the voltage at the noninverting input is equal to the input voltage VIN. This allows us to express the current i1 flowing through resistor R1 as i1 ¼ VIN =R1. Since the noninverting input does not draw current (Rule 2), we infer that the current i2 through resistor R2 is equal to i1. As a result, the voltage at the output becomes: VIN R2 VOUT ¼ VIN þ R2i2 ¼ VIN þ R2 ¼ VIN þ 1 ¼ VIN GCL ð7Þ R1 R1 This circuit is known as a noninverting amplifier since it provides an amplification gain GCL while preserving the phase (sign) of the input voltage VIN. 116 5 Signal Conditioning and Preprocessing

5.3.2 Buffering

The first and simplest application of op-amps is buffering, which is required to isolate different electronic stages and avoid impedance-loading errors. An analog buffer can be implemented with the voltage-follower circuit shown in Fig. 5.7c. This circuit provides (assuming an ideal op-amp) infinite input impedance and zero output impedance.

5.3.3 Amplification

An amplification or gain stage is typically required to bring the signal of the interface circuits to a level that is suitable for the dynamic range of a subsequent analog-to- digital converter. Amplifier circuits can be broadly classified into single-ended or dif-

ferential. A single-ended signal VIN, such as the one from a voltage divider, can be amplified with the noninverting amplifier described earlier in Fig. 5.7b or its inverting counterpart shown in Fig. 5.7d, in which the feedback resistor has been replaced by a potentiometer to allow for manual adjustments of the gain. In the case of Wheatstone bridge interface circuits, a differential amplifier stage, such as the one shown in Fig. 5.7e, may be used. This simple design, however, presents two basic drawbacks. First, the input impedance is significantly reduced since

the R1 resistors are in series with the input signals. Second, accurate matching of the resistor pairs (RA1 ¼ RB1) and (RA2 ¼ RB2) is required to ensure that the differential gains are similar and, therefore, provide good common-mode rejection. Due to these limitations, the so-called ‘instrumentation amplifiers’ are commonly used as difference stages. Fig. 5.7f shows a classical instrumentation amplifier design with three op-amps that can achieve high input impedance and common-mode rejection ratio without critical resistor matching [23]. The two op-amps at the input stage provide high differential gain and unity common-mode gain, whereas the second stage generates a single-ended output. Integrated instrumentation amplifiers are conveniently available from several manufacturers, with all components internal to the chip except

for R2, which can be connected externally to provide a programmable gain.

5.3.4 Filtering

Analog filters are used to remove unwanted frequency components from the sensor signals. Filters can be broadly grouped into four classes according to their frequency response [30, 31]: low-pass, high-pass, band-pass, and band-reject (Fig. 5.8). Low-pass filters allow frequencies below a cutoff frequencye to pass, while blocking frequencies

e) The cutoff frequency is defined as the frequency at which the gain is reduced by 3 dB (or a signal ratio of 0.707) 5.3 Signal Conditioning 117

Fig. 5.8 Frequency response of analog filters

above the cutoff. High-pass filters perform the opposite function, passing only frequencies above a cutoff. Band-pass filters allow passage of frequencies within a band. Band-reject (or notch) filters allow passage of all frequencies except for those within a, typically narrow, band. These analog filters can be implemented using passive or active circuits. Passive filters consist of networks of resistors, capacitors, and inductors, whereas active filters utilize active components (e.g., op-amps, transistors), in addition to passive devices, e.g. resistors and capacitors. Active filters are capable of implementing ‘virtual’ inductors by placing capacitors in the feedback loop, thus avoiding the bulk and nonlinearity of inductorsf. Active filters are suitable for low frequency, small signals, and are preferred over passive filters because they can have gains greater than 0 dB. Conversely, active filters require a power supply and are limited by the bandwidth of the active element. Passive filters have the advantage of being low-noise. Fig. 5.9a shows a pas-

Fig. 5.9 Low-pass first order filters: (a, b) passive and (c) active sive implementation of a first-order Butterworth (low-pass) filter, with a cut-off fre- 1 g quency FC ¼ð2pR2C2Þ and a roll-off slope of 20 dB/decade. Figure 5.9b shows an equivalent implementation with an inductor and a resistor. The active circuit shown in Fig. 5.9c also has a similar frequency response plus a static gain of

R2=R1. Finally, integrated circuits with low-pass, high-pass, band-pass, and band-reject outputs are also available in a single package from several manufacturers. These circuits, known as state-variable filters, are provided with extensive design formulas and tables and can be easily configured using only external resistors. f) Active filters could also use inductors, although g) Steeper roll-offs may be achieved by cascading they usually do not. several filters in series. 118 5 Signal Conditioning and Preprocessing

5.3.5 Compensation

A number of special functions may be implemented with analog circuits to compensate for deficiencies, cross-sensitivities, and nonlinearities in the sensor response, and reduce the computational load of a subsequent digital signal processing stage. These circuits perform various functions including linearization, integration, differentiation, logarithmic and antilogarithmic conversion, peak-to-peak and phase detection, and temperature compensation [29]. We now introduce several interface circuits for chemoresistors that can be used to obtain linear resistance-voltage relationships. These circuits are presented here, rather than in Section 5.2.1 with the remaining interface circuits, because they require familiarity with op-amps and they perform a compensation function. Additional compensation circuits for concentration and temperature are reviewed in Section 5.3.5.2.

5.3.5.1 Linearization of Resistance Measurements Among other shortcomings, voltage dividers have a nonlinear resistance-to-voltage transfer function. As a result, the sensitivity of the circuit is not constant over the dynamic range of the sensor. The resistance-to-voltage relationship can be easily linearized, however, by driving the sensing element at constant-voltage or constant-current. Figure 5.10a illustrates a constant-voltage measurement circuit that employs a virtual ground at the inverting input of the operational amplifier to apply a constant

voltage VCC across the sensor RS [20]. Negative feedback through a load resistor generates an output that changes linearly with the sensor conductance GS (the inverse of sensor resistance RS):

VCC VOUT ¼ISRL ¼ RL ¼VCCRLGS ð8Þ RS

An additional advantage of this circuit is that the load resistor RL can be chosen to provide different amplification gains.

Constant-current excitation is illustrated in Fig. 5.10b. The current IS through the sensor is entirely determined by the load resistor since the voltage at the op-amp in-

verting input is constant and equal to VCC [4]. The differential voltage across the sensor is then linearly proportional to the sensor resistance:

VCC VOUT ¼ RSIS ¼ RS ð9Þ RL

A similar constant-current arrangement can be used to provide a linear resistance- voltage relationship in Wheatstone bridges, as shown in Fig. 5.10c [4]. The operational amplifier provides a virtual ground to the midpoint of the sensor arm, generating a constant current through the sensor:

VCC IS ¼ ð10Þ R0 5.3 Signal Conditioning 119

Fig. 5.10 Linearizing a voltage divider through constant-voltage (a) or constant-current (b) measurements. Linearization of a Wheatstone bridge with a constant-current arrangement (c)

The voltage at the output of the op-amp is then proportional to the sensor resistance:

VCC V0 ¼RSIS RS ð11Þ R0 and the output of the circuit becomes: 1 RS 1 R0ð1 aÞ 1 VOut ¼ VCC 1 ¼ VCC 1 ¼ VCCa ð12Þ 2 R0 2 R0 2

5.3.5.2 Miscellaneous Functions A number of miscellaneous compensation functions may be performed with analog circuits. Figure 5.11a shows a logarithmic amplifier that may be used to compensate for the power-law concentration-resistance relationship R /½Cb of metal-oxide chemoresistors [32] and provide an output voltage proportional to the log concentration log[C] of the analyteh. Figure 5.11b illustrates a circuit that is employed in commercial

Fig. 5.11 Special functions: (a) logarithmic amplifier and (b) temperature compensation [1] h) The relationship VBE / logðICÞ may be used compensation for oscillations and ambient to derive the logarithmic transfer function. temperature [29]. This simple circuit, however, requires additional 120 5 Signal Conditioning and Preprocessing

gas alarm circuits to compensate for temperature [1, 2]. The circuit includes a thermis-

tor RTH (temperature dependent resistor) that adapts the alarm reference voltage VREF according to ambient temperature. The schematic in Fig. 5.11b uses a voltage regulator (7805) to provide a stable 5 V DC supply voltage to the heater and the voltage divider. Finally, the output of the comparator is current-boosted with an NPN transistor in order to drive an alarm.

5.4 Signal Preprocessing

Following an appropriate conditioning stage, the sensor array signals are digitized and either processed online or stored for future analysis. Due to space constraints, the reader is referred to the existing literature [30, 33] for a review of data acquisition for sensor systems (e.g., sample/hold, anti-aliasing, and analog-to-digital conversion). It is important to mention, however, that in order to avoid aliasing effects, the sampling rate during data acquisition should be at least twice the highest frequency in the sensor response. This is known as the Nyquist sampling theorem [34]. With this in mind, we focus our attention on signal preprocessing, the first computational stage after the sensor array data has been sampled and stored into computer memory. The goal of signal preprocessing is to extract relevant information from the sensor responses and prepare the data for multivariate pattern analysis (covered in Chapter 6). The choice of signal preprocessing is critical and can have a significant impact on the performance of subsequent modules in the pattern analysis system [35]. Although signal preprocessing is somewhat dependent on the underlying sensor technology, three general stages can be identified [36]: baseline manipulation, compression, and normalization.

5.4.1 Baseline Manipulation

The first stage of preprocessing consists of manipulating the sensor response with respect to its baseline (e.g., response to a reference analyte) for the purposes of drift compensation, contrast enhancement and scaling. Considering the dynamic

response of the sensor xSðtÞ shown in Fig. 5.12a, three techniques are commonly employed [3]:

* Differential: the baseline xSð0Þ is subtracted from the sensor response. As a result, any additive noise or drift dA that may be present in the sensor signal is effectively removed from the preprocessed response ySðtÞ:

ySðtÞ¼ðxSðtÞþdAÞðxSð0ÞþdAÞ¼xSðtÞxSð0Þð13Þ 5.4 Signal Preprocessing 121

* Relative: the sensor response is divided by the baseline. Relative measurements

eliminate the effect of multiplicative drift dM and provide a dimensionless response ySðtÞ:

xSðtÞð1 þ dMÞ xSðtÞ ySðtÞ¼ ¼ ð14Þ xSð0Þð1 þ dMÞ xSð0Þ

* Fractional: the baseline is subtracted and then divided from the sensor response. Fractional measurements are not only dimensionless but also normalized since the

resulting response ySðtÞ is a per-unit change with respect to the baseline, which compensates for sensors that have intrinsically large (or small) response levels:

xSðtÞxSð0Þ ySðtÞ¼ ð15Þ xSð0Þ

The choice of baseline manipulation technique and response parameter xSðtÞ (e.g., resistance, conductance, frequency) is highly dependent on the sensor technology and the particular application, but a few general guidelines can be extracted from

Fig. 5.12 Gas sensor transient response to an odor pulse (a). Transient analysis approaches: (b) sub-sampling, (c) parameter-extraction, and (d) system-identification 122 5 Signal Conditioning and Preprocessing

the literature. Gardner et al. [37, 38] have shown that the fractional change in conduc-

tance ySðtÞ¼ðGSðtÞGSð0Þ=GSð0Þ provides the best pattern-recognition performance for (n-type) MOS chemoresistors, compensating for temperature cross-sensitivity and nonlinearities in the concentration dependence [39]. Fractional methods for MOS chemoresistors are also widely used [40, 41]. In the case of conducting polymer chemoresistors, fractional changes in resistance are commonly employed, both in research prototypes and in commercial instruments [42, 43]. For piezo-electric

oscillators, where the response xSðtÞ being monitored is a frequency, differential measurements with respect to a reference analyte (and/or an uncoated reference sensor) are commonly used [12, 44]. Differential measurements are also widely used for

MOSFETs [45, 46], where the response xSðtÞ is a voltage shift in the I(V) curve as described in Section 5.2.3. Finally, a number of variations of these three basic baseline-manipulation techniques have been proposed in the literature, including data- driven procedures to optimize the baseline-manipulation stage for specific applications [35, 36, 47].

5.4.2 Compression

The second stage in preprocessing is aimed at compressing the sensor-array response down to a few descriptors to form a feature vector or fingerprint. In most cases this is performed by extracting a single parameter (e.g., steady-state, final, or maximum response) from each sensor, disregarding the initial transient response, which may be affected by the fluid dynamics of the odor delivery system (covered in Chapter 3). How- ever, with careful instrument design and sampling procedures, transient analysis can significantly improve the performance of gas sensor arrays:

* Improved selectivity. The dynamic response to an odor exposure (and the subsequent odor recovery) carries a wealth of odor-discriminatory information that cannot always be captured with a single parameter. In some situations, transient parameters have also been reported to exhibit better repeatability than static descriptors [48–50]. Therefore, sensor transients can be used as dynamic fingerprints to improve selectivity by pattern-recognition means. * Reduced acquisition time. The duration of the acquisition cycles can be significantly shortened if the initial sensor transients contain sufficient discriminatory information, avoiding the lengthy acquisition times required to reach steady state [51]. As a consequence, the sensors also require less time to recover their baseline, a process that can be particularly slow when the target odors have high concentrations. * Increased sensor lifetime. By reducing the duration of the odor pulse and, therefore minimizing irreversible binding, the lifetime of the sensors can also be increased.

For these reasons, transient analysis has received much attention in recent years. Ac- cording to the procedure employed to generate the dynamic fingerprint, transient compression methods can be broadly grouped into three classes: 5.4 Signal Preprocessing 123

* Sub-sampling methods: As depicted in Fig. 5.12b, these methods exploit dynamic information by sampling the sensor transient response (and/or its derivatives) at different times during the odor exposure and/or odor recovery phase [36, 49, 52, 53]. * Parameter-extraction methods: These methods compress the transient response using a number of descriptors, such as rise times, maximum/minimum responses and slopes, and curve integrals. [48, 54–56]. * System-identification methods: These methods fit a theoretical model (e.g., multiexponential, auto-regressive) to the experimental transients and use the model parameters as features [55, 57, 58].

Exponential curve-fitting methods can result in nearly lossless compression of the sensor transients, but are computationally intensive [57, 59]. For these reasons, sub- sampling and parameter-extraction methods are more commonly employed. A final word of caution regarding the use of transient information: a large number of dynamic parameters will require an exponentially increasing number of training examples in order to prevent the pattern recognition system from over-fitting the data. Alterna- tively, one may use resampling techniques (e.g., cross-validation, bootstrap) or regu- larization (e.g., shrinkage, weight decay) to control the complexity of the model. Further details on small-database considerations and dynamic pattern-recognition methods may be found in Chapter 12 of this Handbook.

5.4.3 Normalization

Normalization constitutes the final stage of digital preprocessing prior to multivariate pattern analysis. Normalization techniques can be broadly grouped in two classes: local and global methods. Local methods operate across the sensor array on each individual “sniff” in order to compensate for sample-to-sample variations caused by analyte concentration and sensor drift, among others. Global methods, on the other hand, operate across the entire database for a single sensor (e.g., the complete history of each sensor), and are generally employed to compensate for differences in sensor ðk scaling. In what follows, we will denote by xS the response of sensor ‘s’ to the k-th example in the database.

5.4.3.1 Local Methods The most widely used local method is vector normalization, in which the feature vector of each individual ‘sniff’ is divided by its norm and, as a result, forced to lie on a hyper- sphere of unit radius, as shown in Fig. 5.13d,e:

ðk ðk rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffixS yS ¼ P ð16Þ ðk 2 ðxS Þ S 124 5 Signal Conditioning and Preprocessing

Fig. 5.13 Normalization procedures: (a,d) raw data, (b) sensor autoscaling, (c) sensor normalization and (e) vector normalization

Vector normalization can be employed to compensate for differences in concentration ðk ðk b between samples. Assuming the power-law relationship xs;a ¼ as;a½Ca of metal-oxi- ðk de chemoresistors [32], where xs;a is the response of sensor ‘s’ to the k-th sample of ðk odor ‘a’, as;a is the sensitivity of sensor ‘s’ to odor ‘a’, and ½Ca is the concentration of the k-th sample of odor ‘a’, then: hi b ðk ðk a C x ; s;a a a yðk ¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffis a ¼ sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffis;a ð17Þ s;a P 2 hi P 2 ðk P b 2 xs;a a ðk as;a s s;a Ca s s

To the extent that these simplifying assumptions hold, vector normalization can therefore be used to compensate for sample-to-sample variations in concentration. In this context, vector normalization can be applied in situations when each odor has a unique concentration, but discrimination is to be performed on the basis of odor quality (e.g., !ðk the direction of the response vector x a ) rather than odor intensity (e.g., the magni- !ðk tude of x a ). Conversely, this method should not be used when the vector amplitude is known to carry relevant information. 5.5 Noise in Sensors and Circuits 125

5.4.3.2 Global Methods Two global procedures are commonly employed in electronic nose systems:

* Sensor autoscaling, in which the distribution of values for each sensor across the entire database is set to have zero mean and unit standard deviation:

ðk ðk xs mean½xs ys ¼ ð18Þ std½xs

* Sensor normalization, in which the range of values for each individual sensor is set to [0,1]. This is simply done by subtracting the minimum and dividing by the range of the sensor across the entire database:

ðk ðk ðk xs min8k½xs ys ¼ ðk ðk ð19Þ max8k½xs min8k½xs

Global methods are typically used to ensure that sensor magnitudes are comparable, preventing subsequent pattern-recognition procedures from being overwhelmed by sensors with arbitrarily large values. For instance, nearest-neighbors procedures are extremely sensitive to feature weighting, and multilayer perceptrons can saturate their sigmoidal activation functions for large inputs. Sensor normalization makes full use of the input dynamic range but, as illustrated in Fig. 5.13a,c, is very sensitive to outliers since the range is determined by data outliers. Autoscaling, on the other hand, cannot provide tight bounds for the input range but is robust to outliers. However, it must be noted that both techniques can amplify noise since all the sensors (particularly those which may not carry information) are weighted equally. Logarithm metrics have also been used to compensate for highly nonlinear concentration effects [41]. It is also worth mentioning the Box-Cox transform [60], which could be employed to compensate for nonlinearities, as well as compress the dynamic range of the sensors: 8 <> ðk k ðxs Þ 1 ðk k k 6¼ 0 ys ¼ ð20Þ :> ðk ln xs k ¼ 0

5.5 Noise in Sensors and Circuits

Noise is generally considered to be any unwanted effect that obscures the detection or measurement of the desired signal. As shown in Fig. 5.14a, noise can arise at various stages in the measurement process, including the quantity under measurement itself, the sensors, the analog processing system, the data acquisition stage and the digital signal processing system. Among these, noise in the early measurement stages is clearly most harmful as it propagates and can be potentially amplified through the 126 5 Signal Conditioning and Preprocessing

Fig. 5.14 (a) Sources of noise in sensor systems. (b) Power spectral density of white and 1/f noise. (c) Quantization noise in A/D conversion

subsequent stages in the signal pathway [61]. Several noise sources, such as thermal and shot noise, are inherent to the underlying physics of the sensors or electronic components and are, therefore, irreducible. Other types of noise, conversely, are originated from processes that could be avoided, and include 1/f noise, transmission and quantization noise. Thermal noise, also known as Johnson or Nyquist noise, arises in any medium that dissipates energy, such as a conductor. This means that even a simple resistor is a noise source.pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi The open-circuit noise voltage generated by a resistance R is Vnoise ¼ 4kTRDf , where k is Boltzman’s constant, T is the absolute temperature (Kelvins), and Df is the bandwidth (Hz) over which the measurement is made [23]. Therefore, the larger the resistance, the more noise it can introduce. Thermal noise has a flat power spectral density (PSD), and is oftentimes called white noise in analogy to white light, which has a flat distribution of all frequencies in the visible spectrum. In addition, the amplitude distribution of thermal noise is Gaussian [23]. Shot or Schottky noise arises from the random fluctuations in the number of charge carriers (electrons and holes) that cross a potential barrier in the charge flow, and is typical of p-n junctionspffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi in diodes and transistors. The shot-noise RMS current fluctuation is Inoise ¼ 2qlDCDf , where q is the electron charge, IDC is the average current through the barrier, and Df is the bandwidth. Shot noise is also white and Gaussian [4]. 1/f (read ‘one-over-f’) or flicker noise is considered to arise from imperfections in the manufacturing process of electronic components. As the name indicates, 1/f noise has 5.5 Noise in Sensors and Circuits 127 a PSD that is inversely proportional to frequency. For this reason it is also known as low-frequency or pink noise (red is at the low side of the visible spectrum). It is also referred to as excess noise because it appears in addition to white noise, as illustrated in Fig. 5.14b. 1/f noise is most pronounced at frequencies below 100 Hz, where many sensors operate, and becomes barely noticeable at frequencies above a few hundred KHz where white noise dominates. In contrast with thermal noise, which equally affects a cheap carbon resistor or the most carefully made resistor, 1/f noise can be reduced by using good quality metal film or wire-wound resistors at the early stages of sensor interface circuits [23]. Noise can also be transmitted from interferences such as fluctuations in the DC power supply, 50–60 Hz pickup, changes in ambient temperature, capacitive or inductive couplings, and ground loops. A careful layout and construction of the electronics, with proper shielding and grounding, must be used to reduce electromagnetic interference noise to acceptable levels [23]. In addition, differential measurements, such as the ones in Fig. 5.4d,e, can be employed to compensate for noise effects that are additive in nature. Multiplicative effects, on the other hand, can be reduced by means of ratiometric measurement techniques [5]. Analog filtering (Section 5.3.4) and digital signal preprocessing (Section 5.4) can also be employed to further reduce noise. For instance, differentiation can be used to reduce low-frequency noise (e.g., drift) at the expense of amplifying high-frequency components. Conversely, integration or averaging reduces high-frequency noise while amplifying low-frequency components. As mentioned earlier, noise can also arise in the latter stages of the signal pathway, primarily during analog-to-digital conversion, when the continuous sensor signals are converted into a discrete subset of values and stored in computer memory. This process introduces nonlinear quantization errors that can be treated as an additional noise source, as depicted in Fig. 5.14c. Quantization noise must be controlled by selecting an appropriate gain in the signal conditioning circuits to fully utilize the dynamic range of the analog-to-digital converter, and by employing differential measurements to remove uninformative baseline offsets in the sensor response [62]. Limitations in machine precision and fixed-point arithmetic can also introduce digital noise in the signal pathway. For a systematic treatment of quantization and finite word-length noise, the reader is referred to the literature [34]. Finally, it is important to notice that the inherent drift and poor repeatability of the sensor responses can sometimes be significantly larger than most of the other noise sources described in this section, effectively limiting the sensitivity of electronic nose systems. As proposed previously [61], the global effect of all these noise sources can be combined into a single parameter called the noise-equivalent concentration, which indicates the gas concentration that results in a unit signal-to-noise ratio. 128 5 Signal Conditioning and Preprocessing

5.6 Outlook

From their original conception as arrays of homogeneous gas sensors with overlapping selectivities, electronic-nose systems, including those commercially available, are slowly evolving towards hybrid arrays that take advantage of several sensor technologies [63]. The use of sample preconditioning such as thermal-desorption units or chromatographic columns, is also becoming increasingly popular as the means to increase the sensitivity and selectivity of the instrument [64–66]. An additional trend in electronic-nose systems has become the measurement of multiple parameters from the same sensing membrane [67]. We focus our attention on the latter, since the use of hybrid systems does not introduce conceptual problems other than the integration of the various sensor technologies into a single package, and sample preconditioning methods are covered in Chapter 3 of this Handbook. Multiparameter sensing approaches can be broadly grouped in three categories:

* Similar sensing layer but different transduction principles: these systems extract multiple physical parameters from the same sensing layer, such as work function and conductance on MOS sensors, or resistance and mass changes in conducting polymer sensors. * Similar sensing layer and transduction principle but different operating modes: in this case, the selectivity of the sensor is modified by modulating the operating conditions, such as temperature cycling in MOS sensors or AC impedance spectroscopy in MOS or conducting polymer sensors. * Similar sensing layer, transduction principle, and operating modes but different features: A third possibility is to extract multiple parameters from the sensor transient response.

In this section, we review a multiparameter technique for metal-oxide sensors that has received much attention in recent years: temperature modulation. AC impedance spectroscopy and transient analysis, which can also been used as multiparameter approaches to improve the selectivity of gas sensors, were covered in Sections 5.2.1.3 and 5.4.2, respectively. For additional material on multiparameter sensor systems the reader is referred to the authoritative review of Weimar and Go¨pel [67].

5.6.1 Temperature Modulation

The selectivity of metal-oxide sensors is greatly influenced by the operating temperature of the device, since the reaction rates for different volatile compounds and the stability of adsorbed oxygen species are a function of surface temperature [68]. This temperature-selectivity dependence can be utilized to improve the performance of MOS sensors. Rather than maintaining a constant operating point, as described in Section 5.2.4, the temperature of the sensor may be cycled during exposure to an odor 5.7 Conclusions 129

Fig. 5.15 Left: Sensitivity-temperature profile for Pt- and Pd-doped tin-oxide sensors [70]. Right: conductance-temperature response of a tin-oxide gas sensor in (a) air, (b) methane, (c) ethane, (d) propane, (e) n-butane, (f) isobutene, (g) ethylene, (h) propylene, and (i) carbon monoxide [71]

to obtain a multivariate dynamic signature. Figure 5.15a illustrates the sensitivity profiles of several doped tin-oxide gas sensors at different temperatures when exposed to various analytes. If maximum sensitivity to a particular analyte, say C3H8, were needed, a constant temperature of 250 8C for the Pd-doped sensor would then be most suitable. For machine olfaction applications, however, where the analyte detection range is broader, it would be advantageous to capture the response of the sensor over the entire temperature range. Figure 5.15b shows the conductance-temperature dynamic response to various analytes when a sinusoidal voltage (2–5 V, 0.04 Hz) is applied to the heater of a commercial SnO2 sensor (Figaro TGS813). It can be observed that not only the magnitude of the conductance but also the shape of the dynamic response is unique to each analyte. An excellent survey of temperature modulation in semiconductor gas sensing may be found in [69].

5.7 Conclusions

This chapter has presented the hardware and software components that constitute the interface between chemical sensor arrays and pattern analysis techniques, the two critical building blocks in odor-sensing systems. We have surveyed a number of interface circuits that can be used to generate electrical signals for the most popular gas sensing technologies: chemoresistive, acoustic wave, and field effect sensors. Analog signal conditioning of the resulting electrical signals has also been outlined, including a gentle review of operational amplifiers. Various approaches for controlling the 130 5 Signal Conditioning and Preprocessing

operating temperature of metal-oxide sensors have also been presented. Finally, preprocessing algorithms to prepare sensor-array data for multivariate pattern analysis have been described. Although often overlooked, careful selection of sensor interface circuits, signal conditioning, and preprocessing is critical for achieving optimal performance in odor-sensing systems.

5.8 Acknowledgements

This work was partially supported by the award NSF/CAREER 9984426. The authors are grateful to J. W. Gardner and T. C. Pearce for helpful suggestions during the review process of this manuscript.

References

1 Figaro, General Information for TGS Sensors, Design and Physico-Chemical Applications, Figaro Engineering, Inc., Osaka, Japan, Academic Press, San Diego, CA, 1997. 1996. 14 A. Russell. Odour detection by mobile robots, 2 FIS, Products Review, FIS Inc., Osaka, Japan, World Scientific, Singapore, 1999. 1998. 15 R. A. Russell, L. Kleeman, S. Kennedy. 3 J. W. Gardner, P. N. Bartlett. Electronic Noses, Proceedings of the 2000 Australian Conference Principles and Applications, Oxford Uni- on Robotics and Automation, Melbourne, versity Press, Oxford, UK, 1999. Aug. 30-Sept. 1, 2000, 87–92. 4 R. Pallas-Areny, J. G. Webster. Sensors 16 A. Spetz, F. Winquist, H. Sundgren, and Signal Conditioning,2nd Edition, Wiley, I. Lundstrom. 1992, in Gas Sensors New York, 2001. (Ed.: G. Sverveglieri), Kluwer Academic 5 J. Fraden. Handbook of Modern Sensors. Publishers, 1992, 219–279. Physics, Designs and Applications, 2nd 17 J. V. Hatfield, J. A. Covington, J. W. Gardner. Edition, American Institute of Physics, Sens. Actuators B, 2000, 65(1–3), 253–256. Woodbury, New York, 1997. 18 I. Lundstrom, E. Hedborg, A. Spetz, H. 6 W. Go¨pel, K. D. Schierbaum. Sens. Sundgren, F. Winquist. In Sensors and Actuators B, 1995, 26–27, 1–12. Sensory Systems for an Electronic Nose, 7 U. Hoefer, K. Steiner, E. Wagner. Sens. (Eds.: J. W. Gardner, P. N. Bartlett), Kluwer Actuators B, 1995, 26–27, 59–63. Academic Publishers, Dordrecht, 1992, 8 U. Weimar, W. Go¨pel. Sens. Actuators B, 303–319. 1995, 26–27, 13–18. 19 R. C. C. Li, P. C. H. Chan, P. W. Cheung. 9 G. Sberveglieri. Sens. Actuators B, 1995, 23, Sens. Actuators B, 1995, 28(3), 233–242. 103–109. 20 K. Ikohura, J. Watson. The Stannic Oxide Gas 10 M. E. H. Amrani, K.C. Persaud, P. A. Payne. Sensor, Principles and Applications, CRC Meas. Sci. Technol., 1995, 6(10), 1500–1507. Press, Boca Raton, FL., 1994. 11 J. W. Grate, G. C. Frye. In Sensors Update 21 M. Benammar, W. C. Maskell. J. Phys. E: Vol. 2, (Eds.: H. Baltes, W. Go¨pel and Sci. Instrum., 1989, 22, 933–936. J. Hesse), VCH, Weinheim, 1996, Chapter 2. 22 P. McGeeghin, P. T. Moseley, 12 J. W. Grate, S. J., Martin, R. M. White. Anal. D. E. Williams. Sensor Review, 1994, 14(1), Chem., 1993, 65(21), 940–948. 13–19. 13 D. S. Ballantine, R. M. White, S. I. Martin, 23 P. Horowitz, W. Hill. The art of electronics, A. J. Ricco, E. T. Zellers, G.C. Frye, Cambridge University Press, Cambridge, H. Wohltjen. Acoustic Wave Sensors. Theory, UK, 1989. 5.8 Acknowledgements 131

24 A. Heilig, N. Barsan, U. Weimar, W. Go¨pel. 46 N. Paulsson, F. Winquist. Forensic Sci. Int., Sens. Actuators B, 1999, 58(1–3), 302–309. 1999, 105(2), 95–114. 25 S. Jonda, M. Fleischer, H. Meixner. Sens. 47 T.C. Pearce, J. W. Gardner. The Analyst, Actuators B, 1996, 34(1–3), 396–400. 1998, 123, 2057–2066. 26 A. P. Lee, B. J. Reedy. Sens. Actuators B, 2000, 48 E. Llobet, J. Brezmes, X. Vilanova, 69(1–2), pp. 37–45. X. Correig, J.E. Sueiras. Sens. Actuators B, 27 P. Mielle. Sens. Actuators B, 1996, 34(1–3), 1997, 41(1–3), 13–21. 533–538. 49 S. Roussel, G. Forsberg, V. Steinmetz, 28 W. M. Sears, K. Colbow, F. Consadori. Sens. P. Grenier, V. Bellon-aurel. J. Food Eng., Actuators, 1989, 19, 333–349. 1998, 37, 207–22. 29 J. J. Carr. Designer’s Handbook of Instru- 50 B.W. Saunders, D.V. Thiel, A. Mackay-Sim. mentation and Control Circuits, Academic The Analyst, 1995, 120, 1013–1018. Press, San Diego, CA, 1991. 51 F. Sarry, M. Lumbreras. Sens. Actuators B, 30 H. R. Taylor. Data Acquisition for Sensor 2000, 67, 258–264. Systems, Chapman and Hall, London, UK, 52 J. White, J. S. Kauer, T. A. Dickinson, 1997. D. R. Walt. Anal. Chem., 1996, 68(13), 31 D. C. Ramsay. Principles of Engineering 2191–2202. Instrumentation, Arnold, London, UK, 1996. 53 B. G. Kermani, S. S. Schiffman, H. T. Nagle. 32 P. K. Clifford, D. T. Tuma. Sens. Actuators, IEEE Trans. Instrum. Meas., 1998, 47(3), 1982, 3, 233–254. 728–741. 33 J. Brignell, N. White. Intelligent sensor 54 D. M. Wilson, S. P. DeWeerth. Sens. systems, Institute of Physics Publishing, Actuators B, 1995, 28, 123–128. Bristol, UK, 1996. 55 T. Eklov, P. Martensson, I. Lundstrom. Anal. 34 C. L. Phillips, H. T. Nagle. Digital Control Chim. Acta, 1997, 353, 291–300. System Analysis and Design, Prentice Hall, 56 T. D. Gibson, O. Prosser, J. N. Hulbert, Englewood Cliffs, New Jersey, 1995. R. W. Marshall, P. Corcoran, P. Lowery, 35 J. W. Gardner, M. Craven, C. Dow, E. A. Ruck-Keene, S. Heron. Sens. Actuators E. L. Hines. Meas. Sci. Technol., 1998,9, B, 1997, 44(1–3), 413–422. 120–127. 57 R. Gutierrez-Osuna, H. T. Nagle, 36 R. Gutierrez-Osuna, H. T. Nagle. IEEE S. S. Schiffman. Sens. Actuators B, 1999, Trans. Sys. Man Cyber. B, 1999, 29(5), 61(1–3), 170–182. 626–632. 58 T. Nakamoto, A. Iguchi, T. Moriizumi. Sens. 37 J. W. Gardner. Sens. Actuators B, 1991,4, Actuators B, 2000, 71, 155–160. 109–115. 59 E. Llobet, X. Villanova, J. Brezmes, 38 J. W. Gardner, E. L. Hines, H. C. Tang. R. Alcubilla, J. Calderer, J. E. Sueiras, Sens. Actuators B, 1992, 9, 9–15. J. Correig. Meas. Sci. Technol., 1997,8, 39 J. W. Gardner, P. N. Bartlett. Sens. Actuators 1133–1138. B, 1994, 18–19, 211–220. 60 G. E. P. Box, D. R. Cox. J. Roy. Statist. Soc. 40 G. Horner, C. Hierold. Sens. Actuators B, Ser. B, 1964, 26, 211–243. 1990, 2, 173–184. 61 F. Bordoni, A. D’Amico. Sens. Actuators A, 41 H. Abe, T. Yoshimura, S. Kanaya, Y. Taka- 1990, 21–23, 17–24. hashi, Y. Miyashita, S.-I. Sasaki. Anal. Chim. 62 P. Corcoran. Sens. Actuators B, 1994, 18–19, Acta, 1987, 194, 1–9. 649–653. 42 K.C. Persaud, S.M. Khaffaf, J.S. Payne, 63 H. Ulmer, J. Mitrovics, U. Weimar, W. A.M. Pisanelli, D.-H. Lee, H.-G. Byun. Sens. Go¨pel. Sens. Actuators B, 2000, 65(1–3), Actuators B, 1996, 36(1–3), 267–273. 79–81. 43 E. J. Severin, B. J. Doleman, N. S. Lewis. 64 J. W. Grate, S. L. Rose-Pehrsson, Anal. Chem., 2000, 72(4), 658–668. D. L. Venezky, M. Klusty, H. Wohltjen. 44 A. Hierlemann, U. Weimar, G. Kraus, Anal. Chem., 1993, 65, 1868–1881. M. Schweizer-Berberich, W. Go¨pel. Sens. 65 B. Hivert, M. Hoummady, D. Hauden, Actuators B, 1995, 26(1–3), 126–134. P. Mielle, G. Mauvais, J. M. Henrioud. Sens. 45 H. Sundgren, F. Winquist, I. Lukkari, I. Actuators B, 1995, 27(1–3), 242–245. Lundstrom. Meas. Sci. Technol., 1991, 2(5), 66 S. Strathmann. Sample Conditioning for 464–469. Multi-Sensor Systems, Ph.D. Dissertation, 132 5 Signal Conditioning and Preprocessing

Institute for Physical and Theoretical 69 A. P. Lee, B. J. Reedy. Sens. Actuators B, 1999, Chemistry, University of Tu¨bingen, 60, 35–42. Germany, 2001. 70 N. Yamazoe, N. Miura. In Chemical Sensor 67 U. Weimar, W. Go¨pel. Sens. Actuators B, Technology Vol. 4 (Ed.: S. Yamauchi), 1998, 52, 143–161. Chemical Sensor Technology, 1992, 19–42. 68 M. J. Madou, S. R. Morrison. Chemical 71 S. Nakata, S. Akakabe, M. Nakasuji, Sensing with Solid State Devices, Academic K. Yoshikawa. Anal. Chem., 1996, 68, Press, Boston, MA, 1989. 2067–2072. 133

6 Pattern Analysis for Electronic Noses

Evor L. Hines, Pascal Boilot, Julian W. Gardner and Mario A. Gongora

Abstract This chapter provides a detailed description of a comprehensive set of pattern recognition (PARC) techniques that have been employed to analyze electronic nose (EN) data; i.e. well-known and commonly used techniques, research algorithms and future trends in pattern analysis. The problem of pattern analysis of EN data is closely linked to that of multivariate data analysis. Both statistical and non-parametric multivariate analysis techniques are discussed here. The chapter focuses on basic chemometric techniques and so those based on the principles of engineering, mathematics and statistics. We first describe methods that are common conventional statistical methods, such as principal components analysis (PCA), partial least squares (PLS), multiple linear regression (MLR), principal component regression (PCR), discriminant function analysis (DFA) including linear discriminant analysis (LDA), cluster analysis (CA) including nearest neighbor (NN). We then briefly explore the development of biologically motivated non-parametric methodologies, such as artificial neural networks (ANNS) including multi-layer perceptron (MLP), fuzzy inference systems (FIS), self-organizing map (SOM), radial basis function (RBF), genetic algorithms (GAS), neuro-fuzzy systems (NFS) and adaptive resonance theory (ART). There has always been an appeal when working on EN architectures that mimic the human olfactory system, namely to build physiologically inspired PARC systems that imitate the human brain. The classification scheme presented here is made on three levels: first a distinction is made between statistical and biological approaches, then between quantitative and qualitative pattern analysis algorithms, and finally supervised and unsupervised techniques. Together these provide the reader with a comprehensive review of pattern analysis techniques for ENS.

6.1 Introduction

The electronic nose (EN) is an instrument that has been developed to mimic the human organ for smell; i.e. the biological olfactory system presented in Chapter 1. The EN is both a chemical sensing and a data analysis system that can, to some extent, discriminate between different simple or complex odors. The nature of odors and aromatic volatile compounds has been previously discussed in Chapter 1. Gener- ally, the EN is developed as a match-model for the natural nose comprising the various stages between a volatile odorant and its recognition, namely: interaction, signal generation, processing, and identification, as outlined by the parallel between biological and artificial noses in Fig. 6.1. The system comprises a chemical sensor array, together with an interfacing electronic circuitry and a pattern-recognition unit that acts as a signal processing system [1]. However, a simpler model based on an array of sensors and a pattern recognition system was later introduced, which helps to better understand and represent how the nose functions [2]. A discussion of chemosensors and signal pre-processing is given in Chapters 4 and 5, respectively. Both models mentioned above incorporate a pattern recognition system, yet much effort in EN development work has focused on the sensor and instrumentation design while data exploration has perhaps been relatively neglected for long periods. In this chapter, we review the pattern analysis techniques, classification systems, identification methods and recognition algorithms that have been applied to solve olfactory problems. Data analysis, machine learning or chemometrics are being widely used today in physical, chemical, and engineering sciences, so that currently there are a large number of pattern recognition (PARC) techniques available. In order to select appropriate PARC algorithms for EN applications, it is important to understand the fundamental

Fig. 6.1 Basic diagram showing the analogy between biological and artificial noses 6.1 Introduction 135 nature of the data being analyzed. The problem of analyzing EN data sets is one of determining the underlying relationships between one set of independent variables (e.g. outputs from an array of n sensors) and another set of dependent variables (e.g. odor classes and component concentrations) using for example multivariate analysis [3]. The general multivariate problem in odor sensing is commonly referred to as PARC and is used to analyze qualitatively the odor patterns produced by these instruments; but it can also possibly be used quantitatively, for example to compute individual component concentrations. It is envisaged that efficient data processing and pattern analysis will provide more accurate models and better understanding of the data generated. Pattern recognition algorithms and data processing techniques are a critical component in the implementation, development and successful commercialization of ENs.

6.1.1 Nature of Sensor Array Data

Now, let us consider an array of n discrete sensors, as illustrated in Fig. 6.2, where each sensor i produces a time-dependent output signal XijðtÞ in response to an odor j. The electrical sensor signal depends on several physical parameters (e.g. flow rate of odor across sensor, ambient pressure, temperature and humidity), but the sensor outputs are expected to reach constant asymptotic values when presented with a constant input stimulus. It has been common practice to use only the static or steady-state values of the sensor signals rather than the dynamic or transient responses, the response is then simply a time-independent parameter, XijðtÞ!Xij. However, the choice of the response parameter is fundamental to the subsequent performance of the PARC, so the pre-processing technique, which is applied to the response vectors, is usually designed to help analyze data in the context of a specific problem. Generally, in order to extract relevant key features from the data in terms of the static change in sensor parameter (e.g. resistance or conductivity), a good choice is to use a fractional differ- odor 0 0 odor ence model: Xij ¼ðXij Xi Þ=Xi where Xij is the response of the sensor i to the 0 sample odor j, and Xi is the baseline or reference signal, such as the value in ambient room air. The response generated by the n-sensor array to an odor j can then be re- T presented by a time-independent vector: Xj ¼ðX1j; X2j; :::; Xij; :::; XnjÞ . When the same array is presented to a set of m odors, the responses can be regarded as a set of m vectors, which are best represented by a response matrix R: 0 1 X X ::: X B 11 12 1m C B X21 X22 ::: X2m C B C : R ¼ @ . . . A ð6 1Þ . . Xij . Xn1 Xn2 ::: Xnm

Each column represents a response vector associated with a particular odor, whereas the rows are the responses of an individual sensor to the different measurands. As odor sensors are not entirely specific, an individual sensor will respond to a variety 136 6 Pattern Analysis for Electronic Noses

of odors but with varying sensitivity (e.g. speed and intensity of the response). As a result, the off-diagonal terms of R are usually non-zero, and thus, under these conditions, PARC techniques are required to process the data and solve the class prediction problem.

6.1.2 Classification of Analysis Techniques

The responses generated by an array of odor sensors may be processed using a variety of techniques. In Fig. 6.2, where the basic data-processing structure of an EN is presented, the array formed from the sensor outputs is pre-processed and normalized so that the modified response matrix can be fed into a PARC engine (see Chapter 5). The nature of a PARC engine is usually classified in terms of being parametric or non- parametric, and supervised or unsupervised.

* Parametric. A parametric technique, commonly referred to as a statistical approach, is based on the assumption that the spread of the sensor data can be described by a probability density function (PDF). In most cases, the assumption made is that the data follow a normal distribution with a constant mean and variance. These techniques try to find an underlying mathematically formulated relationship between system inputs, odor vectors and its outputs, classes or descriptors. * Non-parametric. Non-parametric methods do not assume any specific PDF for the sensor data and thus apply more generally. This approach to multivariate data analysis has led to the fields of artificial neural networks (ANNS) and expert systems.

Fig. 6.2 Basic architecture of a data processing system for an EN 6.1 Introduction 137

* Supervised. In a supervised learning PARC method, a set of known odors are systematically introduced to the EN, which then classifies them according to known descriptors or classes held in a knowledge base. Then, in a second stage for identification, an unknown odor is tested against the knowledge base, now containing the learnt relationship, and then the class membership is predicted. Unknown odor vectors are analyzed using relationships found a priori from a set of known odor vectors used in an initial calibration, learning, or training stage. The idea of testing a method using unclassified response vectors is well established and is often referred to as cross-validation. * Unsupervised. For unsupervised learning, PARC methods learn to separate the different classes from the response vectors routinely, discriminating between unknown odor vectors without being presented with the corresponding descriptors. These methods are closer to the way that the human olfactory system works using intuitive associations with no, or little, prior knowledge.

6.1.3 Overview

This chapter provides a detailed description of a comprehensive list of PARC techniques that have been employed to analyze EN data; i.e. well-known and commonly used techniques, up-to-date algorithms and future trends in pattern analysis. Both statistical and non-parametric analysis techniques are discussed. The chapter focuses on basic chemometric techniques and so those based on the principles of engineering, mathematics and statistics [4]. Thus we first describe methods that are common conventional statistical methods, such as principal components analysis (PCA), partial least square (PLS), multiple linear regression (MLR), principal component regression (PCR), discriminant function analysis (DFA) including linear discriminant analysis (LDA), and cluster analysis (CA) including nearest neighbor (NN). Then we briefly explore the development of biologically motivated methodologies, such as artificial neural networks (ANNS) including multi-layer perceptron (MLP), fuzzy inference systems (FIS), self-organizing map (SOM), radial basis function (RBF), genetic algorithms (GAS), neuro-fuzzy systems (NFS) and adaptive resonance theory (ART). There has always been an appeal, when working on EN that mimic the human olfactory system, to build physiologically inspired PARC systems that imitate the human brain. As stated above, the problem of pattern analysis of EN data is closely linked to the multivariate analysis of data sets. Figure 6.3 summarizes the main multivariate data processing techniques, or PARC algorithms, that have been employed in the field of ENs and which are explored in this chapter. The classification scheme is made on three levels: a first distinction is made between statistical and biological approaches, then between quantitative and qualitative pattern analysis algorithms, and finally between supervised and unsupervised techniques. Specifically, Section 6.2 describes the commonly used conventional or classical statistical pattern analysis techniques, whereas Section 6.3 describes some biologically inspired or ‘intelligent’ PARC models, such as ANNs. Key factors for a comparison of these algorithms are presented in Section 6.4 138 6 Pattern Analysis for Electronic Noses

Fig. 6.3 Classification scheme of the multivariate pattern analysis techniques applied to EN data

together with future trends in EN pattern analysis in terms of the use of dynamical analysis and intelligent sensor systems.

6.2 Statistical Pattern Analysis Techniques

Classical statistical methods, using a probability model, were first developed and used in the field of applied mathematics, now called chemometrics. In this section some mathematical methods that may be applied to the multi-component analysis of odors, are presented. Categorization of classifiers, as presented in the previous paragraph, can be made based on certain features, such as supervised or unsupervised, model- based or model-free, qualitative or quantitative. For example, discriminant function analysis (DFA) is a parametric and supervised learning classifier, which can be used for both qualitative and quantitative analysis. Principal components analysis (PCA) is a non-parametric projection method and is often used to implement a linear supervised classifier, in conjunction with discriminant analysis. 6.2 Statistical Pattern Analysis Techniques 139

6.2.1 Linear Calibration Methods

Linear multivariate calibration methods, using linear algebra, are often used to process sensor array data and obtain the concentrations within a multi-component mixture. This is usually based on two basic assumptions: 1) that the response of each sensor is proportional to the component concentration (linear sensor model), and 2) that the response of a sensor mixture equals the sum of the responses to the individual components (superposition model). The multiple linear regression (MLR) method is commonly used to analyze mixtures of gases and vapors. MLR uses sensor responses vari- s ables Xij to predict component concentrations cj from a regressive equation holding the partial regression coefficients bij [5].

s cj ¼ b1jX1j þ b2jX2j þ ::: þ bijXij þ ::: þ bnjXnj ð6:2Þ

The goal of MLR is to calculate the values of the regression coefficients bij for the sensors, minimizing the sum of squared deviations (gradient descent) between the s predicted component concentration values cj and the actual measured concentration values. MLR has been successfully applied to analyze the response of nine odor sensors to certain organic vapors [6]. It is a technique widely used in chemometrics that works best with orthogonal variables for which sensors are component specific, however it is sensitive to noise and suffers from the considerable degree of co-linearity present in solid-state odor sensors, for example tin-oxide resistors. When it is desirable to determine the individual gas concentrations from a multivariate calibration, two other methods used in preference to MLR, are PLS and PCR, which assume that a linear-inverse model can be applied to the data. In the model, the concentration vector c is related to the response matrix R by c ¼ Rm þ e where m is a regression vector containing all the model parameters, and e is an error vector containing the concentration residuals from other gases. The regression vectors are estimated in PLS and PCR by finding the pseudo-inverse response matrix in terms of orthonor- mal and diagonal matrices [7]. PLS was first described in the mid-1960s and has since been refined and specialized for chemical applications [8]. PLS is often applied to gas mixture analysis because it accepts collinear data, separates out noise from useful sample information, and makes meaningful linear combinations for different concentrations. It is also one of the latest regression procedures, based on the properties of MLR, to be developed for concentration prediction. The main difference between PLS and PCR is that PLS includes information about the concentration vector in the model building while PCR does not. This is important when analyzing data to classify odors rather than to predict chemical concentrations. Since most chemical sensors have a non-linear concentration dependence, these techniques are only approximately valid within a small, or a low (e.g. Langmuir model) concentration range. In order to handle non-linear data and improve the performance of linear PARC, the sensor response against concentration can be linearized using either an appropriate pre-processing technique, or by using a non-linear MLR model [9]. A non-linear PLS for correcting non-linearities after calculations has been 140 6 Pattern Analysis for Electronic Noses

applied to evaluate signals for gas sensor array and used for quantitative multi-component analysis [10]. This type of technique shows good results when applied to binary or tertiary gas mixtures (n = 2 or 3) using an array of sensors (n > 3), however the calibration method becomes impossible when working with complex odor samples that may contain tens or even hundreds of different gases or components. Conse- quently, most research has focused on the use of qualitative types of classification methods for EN data, such as discriminant analysis and cluster analysis.

6.2.2 Linear Discriminant Analysis (LDA)

In DFA, a parametric pattern analysis method, it is first assumed that the data are

multinormal-distributed and then the discriminant functions Zp are determined. The set of discriminant functions Zp is calculated from the variables by separating the odor classes, finding the linear combination of the independent sensor responses

Xij in following equation:

Zp ¼ a1pX1j þ a2pX2j þ ::: þ aipXij þ ::: þ anpXnj ð6:3Þ

The coefficients aip are determined so that the F-ratio on the analysis of the variance is maximized subject to Zp being uncorrelated with Zp:::Zp1 within groups. Once the regression coefficients aip have been computed on the known data, following supervised learning, then they can be used to form the classification functions, which predict the group membership of unknown response vectors (referred to as cross-valida-

Fig. 6.4 Results of linear DFA on the analysis of three commercial roasted coffees using a 12-element tin oxide EN. (Reprinted from ref. [12], Elsevier Science, with permission.) 6.2 Statistical Pattern Analysis Techniques 141 tion). There are many ways of performing DFA, but the classical approach is LDA, for which a straight-line hyperplane passing through the data is found using different criteria [11]. However, sometimes, overlap occurs between classes and so there is no exact or crisp cut-off value. LDA has been applied to the discrimination of commercial coffee flavors, as shown in Fig. 6.4, and alcohol vapors with almost 100 % success rate [12]. Figure 6.4 shows the results of applying DFA to the response (fractional change of conductance) of 12 tin oxide gas sensors sampling the headspace of three different coffees. Plots of the first two discriminant functions show reasonable separation of the three classes. The observed classification rate was 81 % when half of the data was used for cross-validation. Other more advanced models have been developed, including quadratic or logistic discrimination that require some assumptions about the original data, but provide better discrimination performance. In Shaffer et al. [13], two examples of LDA are presented. The first one, the Mahalanobis linear discriminant analysis (MLDA) classifier is based on the Mahalanobis distance metric, it is trained by computing a mean vector for each class and the pooled covariance matrix in order to define the class boundaries. To classify a new pattern (Xj), theq Mahalanobisffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi distance to the mean vec- 0 1 tor (Xm) for each class is computed as: djm ¼ ðXj XmÞ S ðXj XmÞ, where d is the distance between pattern vector j and the mean pattern vector for class m and S1 is the inverse of the pooled covariance matrix (estimate of the common covariance of the classes). The classification of the new pattern is assigned to the classification of the closest mean vector, i.e. smallest d. However, the use of the pooled covariance matrix implies that the covariance matrices for each class are not significantly different. The second method, the Bayes linear discriminant analysis (BLDA) is based on the Bayes strategy for minimizing the risk associated with the classification decision. The training is performed using the mean vector for each class, and the pooled covariance matrix to position a linear separating surface. The assignment of class membership for a new pattern is determined by the side of the discriminant in which it lies using a scalar dot product of the pattern with each linear separating surface.

6.2.3 Principal Components Analysis (PCA)

PCA is a linear unsupervised method that has been widely used by various researchers to display the response of an EN to simple and complex odors (e.g. alcohols, beers, coffees). It is a multivariate statistical method, based on Karhunen-Loe´ve expansion, used by classification models to produce qualitative results for EN PARC. The method consists of expressing the response vectors Xj in terms of linear combinations of orthogonal vectors along a new set of coordinate axes, and is sometimes referred to as vector decomposition and thus helps to display multivariate data in two or three dimensions. Along the new axes the sample variances are extremes and uncorrelated so that an analysis in terms of principal components can show linear interdependence in data. Each orthogonal vector, principal component, accounts for a certain amount of variance in the data with a decreasing degree of importance. The scalar product of the 142 6 Pattern Analysis for Electronic Noses

orthogonal vectors with the response vectors gives the value of the pth principal component:

Zp ¼ a1pX1j þ a2pX2j þ :::aipXij þ ::: þ anpXnj ð6:4Þ

The variance of each principal component score, Zp, is maximized under the constraint that the sum of the coefficients of the orthogonal vectors or eigenvectors

ap ¼ðaip; :::; ajp; :::; anpÞ is set to unity, and the vectors are uncorrelated. The corresponding eigenvalues give an indication of the amount of information the respective principal components represent. The eigenvector associated with the largest eigenvalue has the same direction as the first principal component. The eigenvector associated with the second largest eigenvalue determines the direction of the second principal component. Since there is often a high degree of sensor co-linearity in EN data, the majority of the information held in response space can often be displayed using a small number of principal components. PCA is in essence a data dimensionality reduction technique for correlated data, such that a two- or three-dimensional plot can describe an n-dimensional problem. It can be applied to high dimensional data-sets to explore the nature of the classification problem in gas sensor applications and determine the linear separability of the response vectors. However, if the sensor output parameters are not linear, the results of PCA are not straightforward and the interpolation of features may be suspect, sometimes referred to as ‘pure created artifacts’. PCA is a linear technique that treats all sensors equally, thus the sensors may unduly influence its performance [3]. Figure 6.5 shows the results of applying PCA to an array of four tin-oxide sensors when applied to aromatic headspace of bananas [14]. Since metal oxide sensors gen-

Fig. 6.5 Results of PCA analysis of the response of a four-element tin-oxide sensor based EN to bananas aromas, showing clusters of increasing ripeness, a–g. (Reprinted from ref. [14], IOP Publishing Ltd, with permission.) 6.2 Statistical Pattern Analysis Techniques 143

Fig. 6.6 Results of PCA analysis of the response of a 32 carbon- polymer composite sensor based EN to bacteria causing eye infections. (Reprinted from ref. [16], IOP Publishing Ltd, with permission.) erally respond in a similar manner (i.e. correlated), over 99 % of the variance is typically described by the first two principal components [15]. Seven clusters or categories are apparent and are associated to seven states of ripeness; from left to right the groups appear according to increasing ripeness. However, the occurrence of complex boundaries suggests that a non-linear classification method is needed in order to obtain good performance in terms of PARC, rather than linear methods. Figure 6.6 illustrates how well this technique works when analyzing the response of an array of 32 carbon-polymer composite sensors. The system is being used to identify six species of bacteria, commonly associated with eye infections [16]. Most of the variance in the data is explained by considering only the first principal component, which implies that the sensor responses are again highly correlated. It can be seen in Fig. 6.6 that six groups exist and are associated with the different bacteria species. Although PCA is useful as a tool with which to assess the performance of an EN, CA presented in the following section, is perhaps a more appropriate tool because it is an unsupervised technique for enhancing the differences between the response vectors.

6.2.4 Cluster Analysis (CA)

Clustering is the separation of a data set into a number of groups, called clusters, based on measures of similarity. CA is an unsupervised, non-parametric technique that is widely used to discriminate between response vectors in n-dimensional space by enhancing their differences. It is also used to identify clusters or groups to which unknown vectors are likely to belong. The goal is to find a set of clusters for which samples within a cluster are more similar than samples from different clusters. Commonly 144 6 Pattern Analysis for Electronic Noses

clusters are allowed to merge and split dynamically by the clustering algorithm. CA is a model-free qualitative analysis that generally undergoes an unsupervised learning phase. Four basic types of clustering methods have been identified [17]: hierarchical, optimization-partitioning, clumping, and density-seeking. Hierarchical and partitioning methods are the most popular. Hierarchical techniques calculate the multivariate distances d of each individual to all others, and then cluster them using a process of either agglomeration (bottom up) or division (top down). The agglomeration techniques, among which are nearest neighbor (NN), furthest neighbor, fusion and Ward’s method, assumes that all individuals start off being alone, i.e. in a group of one, the nearest groups are then merged and this process continues until all patterns form suitable groups. The partitioning technique works on the opposite principle, it assumes that all the individuals start in one group and then splits them into two, and so on until all are in a group of their own. Hierarchical techniques produce a structured tree or dendrogram depending upon the definition of the distance-metric d and the way closeness and proximity of individuals are defined. The grooping is based on the proximity of the vectors in feature space. To do so a multi-distance metric

dij is calculated between data points i and j according to the expression: ! = XN 1 N N dij ¼ ðXik XjkÞ ð6:5Þ k¼1

N is normally set to 2 and so the Euclidean (linear) metric is used, there seems to be little advantage to be gained from using a non-linear metric when analyzing most EN data. To classify a new pattern, the Euclidean distance between the new pattern and each pattern in the training set is computed. The proximity of all points relative to each

other is then found by computing a so-called similarity value Sij, such that:

Sij ¼ 1 ðdij=maxfdijgÞ ð6:6Þ

This is called complete linkage because the distance metric is divided through by the maximum separation between all data points. Thus the similarity value is zero for the furthest neighbors and close to unity for the nearest neighbors. Other definitions can be considered for the similarity value but the choice of metric and linkage has a mar- ginal effect on the results. Many techniques exist, such as the one that links together groups in which the average distance, median distance, or distance between centroids is small enough, (Ward’s method and k-nearest neighbors). The proximity can be represented by plotting either the multivariate distance d or the similarity index S.It should be mentioned that the Euclidean distance can sometimes produce unexpected results unless the pattern vector is normalized (or scaled), so that CA is very sensitive to data pre-processing methods. Figure 6.7 shows the results of a CA (Euclidean metric, complete linkage) of the response of a metal-oxide EN to different alcohols [15]. The dendrogram connects up response vectors with the nearest similarity value and thus illustrates how the odors are interrelated. CA is a method easy to use and rapidly provides the user with pertinent information, and is widely used in the field of EN pattern analysis. PCA is used to identify groups or 6.3 ‘Intelligent’ Pattern Analysis Techniques 145

Fig. 6.7 Dendrogram showing results of CA on responses of 12-element tin oxide EN to five alcohol samples, resulting in clusters A, B, C, D, E. (Reprinted from ref. [15], Elsevier Science, with permission.)

clusters of points in feature space. However, the nature of EN data is such that it is often desirable to use a more powerful pattern analysis method. Typically, a method is required that not only copes with non-linear, non-parametric data but also generates a metric, which can adapt locally to regions of closely-packed response vectors and so give improved predictive performance. This has led to the rapid and widespread application of ANN to the analysis of patterns generated by EN. More ‘intelligent’ techniques will be considered in the following section.

6.3 ‘Intelligent’ Pattern Analysis Techniques

The nature of EN data is such that it is often desirable to use a more powerful PARC method that is able to cope with non-linear data, and has further advantages, over more conventional methods, such as learning capabilities, self-organizing, generalization and noise tolerance [18]. When the objective is to develop an EN that mimics the human olfactory system there is always an intellectual appeal to work on physiologically 146 6 Pattern Analysis for Electronic Noses

inspired PARC systems that imitate the human brain by learning from patterns. Re- cent interest in learning from data has resulted in the development of biologically motivated methodologies such as ANN, FIS, SOM, GA, NFS, and ART. ANNS, sometimes called neurocomputers, consist of parallel interconnected and usually adaptive processing elements that are attractive as they, to a certain extent, mimic the neuro- biological system [1]. The processing elements represent the biological brain cells or neurones, and their interconnections, the synaptic links. The pattern recognition ability of ANNS is potentially higher than the classical PARC paradigms described previously, due to parallel signal processing and great tolerance to sensor drift and noise. For a historical review of ANNS, see for example Haykin [19], the following section gives a comprehensive review of ANN-based EN systems and applications.

6.3.1 Multilayer Feedforward Networks

ANN generally give results quickly, are efficient with information processing, and learn by presenting examples; however it is sometimes difficult to choose the optimal network parameters and training procedures. Recently, ANNS have been widely used in odor recognition and many different ANN paradigms have been applied in this context. Since three-layered networks have sufficient computational degrees of free- dom to solve any classification problem [20], most EN workers have adopted this topology of network for implementing MLPs. Other feedforward networks can be used and the main ones are presented in this section, these include RBF and probabilistic neural networks (PNN). MLP, as a three-layered feedforward back-propagation (BP) trained network, is the most popular arrangement of neurones in odor classification and was the first one to be applied to EN [1]. In a network, the processing elements are organized in a regular architecture of three distinct groups of neurones: input, hidden, and output layers. Only the units in the hidden and output layers are neurones and so a MLP has two layers of weights. The number of input nodes is typically determined to correspond to the number of sensors in the array. The number of neurones in the hidden layer is determined experimentally and the number of odors analyzed generally determines the number of output neurones. When using a one-in-N coding scheme, there is one output neurone for each potential odor class. There are more efficient coding schemes but this is the simplest. A MLP has a supervised learning phase, which employs a set of training vectors, followed by the prediction, test or recall phase of unknown input vectors. Figure 6.8 shows the topology of a network used to identify five alcoholic odors using a twelve-element tin-oxide sensor EN [1]. MLP with BP learning algorithm has been applied to the prediction of bacteria type and culture growth phase using an array of six different metal-oxide semiconductor gas sensors [21]. Re- sults show that the best MLP was found to classify successfully 96 % of unknown samples on the basis of 360 training vectors and 360 test vectors. Using BP to train the network, it is necessary to provide it with a number of sample inputs (training set) with their corresponding target outputs. Each neurone computes 6.3 ‘Intelligent’ Pattern Analysis Techniques 147

Fig. 6.8 Structure of a fully connected three-layer (layer i are o, 1 and 2) backpropagation network used to process data from a 12-element in oxide EN for five alcoholic odors. (Reprinted from ref. [1], Institute of Physics Publishing, with permission.)

its weighted inputs and performs a non-linear transformation of this sum using a given activation function, for example a sigmoid transfer function which constrains the output to a value between [0,þ1] or [1,þ1]. Considering a neurone h, with n inputs, [1,…,i,…,n] and an input vector j, the summation function ajh multiplies and sums the input signals Xij with associated adaptable weights whi considering a fixed weight called a bias, hh0, which is then transformed by a non-linear activation function f(.) (e.g. sigmoid) to produce the single output zjh; the overall computation follows: ! Xn

zjh ¼ f ðajhÞ¼1=ð1 þ expðajhÞÞ ¼ f ðwhiXij hh0Þ ð6:7Þ j¼1

The calculation is carried out for each neurone and each layer feeding the values through to the output layer, forward pass. During this learning phase, the weights are adjusted to minimize the difference between the actual output zjh and the ideal or target output tjh for the considered input vector j using the expression djh ¼ zjh tjh. The error term is often called delta, and the widely used parameter-updating scheme is known as the delta learning rule [22], the component difference vector is calculated using the expression djh ¼ðtjh zjhÞð1 tjhÞ. In the backward pass computations; the stochastic approximation procedure updates synaptic weight values during each presentation of the jth training sample on each iteration (or epoch) s using, for example, the gradient descent method with momentum:

ðsÞ ðs1Þ ðsÞ ðs1Þ ðsÞ ðs2Þ wkh ¼ wkh þ Dwkh ¼ wkh gdjhzjh þ lDwkh ð6:8Þ

ðsÞ Basically, the new set of weights, wkh , is made of a combination of the old weight ðs1Þ ðsÞ values, wkh (from the previous epoch) and a weight update or delta, Dwkh . The change in weights is based on two parameters for this example: 148 6 Pattern Analysis for Electronic Noses

1) g, the learning rate, a small positive number (default is generally 0.9) that determines the rate of convergence to a solution of minimum error. 2) l, the momentum term, a small positive number (default is generally 0.5) is often added to improve the speed and stability of the learning.

Thus the change in adjustable synaptic weights is proportional to the error and to the activation of the input unit. Using BP, the weights and biases associated with the neurones are modified to minimize the mapping error, when stabilized, the network is said to be trained. The total sum squared error can be used to measure the network performance. The updating procedure is repeated for a number of epochs until the network error has fallen to a small constant level. Once the network is trained, it can be used to predict the membership of novel, unseen and untrained samples in a validation set. The classification of new patterns is accomplished by propagating the new pattern through the network and the output neurone with the highest score indicates the class. The success of the training process, in terms of a fast rate of convergence and good generalization, can be affected by the choice of architecture and initial parameters (e.g. learning rate and initial weights). Various learning paradigms are available to train a BP MLP network; Boilot et al. [16] used both gradient descent with momentum and Levenberg-Marquardt variations as supervised learning algorithms. For all architectures of the MLP networks tested, the latter paradigm outperformed gradient descent. Since architecture and parameters are to be determined experimentally, much time may be spent searching for the optimal ANN. An often employed rule- of-thumb is to set the number of inputs equal to the number of sensors or the number of extracted features considered for the sensor array, the number of output nodes not greater than the number of species or compounds to be discriminated (in a one-in-N coding), and a hidden layer not larger than the largest of the two other layers. It is also recommend having twice as many training vectors as there are weights in the network developed in order to achieve good generalization. An alternative method to optimize the ANN design is to use a GA to determine automatically a suitable network architecture (e.g. growing or pruning the network) and a set of parameters (e.g. learning rate, momentum term) from a restricted region of design space [23]. GA are heuristic search algorithms based on the mechanics of natural selection. The structure and parameters of the neural network, learning rate, initial weights, number of layers, number of neurones per layer, and connectivity, are coded using binary strings, which are concatenated to form chromosomes. GA are then applied to search populations of chromosomes using defined typical genetic operators such as parent selection, crossover and mutation. The performance of the network

represented by each chromosome ci is evaluated using a fitness function; FðciÞ¼auðciÞþb where F is the fitness function, u is the objective function to optimize, and a and b are transformation parameters that are dynamically adjusted to avoid premature convergence. The objective function is generally a weighted sum of the various performance measures. In the sensor data classification problem, the performance measures used in the objective function are based on, for example, the network 6.3 ‘Intelligent’ Pattern Analysis Techniques 149 prediction error, speed of convergence, size of the network and degree of generalization [24]. The probabilistic neural network (PNN) operates by defining a probability density function (PDF) for each data class based on the training data set and an optimized kernel width [13]. A multivariate estimate of the PDF for each class can be expressed as the sum of individual training pattern Gaussian-shaped kernels. The PDF defines the boundaries for each data class, while the kernel width determines the amount of interpolation between adjacent kernels. The classification of new patterns is performed by propagating the pattern vector through the PNN; the input layer is used to store the new pattern while it is serially passed through the hidden layer. The dot product distance between the new pattern and the training set pattern stored is computed at each neurone in the hidden layer. The summation layer consists of one neurone for each data class and sums the outputs from all hidden neurones of each respective data class. The products of the summation layer are forwarded to the output layer where the estimated probability of the new pattern, being a member of that data class, is computed. RBFs are attractive when other ANN methods fail to get a good classification due to a significant difference between classes in terms of shape, volume or density, of overlapping classes. RBF networks are supervised learning paradigms very similar to MLP except that they use radial basis transfer functions for the hidden layer rather than linear or sigmoid ones. Hence they classify data using hyper-spheres rather than hyper-planes [25]. The purpose of RBF is to allow the screening of the input space with overlapping receptive fields. The non-adaptive RBF is a fast two-stage training procedure using a hybrid-learning rule:

1) Unsupervised learning in the input layer for the determination of the receptive field centers and widths. 2) Supervised learning of weights in the output layer simply using the delta learning rule via linear least squares.

Hence RBF implementations differ mainly in the choice of heuristics used for selecting basis function centers and widths. For example, taking every sample as a center (may result in over-fitting), selecting centers as representative prototypes using the generalized Lloyd algorithm (GLA) and Kohonen’s SOMs, or adding new basis functions centered on one of the training samples sequentially. Although RBF networks do not provide error estimates, they have an intrinsic ability to indicate when they are extrapolating since the activation function of the receptive fields is directly related to the proximity of the test pattern to the training data. RBF are becoming more and more popular for EN pattern analysis. However, one of the main difficulties when using this type of system is the determination of the optimal architecture – the number of hidden nodes necessary to achieve a good classification. Boilot et al. [16] report on the use of RBF for the prediction of bacteria causing eye infections. Although RBF networks classify bounded regions of sensor space, this can make them more sensitive to sensor drift and so less robust; this is a trade-off between model accuracy and robustness. 150 6 Pattern Analysis for Electronic Noses

6.3.2 Competitive and Feature Mapping Networks

There are many different types of neural network but the ones considered in this chapter are those that have been applied to EN data. One of them is a single-layer neural network with competition, such as Kohonen’s self-organizing map. Competi- tive layers are used in ANN to improve the discrimination process and, unlike traditional network layers, there are connections between the neurones within a layer. The basic principle is that competition enhances the difference between the level of activation of the neurones, sometimes in extreme cases the ‘winner-takes-all’ and one neurone only is allowed to be switched on. A Hamming network is a fixed-weight competitive ANN where the lower network feeds into a competitive network called maxnet. It uses a maximum likelihood classifier, based on the measure of similarity as in statistical clustering technique, defined as n minus the Hamming distance between the input unknown odor vector and the exemplar reference vectors. An N-tuple competitive network was used to classify the responses of a 12-element MOS odor sensor array to both a set of alcohols and to a set of different blends of roasted and ground coffee beans [26]. In this case, the neural network outperformed statistical linear discriminant function analysis with a success rate of 87 %. Another competitive network that has been applied to EN data is the self-organizing neural network or Kohonen network [27]. The SOM algorithm was developed by Ko- honen to transform an incoming signal pattern of arbitrary dimension into a one- or two- dimensional discrete map. SOM is more closely related to the neural structures of the human olfactory cortex than other neural networks presented before because it emulates parts of the brain. SOM applied to EN systems typically contain a two-dimensional single layer of neurones in addition to an input layer of branched nodes. If the system is left for learning in the environment of interest, the learning algorithm of the network processes the sensor outputs step by step, and constructs an internal representation of the environment [27]. SOM accumulate a lot of statistical information in an unsupervised fashion, using a competition layer in the form of a Kohonen organizing map so that all weight vectors of the winner and adjacent neurones are updated. They are interesting for EN systems because of their inherent features such as dimensionality reduction and invariance to drift and transitory noise [28]. We assume here that there are m neurones in the Kohonen neural layer, typically arranged as the knots of a square lattice, and each one has a parameter weight vector VðlÞ of dimension n, which is the same as the input feature vectors (i.e. the number of vectors). A vector describes each neurone in this layer so that the vector components are the knot coordinates in the lattice. The weight vectors are randomly initialized at

the beginning. One input vector Xi is selected from the dataset and put into the net- ðlÞ work, so that the distances between Xi and each V are computed using the components:

Xn 2 ðlÞ 2 dil ¼ ðXij Vj Þ l ¼ 1; :::; mj¼ 1; :::; n ð6:9Þ j¼1 6.3 ‘Intelligent’ Pattern Analysis Techniques 151

The minimum distance dil* is then determined to obtain the neurone l that is the winner over the others. In a winner-takes-all strategy, only the winning neurone ðl*Þ ðl*Þ ðl*Þ weights are updated using Vnew ¼ Vold þ gðXi Vold Þ where g is the step gain or learning rate, whereas all other neurones keep their old weights. In another strategy, all neurones s that are close to the winner are updated using ðsÞ ðsÞ ðsÞ * 2 2 Vnew ¼ Vold þ ghsl* ðXi VoldÞ. hsl* ¼ exp ks l k =2r is called the excitatory response and is only appreciable for the neurone that coincides with l* and its neighbors. r is the length scale of the proximities to l and is generally fixed to a value in the range of 2 to 5 lattice units. It is desirable that after a number of iterations the weights no longer change, and therefore the map is able to stabilize asymptotically in an equilibrium state, with g decreasing to zero. In a supervised learning scheme, the SOM is provided with the desired output functions; it is called learning vector quantization (LVQ) and integrates supervised learning techniques in a self-organizing feature map [29]. It combines some of the features of nearest neighborhood and competitive learning to define a smaller set of reference vectors that span the same space as the original training set pattern. Figure 6.9 shows a schematic diagram of a LVQ network, the hidden layer in the network is a Kohonen layer, which does the learning and classifying. The LVQ scheme has phases that consist of LVQ1 and LVQ2 algorithms. LVQ1, is the basic LVQ learning algorithm, which helps all processing elements to take an active part in the learning. LVQ2 is a fine- tuning mechanism, which refines class boundaries. Therefore the output from LVQ2 is the final encoded version of the original input signal applied to LVQ1. The number of training patterns to ensure equal accuracy to other approaches could dramatically decrease because the given calibration data set is not the unique source of

Fig. 6.9 Schematic diagram of LVQ with Kohonen a layer 152 6 Pattern Analysis for Electronic Noses

information collected by the system during unsupervised learning. However, an important limitation of this approach is that lengthy computation is required when applied to real problems. SOM have been applied to a wide variety of applications including, with some degree of success, classification of odors and patterns generated by an EN [30]. Hines et al. [28] used the supervised Kohonen SOM on the alcohol and coffee data sets and found good performance results in terms of both accuracy and generalization. Shin et al. [31] used LVQ to classify the strain and growth phase of cyanobacteria using a 6-element EN with excellent results. When trained on two classes of a gas mixture, after a short period of time the weights of the network appear to be strictly correlated with the assigned classes. The network does not have any direct information about the classes, except for the sensor outputs [32].

6.3.3 ‘Fuzzy’ Based Pattern Analysis

Fuzzy set theory (FST) was invented by Zadeh [33] to provide a mathematical tool for dealing with the linguistic variables and imprecise language used by humans (for example hot, cold, slow, quite slow). A fuzzy set is defined as a set whose boundary is not sharp. Fuzzy logic has been applied to EN pattern analysis and attempts have been reported to use fuzzy functions in order to identify odors. FST is therefore attractive in the field of machine olfaction in which odor samples are described by olfactory descriptors, such as peppery, floral, or sweet, and intensity attributes, such as quite, very, or strong. Gardner and Bartlett [3] describe three principal approaches when fuzzifying the EN classification problem:

1) Sensor space can be defined using fuzzy functions. 2) The pattern recognition algorithm can be fuzzified. 3) Classification space can be defined using fuzzy functions.

Fuzzy clustering essentially deals with the task of splitting a set of patterns into a number of classes with respect to a suitable similarity measure of the pattern belonging to a given cluster. Fuzzy clustering provides partitioning results with additional information supplied by the cluster membership values indicating different degrees of belongingness. Fuzzy clustering can be precisely formulated as an optimization problem of class centers and spreads. The fuzzy c-means (FCM) algorithm, for example, provides an iterative approach for this optimization. Most of the FCM approaches to EN pattern analysis need to be given the correct number of clusters but can prove to be very attractive for finding patterns in data sets or can even be applied to clusters extracted from data with PCA. Yea et al. [34] used fuzzy logic to fuzzify sensor space by assigning the steady-state voltage of three gas sensors to one of the three odor classes, giving an excellent classification rate. Another approach is to use fuzzy logic to fuzzify the neuronal weights and weight calculations in a multi-layer neural network. Conventional networks are trained using 6.3 ‘Intelligent’ Pattern Analysis Techniques 153 randomly initiated weights, which may be a problem for the overall training process. This is because the search for the best set of weights to both classify the training patterns and identify new ones usually starts from a poor point that may never reach the desired optimal point. On the other hand, a suitable starting point, depending on the nature of data, is desirable to speed up the process and reduce the likelihood of settling in local minima. A type of fuzzy neural network (FNN) can be used to make use of possibility distributions to determine the initial weights using membership- class restrictions imposed on a variable defining the range of values [35]. Possibility distributions, based on fuzzy logic theory, are often triangular and so they are similar in shape to normal distributions with the means having the highest possibility of occurrence. In FNN, the signal conditioning that occurs during fuzzification and defuz- zification translate many properties of overlapping sensor arrays into parameters that are better handled by the classifier. In Singh et al. [36], the use of fuzzy neuronal tree computing is reported when used on coffee and tainted-water data from an EN. Their version of a FNN proved to be better than classical ANN. Ping and Jun [37] used a combined neural network (RBF) with a fuzzy clustering (FCM) algorithm and were able to demonstrate the unusual effectiveness and the good recognition performance of their method. FNN are becoming more and more popular and represent a considerable potential improvement in the analysis of certain EN problems. ART was introduced as a theory of human cognition in information processing [38] and it is based on the fact that the human brain can learn new events without necessarily forgetting those learnt in the past. ART networks are intelligent systems that are capable of autonomously adapting in real time to changes in the environment, and that are stable enough to incorporate new information without destroying the memories of previous learning. ART networks have been applied to metal-oxide sensor based EN with results very similar to those obtained with BP trained MLP, but with a shorter training time on small data sets [39]. Carpenter et al. [40] introduced Fuzzy ARTMAP for incremental supervised learning and non-stationary PARC problems. Fuzzy ARTMAP carries out supervised learning, like BP MLP, but it is self-organizing, self-stabilizing and suitable for incremental learning. It can deal with uncertainty or fuzzy data, a key element in many measurement systems and generally shows superior performance in learning compared with MLP, exhibiting fast learning for rare events. Figure 6.10 shows the schematic architecture of a Fuzzy ARTMAP neural network that consists of two ART modules interconnected by an associative memory and internal control structures. The orienting subsystem is responsible for generating a reset signal while the gain control sums the input signal. One of the main advantages of Fuzzy ARTMAP is that it is able to perform real-time learning without forgetting previously learnt patterns and so there is potentially no off-line training phase like MLP. This is very important from a practical point of view because the data-set used to train the network may be increased during the development phase by adding new measurements. Some earlier work by Llobet et al. [41] showed that Fuzzy ARTMAP is a promising technique for EN data analysis. Llobet et al. used it to analyze the state of ripeness of bananas and obtained results that exceeded those for MLP [14]. Shin et al. [31] used it to classify the strain and growth phase of bacteria and once again it outperformed MLP. 154 6 Pattern Analysis for Electronic Noses

Fig. 6.10 Architecture of a Fuzzy ARTMAP neural network

6.3.4 Neuro-Fuzzy Systems (NFS)

Both FIS and ANN are branches of an emerging research area of artificial intelligence called soft computing. This approach can be used, with some restrictions, as non- algorithmic model-free (i.e. heuristic) estimator for data processing purposes. Fuzzy systems can be built to express knowledge in the form of fuzzy linguistic If-Then rules and perform some fuzzy clustering analysis, while neural networks can be used to learn from data and perform pattern recognition and classification. NFS are one of the most promising approaches that have been developed to deliver the benefits of both and overcome their limitations, combining or fusing these two complementary techniques into an integrated system [42]. Boilot et al. [43] report on several software-based hybrid neuro-fuzzy systems used for specific real world applications linked to data processing. They focus on the extraction of knowledge from a representative data set of alcohol test vectors, collected using a 12-element metal-oxide EN. The paper also introduces a classification scheme for grouping the various software, dis- cussing their merits and demerits, drawing upon a comparison of delineated criteria for evaluating their efficacy (i.e. performances) and interpretability (i.e. semantics). Using these techniques for data exploration, the results from NFS-based EN may be viewed with more confidence because they provide a better representation of the information embedded within data-sets. Users will find it helpful to generate NFS in the context of extracting knowledge from EN data-sets, and representing it as a clear and interpretable set of fuzzy rules. The exploration of EN data and pattern analysis using ‘intelligent’ systems has so far mainly been done using ANN, yet when they perform a classification of various odors they give little or no insight into the true nature of the data. Using NFS for data processing and exploration does not only provide an opportunity to discover unknown dependencies and relationships, but also allows us to present them as a set of rules that are more interpretable than the weight matrices returned by ANN. 6.4 Outlook and Conclusions 155

6.4 Outlook and Conclusions

An EN detects simple and complex odors using an array of non-specific chemical sensors. Essentially, each odor generates a characteristic fingerprint or smell-print of responses from the sensor array and so known odors can be used to build up a database and used to train a pattern recognition system. It is impractical to have specific sensors when an odor may contain hundreds or even thousands of compounds, and so the solution is to use a PARC system to classify smell-prints or patterns. PARCs are therefore a critical component in the successful implementation of ENs. The objective of pattern analysis is to train or configure the recognition system in order to produce unique classifications, or clusterings, of each odorant so that automated identification can be implemented. Many different pattern analysis techniques have been applied to EN patterns in recent years. In this section we summarize the various considerations relating to EN pattern analysis. First, we discuss the basic criteria for the comparison of the various PARC paradigms with respect to both quantitative and qualitative pattern analysis. Next, we discuss sensor modeling and ‘intelligent’ sensor systems. Finally, we draw some conclusions regarding the application of pattern recognition to ENs.

6.4.1 Criteria for Comparison

Compared to other applications, chemical sensor array pattern recognition or EN system pattern analysis has a unique set of requirements and needs [13]. The pattern dimensionality for a sensor array (typically < 40) is considerably smaller than for many other applications of PARC in science and engineering (e.g. spectroscopy or chromatography), thus the computational load on the grouping algorithm and the resources needed to learn the classification rules are greatly decreased. Therefore, many of the accepted procedures that are used in traditional pattern recognition and chemometrics in general may not be pertinent or relevant when applied to EN pattern analysis in particular. EN are expected to be operated in various types of environments and situations, and the pattern analysis paradigm should be able to cope with these conditions. For example, when a system is used in field measurements, additional challenges not seen in the laboratory or a controlled environment are likely to occur, and the system is still expected to detect and identify the target analytes while in the presence of large concentrations of unknown interfering species. As suggested by Shaffer et al. [13], there are a few criteria or qualities that an ideal pattern recognition algorithm should have, such as accuracy, speed or ability to cope with uncertainty.

* High accuracy. For application of an EN in the field, the PARC algorithm must accurately classify new patterns, with a low false alarm rate (true negative) and ideally no missed detection (false positive). For military applications, such as detection of toxic chemical vapors, classification rates should be higher than 90 % accuracy even for low concentration of compounds. 156 6 Pattern Analysis for Electronic Noses

* Fast speed. When used in real-time applications, the PARC algorithm must be able to classify patterns quickly, so that computationally intense paradigms are not practical. * Simple to train. The classification rules and the classification itself must be learned quickly and the training patterns database of the system will need to be updated periodically, therefore the paradigm should be able to ‘relearn’. This procedure must be performed as simply and quickly as possible, keeping the learning outcome simple for the user to be able to understand it. * Low memory requirements. In field applications, the hand-held EN requires on board pattern analysis, so the algorithm should be able to be transferred, embedded and run on a simple micro-controller with limited memory resources. Thus high computational power and large memory requirement algorithms are not appropriate for field units. * Robustness to outliers. When used in uncontrolled environments, the PARC algorithm must be able to differentiate between sensor signals it was trained on and those that it was not, recognizing all the important compounds and ignoring parasitic, noisy or ambiguous signals. * Produce a measure of uncertainty. For most applications, the PARC paradigm must be able to produce a measure of the certainty of the classification results, expressed either as a percentage, a confidence factor or a category.

Unfortunately, no PARC algorithm is able today to meet all of these requirements, but researchers, in an attempt to determine the optimal classifier, have performed comparative studies. The qualitative comparison performed by Derde and Massart [44] on several popular chemometric classifiers focused on technical aspects, such as optimal decision boundaries, overlapping regions, degree of uncertainty and outliers, and practical aspects, such as updates, variables of mixed type, irrelevant parameters and ease of use, of supervised PARC. They conclude with a confirmation of the need for an application specific choice of algorithm and the potential that hybrid approaches can bring. The book published by Michie et al. [45] is probably the most comprehensive and complete comparison study published as they present 23 types of machine learning, statistical and neural-classification methods and conclude by presenting the relative merits and demerits, and on the choice of an appropriate algorithm for a given application. The book of Cherkassky and Mulier [4] provides a treatment of the principles and methods for learning dependencies from data using statistics, neural networks and fuzzy logic oriented around case studies and examples. It also provides a detailed description of the new learning methodology called support vector machines (SVM). To date, a comparison study published by Schaffer et al. [13] is the only one on EN data. It focuses on qualitative criteria together with one quantitative measurement, namely, the classification accuracy, and proposes the use of a combination of LVQ and PNN in order to exploit the advantages of both methods. The study of NFS for EN data processing presented by Boilot et al. [43] again reinforces the potential of hybrid techniques and their practical implementation on micro-controllers. 6.4 Outlook and Conclusions 157

6.4.2 Intelligent Sensor Systems

The modeling techniques used so far in the context of ‘intelligent sensor’ systems aim to enhance the sensor selectivity, reduce the time necessary for calibration, and counteract drift [18]. A careful exploration or analysis of the system is required before applying any dynamic model; unimportant sensors should be discarded using, for example, PCA loadings at this preliminary data exploration stage.

* Enhancing sensor selectivity. To date, models using parameters estimated from the transient sensor responses can enhance selectivity. These parameters may be related to the physical and chemical properties of the sensing material and thus are based on physical models giving some insights into the dynamic behavior of the sensor. However, transient signals can be influenced by previous measurements (short-term memory effect), by drift due to ageing of the system, or variations in ambient temperature and humidity, so that models that do not consider these issues will deteriorate over time. * Calibration time reduction. Some applications of time-varying sensor signals offer a reduction in the time necessary to calibrate the sensor array to odors of interest. Results with ARMA and ad hoc multi-exponential models applied to the dynamic response of tin-oxide sensor arrays have been reported [46]. In this application, the prediction of the static response from the initial part of the dynamic response permits a reduction of the calibration time by a factor of four. * Response models. Dynamic measurements are interesting when changes in either the odors or conditions are of the same time-scale as the sensor response. The correlation approach is a modeling method used to deal with noise, calculating linear systems impulse response and non-linear systems Weiner kernels. How- ever, models constructed using black-box models based on input-output data only, do not give enough insight into the inner structure of the sensor and the model cannot be defined in terms of physical and chemical properties. On the other hand, block-structured models are more related to the intrinsic characteristics of the sensing mechanisms. The use of non-parametric approaches (e.g. cross- correlation) to estimate the impulse response with low errors requires long data sequences and can be rather time-consuming. * Drift counteraction. All approaches described include memory effects and thus can assess the problem of short-term drift. Long-term drift caused by sensor poisoning (or system ageing) implies non-stationary measurements with which most of the techniques, apart from ANN, cannot cope. SOM with residual plasticity can help to maintain the PARC ability of a sensor system affected by drift. ARTMAP and Fuzzy ARTMAP contain a self-stabilizing memory that permits accumulating knowledge of new events in a non-stationary environment; the short term memory gives the network some plasticity to adapt to sensor drift, while the long term memory gives the necessary rigidity to avoid forgetting previously learnt patterns [47]. 158 6 Pattern Analysis for Electronic Noses

6.4.3 Conclusions

For most scientists working in the field of EN systems, the two most commonly used pattern analysis techniques are first PCA to display known odors, explore how the data cluster in the multi-sensor aroma space and assess their linear separability, and MLP BP trained ANNS to provide a predictive classification of unknown odor vectors. How- ever, PCA can only be used to give a linear representation of the clusters and not for classification purposes, moreover its outcomes can sometimes be criticized. BP trained ANNS have arguably been the most successful for many applications to date that focus upon the discrimination of quite dissimilar simple or complex odors, or the staling of a specific complex odor. MLP, even if it is a powerful non- linear classification paradigm and has proven to perform well with EN data, can sometimes fail to achieve high levels of correct classification, moreover it is difficult to interpret its results as the system appears as a black box to the user. Among the algorithms presented in this chapter, neural network approaches (MLP, RBF, PNN, LVQ) are the most accurate classifiers and can cope with overlapping clusters observed with linear techniques. For these applications there is no need to use more complex or sophisticated PARC techniques, and this is why so many commercial EN available today provide a standard BP network as the predictive classifier. Nowadays, researchers have turned to more reliable and advanced techniques to perform pattern analysis for field EN and handheld units, such as cluster analysis based on fuzzy clustering or nearest neighbors. Even in the field of neural networks, the performances of the popular MLP are often outperformed by LVQ and RBF networks in terms of sensitivities and specificities. These two techniques together with other forms of self-organizing techniques are being seen as the benchmark for predictive classifiers in EN applications. It is our belief that the best strategy to perform pattern analysis on EN data is to employ algorithms that can cope, up to a certain extent, with a degree of fuzziness like the human olfactory system and that presents attractive features. In this context, Fuzzy ARTMAP networks, for example, are very attractive for pattern classification in the context of field instruments because they are able to perform incremental learning and offer self-organizing and self-stabilizing potential. We believe that NFS will be increasingly used in more challenging EN applications because they include both fuzzy and neural capabilities and so produce a classification based on an understandable set of rules. However it is always dangerous to try and predict future events! First generation of commercial EN have existed since the early 1990s and now there are more than 15 manufacturers with applications covering food, cosmetic, environment and medical domains [48]. More sophisticated pre-processing and PARC methods are needed in more challenging applications of EN, such as detecting sub-ppm taints of components, and in the development of hand-held units. The PARC techniques employed in a hand-held EN are likely to mimic more closely the signal processing present in our own olfactory system. The next generation of EN are being developed in university laboratories and research institutions using more biologically inspired models of the olfactory system. They will need to be more flexible and able to work in less controlled environments, incorporating all the sensors, signal proces- 6.4 Outlook and Conclusions 159 sing and neuro-inspired models of olfaction to identify and analyze a wide variety of odors in a constantly changing background. However, we believe that it will be several years before dynamical neural networks are developed with the enormous discriminating power and sensitivity of our truly remarkable olfactory system. The human nose is a complex differential (i.e. adaptive) signal processor that can detect an increase or decrease in the intensity of a smell, and thus an EN mimicking it may require the use of sophisticated adaptive filter combined with fuzzy classification functions.

Acknowledgements Pascal Boilot gratefully acknowledges financial support from EPSRC (award number 99310943) and the University of Warwick during his stay and his studies as a PhD student. We thank our colleagues, students, etc. who have contributed directly or indirectly to this work. Finally we would like to thank Roger Granthier for proof reading this document.

References

1 J. W. Gardner, E. L. Hines, M. Wilkinson. 13 R. E. Shaffer, S. L. Rose-Pehrsson, Meas. Sci. Technol.. 1990, 1, 446–451. A. R. McGill. Analytica Chimica Acta, 1999, 2 P. E. Keller, SPIE proceedings series, 1999, 384, 305–317. 3722, 144–152. 14 E. Llobet, E. L. Hines, J. W. Gardner, 3 J. W. Gardner. P. N. Bartlett. Electronic Noses: S. Franco. Meas. Sci. Technol., 1999, 10, Principles and Applications, Oxford Univer- 538–548. sity Press, New York, 1999, Chapter 7. 15 J. W. Gardner, Sens. Actuators B, 1991,4, 4 V. Cherkassky, F. Mulier. Learning from 71–75. Data: Concepts, Theory, and Methods, Wiley, 16 P. Boilot, E. L. Hines, J. Spencer, J. Mitchell, New York, 1998. F. Lopez, J. W. Gardner, E. Llobet, M. Hero, 5 J. W. Gardner, E. L. Hines. in Handbook C. Fink, M. Gongora. In Electronic Noses and of Biosensors and Electronic Noses (Ed.: Olfaction 2000 (Eds.: J. W. Gardner and E. Kress-Roger), CRC Press, Boca Raton, K. C. Persaud), IoP Publishing, Bristol, 2000, 1997, Chapter 27. 189–196. 6 W. P. Carey, K. R. Beebe, B. R. Kowalski. 17 B. S. Everitt, Cluster Analysis, Heinemann, Anal. Chem., 1987, 59, 1529–1534. London, 1981. 7 B. F. T. Manly, Multivariate Statistical 18 E. L. Hines, E. Llobet, J. W. Gardner. IEE Methods, Chapman & Hall, London, 1986. Proc.-Circuits Devices Syst., 1999, 146(6). 8 F. Wold, Festschiftjerzyneymanv, Wiley, 19 S. Haykin S, Neural Networks: a Compre- New York, 1966. hensive Foundation, MacMillan Publishing 9 M. S. Nayak, R. Dwivedi, S. K. Srivastava. Company, New York, 1994. Sens. Actuators B, 1993, 12, 103–110. 20 R. P. Lippmann, IEEE ASSP Mag., 1987, 10 J. Mitrovics, U. Weimar, W. Go¨pel. 4(2), 4–22. Proceedings of the 8th International Conference 21 J. W. Gardner, M. Craven, C. Dow, on Solid-State Sensors and Actuators and E. L. Hines. Meas. Sci. Technol, 1998,9, Eurosensors IX, 1995, 1, 707–710. 120–127. 11 R. Fisher, Annals of Eugenics, 1936,7, 22 B. Widrow, M. E. Hoff. IRE WESCON 179–188. Convention Record, 1960, 4, 94–104. 12 J. W. Gardner, H. V. Shurmer, T. T. Tan. 23 A. K. Srivastava, K. K. Shukla, Sens. Actuators B, 1992, 6, 71–75. S. K. Srivastava. Microelectronics Journal, 1998, 29, 921–931. 160 6 Pattern Analysis for Electronic Noses

24 A. A. Fekadu, E. L. Hines, J. W. Gardner. In 38 G. A. Carpenter, S. Grossberg. Comput. Vis. Artificial Neural Nets and Genetic Algorithms Graph. Image Process., 1987, 37, 116–165. (Eds.: R. F. Albrecht and N. C. Steele), 39 J. W. Gardner, E. L. Hines, C. Pang. Meas. Springer-Verlag, New York, 1993, 691–698. Control, 1996, 29, 172–178. 25 S. Chen, C. F. N. Cowan, P. N. Grant. IEEE 40 G. A. Carpenter, N. Grossberg, N. Marku- Trans. on Neural Networks, 1991&/hf, 2, zon, J. Reynolds, D. Rosen. IEEE Trans. on 302–309. Neural Networks, 1992, 3, 698–713. 26 J. D. Mason, PhD thesis, University of 41 E. Llobet, E. L. Hines, J. W. Gardner, Warwick, Coventry, UK, 1994. P. N. Bartlett, T. T. Mottram. Sens. Actuators 27 T. Kohonen, Biol. Cybern., 1982, 43, 59–69. B, 1999, 61, 183–190. 28 E. L. Hines, J. W. Gardner, C. E. R. Potter. 42 C.-T. Lin, C. S. G. Lee. Neural Fuzzy Systems: Meas. and Control, 1997, 30, 262–268. A Neuro-Fuzzy Synergism to Intelligent 29 T. Kohonen, University of Technology, Systems, Prenctice Hall P T R, Upper Saddle Helsinki, Finland, 1986. River, 1995. 30 F. Davide, C. Di Natale, A. D’Amico. Sens. 43 P. Boilot, E. L. Hines, J. W. Gardner. In Actuators B, 1994, 18–19, 244–258. Sensors Update (Eds.: H. Baltes, J. Hesse and 31 H. W. Shin, E. Llobet, J. W. Gardner, W. Gopel), Wiley-VCH, Weinheim, 2000, E. L. Hines, C. Dow. IEE Proc. – Sci. Meas. Chapter 4. Technol., 2000, 147(4), 158–164. 44 M. P. Derde, D. L. Massart. Analytica 32 C. Di Natale, F. Davide, A. D’Amico. Chimica Acta, 1986, 191, 1–16. Sens. Actuators B, 1995, 23, 111–118. 45 D. Michie, D. J. Spielgelhalter, C. C. Taylor. 33 L. A. Zadeh, Information and Control, 1965, Machine Learning, Neural and Statistical 8, 338–353. Classification, Ellis Horwood, New York, 34 B. Yea, R. Konishi, T. Osaki, K. Sugahara K. 1994. Sens. Actuators A, 1994, 45, 159–165. 46 C. Di Natale, S. Marco, F. Davide, 35 M. M. Gupta, J. Qi. In Fuzzy Logic A. D’Amico. Sens. Actuators B, 1995, 24–25, for the Management of Uncertainty 578–583. (Ed.: L. A. Zadeh), John Wiley, New York, 47 G. A. Carpenter, N. Grossberg, J. Reynolds. 1992, 479–490. IEEE Trans. on Neural Networks, 1995, 6(6), 36 S. Singh, E. L. Hines, J. W. Gardner. Sens. 1330–1336. Actuators B, 1996, 30, 185–190. 48 J. W. Gardner, K. C. Persaud (eds.). Electronic 37 W. Ping, X. Jun. Meas. Sci. Technol., 1996, Noses and Olfaction 2000, Institute of Physics 7(2), 1707–1712. Publishing, Bristol, 2000. Part B Advanced Instrumentation

7 Commercial Electronic Nose Instruments

E. Vanneste, H.J. Geise

7.1 Introduction

For a long time the human nose has been an important tool in assessing the quality of many products, food products being good examples. Whereas all other parts of production processes, including these of the food industry, were becoming more and more automated, there was still no ‘objective’ means for using the ‘subjective’ information confined in the smell of products. In 1982, when G. Dodd and K. Persaud [1] of the Warwick Olfaction Research Group presented their much-celebrated scientific publication in Nature, it heralded the beginning of a new technology: artificial olfaction. The expression electronic nose (EN), however, appeared for the first time in 1988. Much research is being undertaken in order to find new and more diverse sensors while also improving the pattern recognition engines, and today there are several companies offering ENs. This chapter intends to give the reader a description of the individual companies, and explain the technology used. For a comprehensive and detailed description of the different sensor technologies and data-algorithms used in the commercially available equipment we refer to elsewhere in this book. References to previous reviews can be found here [2–8]. The term EN works as an advantage as well as a disadvantage for the development of the concept towards its applications. One might even venture to refer to the EN du- alism. The advantage is that the expression immediately evokes associations to experts and non-experts alike for a device that measures odors. It appeals to one’s imagination and the term is easily uttered. The disadvantage, however, is that it creates great expectations, perhaps too great, because the expression suggests a faithful imitation of the biological sense, which is utterly incorrect: the biological sense of smell is still far superior over today’s artificial odor recognition. This situation will most likely continue for some time. In the absence of a better term, throughout this chapter we will consider the expressions EN and sensor array system as equivalent. As the new concept grew gradually, more and bigger players entered the market. Presently, the EN market is characterized by three trends. We note a geographical

expansion with a concomitant shift of the financial center, a scientific and technological broadening, and a conceptual extension.

7.1.1 Geographical Expansion

The commercial availability of the initial benchtop systems began in the early 1990s. Initially, the EN business used to be an almost exclusive european matter. The first company active in this field was the British OdourMapper Ltd. (1992), shortly afterwards transformed to Aromascan Ltd, Crewe, Cheshire, UK. Located in the Midi-Pyr- e´ne´es, at a stone’s throw from the Mediterranean, Toulouse is the sunny host for Alpha Multi Organoleptic Systems (M.O.S.) (founded in December 1992), commonly regarded as the present market leader. Gradually more players are participating in this emerging field. Back in the UK, in Stansted, Essex, Neotronics Scientific Ltd. (founded in 1994) developed and sold their apparatus NOSE (Neotronics Olfactory Sensing Equipment). The two latter companies have a mutual origin in the tandem University of Warwick/University of Southampton collaboration (Warwick-Southampton EN Group). In fact, quite a few of the current manufacturers find their cradle at a particular university, relying on them for concomitant scientific support. Also in the year 1994, there was an expansion to the north with Nordic Sensor Technologies AB (Sweden) as a newcomer. Slightly before the turn of the century, we note a geographical displacement to the other side of the Atlantic, where new companies such as Cyrano Sciences and Agilent Technologies (formerly known as Hewlett Packard) entered the market. There is little known on the commercial efforts on the Australian and Asian market, although some competence centers exist.

7.1.2 Scientific and Technological Broadening

The factual starting point of EN science was the NATO advanced research workshop ‘Sensors and Sensory Systems for an Electronic Nose’ [9], held in August 1991, Rey- kjavik, Iceland. In the beginning, conducting polymers (CP) were the pet subject of many researchers and EN producers. The systems built and commercialized by Aromascan and Neo- tronics were both based on these materials. Furthermore, Alpha M.O.S. offers their customers a CP-module as an option alongside their metal-oxide semiconductor (MOS) sensor modules. Soon, MOS materials became widely employed, not least because of their proven usefulness in more classical sensors. Other sensitive detecting systems were devised on the basis of other measuring principles, e.g., MOS field effect transistor (MOSFET) and mass detection with surface acoustic wave (SAW) and quartz-crystal microbalances (QMB). On the global scale, the search is on for new types of chemical sensors to implement in an array, shown by the development of calorimetric sensors [10], optical sensors [11–13], electrochemical sensors [14], com- 7.1 Introduction 163 posite polymer-carbon black polymers [15–17], conducting oligomers [18], and phta- locyanine-based sensors. Hybrid ENs are composed of a diversity of different sensor technologies [19–21]. But progress is not only made on the sensory part: optimized hardware and systems design, and overall increase of better data processing algorithms with drift counteraction features contribute to better performing ENs. The emergence of a new promising sensor technology and its strong technological and scientific foundation motivated existing as well as new companies to enter the EN market. As example, we mention OligoSense n.v. (Belgium) as a starter, which produces sensor materials and sensor arrays. Existing sensor producers such as Quartz Technology Ltd. (UK), HKR Sensorsysteme GmbH (Germany), Bloodhound Sensors Ltd. (U.K.), Marconi Applied Technologies (U.K.), and Microsensor Systems Inc. (U.S.), offer sensor arrays to implement in existing nose platforms as a module and/or gradually profile themselves as producers of complete sensor systems but with a focus on sensors. The former approach is preferential since it avoids perfusion of complete sensor-array systems, which then have to compete on an emerging and too crowded market. Also, it allows an optimal use of available hardware, and no precious time and effort is lost on the repeated design of complete systems. Furthermore, the sensor designer can concentrate on the technology by which implementation of improvements is accelerated, the supply is increased and the design of application specific sensors and arrays is facilitated. Finally, some other companies are constructing systems for a dedicated application such as Element, Iceland (e.g., quality control of fish), Environics Industry Oy, Finland (e.g., military-industrial), and WMA Airsense Analysentechnik, Germany (e.g., environmental). It is of interest to see that, at least for the time being, established classical sensor producers such as Figaro (Japan), Capteur (U.K.), FIS (Sweden), and Dra¨ger (Ger- many) do not take the risk, but that their products find their way to prominent EN constructors.

7.1.3 Conceptual Expansion

One can acknowledge three conceptual displacements. First, the definition by Gardner and Bartlett [5] ‘an electronic nose is an instrument which comprises an array of electronic chemical sensors with partial specificity and an appropriate pattern recognition system, capable of recognizing simple or complex odors’ became inaccurate when mass spectrometric detection (SMart Nose and Agilent Technologies) or (flash-) gas chromatography-based separation adjoining SAW-sensor detection (Electronic Sensor Technology) were introduced. Secondly, handheld devices have their own typical target market (e.g., leakage detection) where the low concentrations of typical odorous molecules are not of primordial importance, but where people are interested in detecting rather high (> 100 ppm) concentrations of (predominantly organic) volatile compounds. Last but not least, the array principle is conveyed to the wet phase where potentiometric and amperometric chemical sensors form the building blocks 164 7 Commercial Electronic Nose Instruments

for an electronic tongue (first record of the expression in 1993 [22]). The commercial availability of such a system is currently restricted to Astree from Alpha M.O.S. Since this new technology falls beyond the scope of this chapter, we would like to refer the reader to the literature [23–38] and Chapter 11.

7.2 Commercial Availability

7.2.1 Global Market Players

See Table 7.1 for a summary of the main players and the basis of their instruments.

Tab. 7.1 Electronic Nose Manufacturers, Models and Sensor Cores

Company Sensor Core System

Agilent Technologies MS 4440 Alpha M.O.S. MOS, CP, SAW Fox, Centauri MS and MS-EN Kronos & Prometheus electronic tongue Astree Applied Sensor MOSFET, MOS, QCM 3320, 3310 4 Â MOS, 8 Â QCM VOCseries1 QCM VOCcheck1 Bloodhound Sensors CP BH114 Cyrano Sciences Inc. CP (composite) Cyranose 3201 Daimler Chrysler Aerospace QCM, SAW, MOS SAM system Electronic Sensor Technology SAW zNose Element MOS FreshSense Environics Industry IMCELL MGD-1 Forschungszentrum Karlsruhe MOS, SAW Sagas HKR Sensorsysteme QCM, MS QMB6 Lennartz Electronic QCM, MOS, electrochemical MosesII Marconi Applied Technologies CP, MOS, QCM e-Nose 5000 Microsensor Systems SAW ProSat Osmetech CP OMA and core sensor module Quartz Technology QCM QTS-1 SMart Nose MS Smartnose-300 WMA Airsense Analysentechnik MOS PEN

1 handheld device CP conducting polymer IMCELL ion-mobility MOS metal-oxide semiconductor MOSFET metal-oxide semiconductor field-effect transistor MS mass spectrometry-based QCM quartz-crystal microbalance SAW surface accoustic wave 7.2 Commercial Availability 165

7.2.1.1 Alpha M.O.S. As mentioned earlier, Alpha M.O.S. is one of the pioneers on the EN market. Since its establishment in 1992, the company has seen a steady growth, which resulted in a stock market quotation on 2 April 1998 in Paris on ‘Le Nouveau Marche´’ (ticker: 6280). It has settlements and branch offices in France, the United States, the United Kingdom, and Germany, while Bran&Luebbe tends to its distribution network in other parts of the world. This makes the company a global player. Without any doubt, Alpha M.O.S. also has the largest range of different systems. We mention the modular laboratory system FOX, which encompasses FOX2000, FOX3000, FOX4000 and FOX5000, systems which contain one, two, three and four sensor arrays, respectively, each of which containing six sensors. A standard array board consists of MOS sensors, of which two are available and can be extended with a QMB and/or CP board. The company engages also in the mass-spectrometric approach of olfactometry with their device a-Kronos. In addition they intend to introduce Centauri, a new hyphenated technique that couples an EN to a mass-spectrometric module. In the near future, they intend to introduce the first commercial electronic tongue, under the name Astree Liquid and Taste Analyser. The software used to interpret the data, called a-Soft, originated as National Instru- ments’ Labview and has now reached its seventh release. It allows techniques like principal component analysis (PCA), projection to latent structure (PLS), and artificial neural networks (ANNs), as well as a transferability utility to convert data from different systems (i.e., slightly different sensors and systems). Alpha M.O.S. took the initiative and in 1993 organized the first ‘International Sym- posium on Olfaction and the EN’. This initiative was taken over by the academic world in 1998.

7.2.1.2 AppliedSensor Group On December 4th, 2000, Nordic Sensor Technology and MoTech announced the merger between them. The new alliance is called AppliedSensor Group, with offices in both Sweden and the USA.

MoTech Sensorik, Netzwerke und Consulting GmbH Founded by a couple of researchers from the MOSES II project at the University of Tu¨bingen, this company initially provided services, sensors, and software to the nose producer Lennartz Electronic GmbH. Their own developments include a scale of portable and handheld sensor array systems based on Tagushi and mass-sensitive devices. Scientific backup and cooperation is provided by the Steinbeis Transfer Center for Interface Analysis and Sensors, and the Institute for Physical and Theoretical Chem- istry at the University of Tu¨bingen. There are four members in the VOCmeter series: VOCmeter MOS, VOCmeter QMB, VOCmeter HYBRID (ranging from Q 11 400–Q 18 900) and the VOCmeter VARIO (priced at Q 7900 exclusive of sensors, individual sensors at Q 490) to measure 8 external sensors. Signal recording and processing of the VOCmeter series is per- 166 7 Commercial Electronic Nose Instruments

Fig. 7.1 AppliedSensor’s handheld VOCcheck. Reprinted with kind permission

formed using a RS232-linked PC with an uncomplicated user interface called Argus. The VOCcheck (Fig. 7.1) is a handheld device based on QMB sensors, allowing rapid (< 10 s) identification of volatile compounds, especially in the field of leakage detection and emission control, and is a comparative method against pre-measured reference samples. OEM modules on the basis of the VOCmeter technology can be adapted for a large variety of different requirements.

Nordic Sensor Technologies AB The origin of this leading company was the research made at the University of Lin- ko¨ping (Sweden) in 1994, known by the name Nordic Sensor. A financial injection in March 1996 lead to the formation of Nordic Sensor Technologies AB. There is still a tight symbiosis with the research group ‘Laboratory of Applied Physics’ and the Swed- ish sensor center ‘S-Sence’: the home of sensing MOSFETs [39] since 1975. The successor for the first NST 3210 Emission Analyser is the NST 3220 Lab Emis- sion Analyser. These systems are available for atline (batch) and online (continuous) quality control measurements. Improvements on this blue-and-gray machine included an uncomplicated carousel, allowing 8 specimens in vials of 250 ml to be sampled. At a second stage, the carousel was thoroughly re-examined. This resulted in a 12-position carousel, allowing heating (up to 65 8C) and cooling of the samples: the NST 3320 EN (Fig. 7.2). Although the gas sampling and system design was largely adapted, the core sensor technology remained the same, based on two arrays of 5 MOSFET sensors (at different operating temperatures 140 8C and 170 8C) and one array of 5 MOS sensors. Optional

sensors include CO2 IR devices (1 % or 10 %) and other in-house mass sensitive devices. Its modular principle allows one to include other sensor technology-based arrays. The company is targeting quality control, process control, environmental analysis, and medical diagnosis. One important breakthrough was reported in the field of on- 7.2 Commercial Availability 167

Fig. 7.2 AppliedSensor’s Electronic Nose Model 3310. Reprinted with kind permission

line monitoring of fermentation and other bioprocesses [40]. In addition, AppliedSen- sor also focuses on OEM (technology platform) and component sales. The proprietary data-acquisition and data-processing software Senstool (current release 2.7.4.26) is a straightforward Windows-based graphical user interface (GUI). It contains PCA, PLS and ANN algorithms. The files are saved in the Microsoft Excel format, allowing easy (re-) processing.

Resulting Merger: AppliedSensor The merger of Nordic Sensor Technology and MoTech leads to a powerful company by bringing together a massive amount of knowledge such as sensor technologies, data processing, hardware, and software. This global player offers a variety of handheld and benchtop sensor array systems. AppliedSensor will be a fearsome opponent of Alpha M.O.S., let the battle begin!

7.2.1.3 Lennartz Electronic Lennartz Electronic GmbH has more than 30 years of experience in physical sensors and high-quality data acquisition systems. Their modular EN is called MOSES (MOd- ular SEnsor System). MOSES II (Fig. 7.3) has been developed in close cooperation with Steinbeis-Transferzentrum Grenzfla¨chenanalytik und Sensorik at the Universita¨t Tu¨- bingen (Center for Interface Analytics and Sensors). Lennartz Electronic GmbH uses a basic sensor configuration, consisting of eight commercially available Tagushi sensors and eight quartz microbalance sensors coated with different polymers. These quartz microbalances are manufactured at the Steinbeis-Transferzentrum Grenzfla¨chenanalytik und Sensorik. Under current investigation is a calorimetric module.

7.2.1.4 Marconi Applied Technologies (now ELV Technologies) Marconi Applied Technologies is a general designer and manufacturer of electronic components. It acquired EEV Chemical Sensor Systems, which was formerly known as Neotronics Scientific. The eNOSE 5000 range of instruments was originally de- 168 7 Commercial Electronic Nose Instruments

Fig. 7.3 Modular Sensor Sy- stem II from Lennartz Electronic. Reprinted with kind permission

signed for the laboratory-based profiling of samples through measurement of the characteristic sample headspace. Marconi Applied Technologies no longer markets the eNOSE 5000 instruments for general-purpose laboratory applications. Instead it has developed a real-time monitoring system based on chemical sensor-array technology ProSAT (for atline or online monitoring), predominantly for bioprocessing, fermentation monitoring, food industrial applications, water and wastewater treatment, and the chemical industry. As a consequence of its expertise in sensor development, it also has proprietary libraries of discrete sensors using conducting polymers (CP),

MOS (SnO2; CrTiO2), SAW (@260 MHz), and QMB (@10 MHz) and hence can form custom-made arrays. Investigation is focused towards the development of mo- lecularly specific sensors. In the discrete sensor product range a variety of sensors are offered such as pellistor-type catalytic gas sensors, thermal conductivity and infrared- based sensors (about 35 proprietary sensors in total). An additional benefit of all Mar- coni’s sensor designs is that individual sensors can be substituted or replaced within an array, allowing for array optimization. Tight manufacturing control ensures that sensor reproducibility is high and preserves training model validity when sensors are replaced. Typical sensor arrays contain between 4 and 12 sensors, with 8 in the standard configuration. Some of these are based on standard multivariate techniques such as PCA, multiple discriminant analysis (MDA), canonical analysis (CA), and ANNs. Advanced calibration algorithms are used to compensate for long-term sensor drift and to ensure validity of data sets from module to module.

7.2.1.5 Osmetech plc From as early as 1980, research has been conducted at the University of Manchester Institute of Science and Technology (UMIST) to come up with an instrumental equivalent of the biological nose. The great originality of the project was the use of CP sensors with their broad sensitivity to various vapors coupled to an extensive data processing system. The work led to the first operational prototype in 1990. The establishment of the spin-off company, OdourMapper Ltd, by a group of researchers related to this 7.2 Commercial Availability 169

Fig. 7.4 The heart of Osmetech’s sensor module is this substrate equipped with 48 sensors

project followed in 1992. As early as 1994, the spin-off went to the Alternative Invest- ment Market (now London Stock Exchange, TechMARK, ticker OMH). It collected £11 million in this stock market quotation, while converting into AromaScan plc. The company became market leader in its field, and was awarded the Prince of Wales’ Award for Innovation. Its competence shows in more than 25 patents and publications on data processing [41] as well as on sensor design and development [42, 43]. They patented a method and apparatus for detecting microorganisms, and entered the area of biomedical applications [44]. After some wanderings in various parts of the quasi-infinite number of possible areas of applications, the company focused on biomedical uses, and changed its name to Osmetech plc. (1999) in the process. Their emphasis is now on the detection of volatile metabolites excreted in bacterial infection of the urinary tract, bacterial vaginosis, early diagnosis of bacterial pneumonia, and bacterial pharyngitis. In addition to these biomedical applications the company sells industrial systems applicable to the quality control of basis products used in e.g., health and body care, plastics, and polyurethane foam. The old AS32 systems, with their external humidity controller and 20 to 32 organic- based sensors (see Fig. 7.4), have been replaced by a new and upgraded line of apparatus, Osmetech Microbial Analyser (OMA). The new apparatus houses 50 glass vials, capped with a spectrum through which the headspace can be purged (dynamic headspace sampling). The sensor section of these systems is constructed as an independent module, the so-called Core Sensor Module (CSM). The CSM contains up to 48 sensors situated on a circular substrate. It also contains the essential temperature controller and the electronics for signal processing and the data acquisition interface. Dedicated sensor arrays specifically designed for certain clinical infections are offered, while a universal CSM suffices for the other areas of applications. 170 7 Commercial Electronic Nose Instruments

7.2.2 Handheld Devices

7.2.2.1 AppliedSensor Group For a detailed description see Section 7.2.1.2.

7.2.2.2 Cyrano Sciences, Inc. Cyrano Science was founded in 1997, and raised over $12 million to further develop the patented [16] original composite polymer technology, elaborated by Nathan Lewis et al. at the California Institute of Technology. The company holds 10 US and 1 Eur- opean patents in total, the last one in the emerging and remunerative field of medical and biomedical applications. The Cyranose 320 (Fig. 7.5) is a handheld device comprising a 32-polymer composite (polymers filled with the conductive particles carbon black or another conductive filler) sensor array (Nose Chip). The launch of this $9,000- priced handheld device took place at the technology exhibition Pittcon2000 in New Orleans. The detection limit of the Cyranose320 for different volatile compounds is estimated roughly at 0.1 % of the standard vapor pressure. The dedicated on-board firmware (current release 30.1) is capable of differentiating six different classes for each method stored. The instrument settings, defined methods and raw data can be swapped, stored and further processed on a Windows-based PC using PCnose software (current release 6.5). To share the knowledge optimally, collaboration agreements were signed with Agilent and Osmetech. The Osmetech agreement comprehends a Healthcare Collaboration Agreement, in which Osmetech polymer sensors will be implemented in the Cyranose 320 and used for validation on the detection of the presence of bacteria in urine causing urinary tract infections. Agilent has signed a collaborative research agreement with Cyrano Sciences, sharing among other things the Infometrix Pirouette software.

Fig. 7.5 The Cyrano 320 handheld device from Cyrano Sciences 7.2 Commercial Availability 171

7.2.2.3 Microsensor Systems, Inc. Ever since 1979, Sawtek has been a dedicated SAW device developer for a countless number of applications in communications, cellular wireless data transmission, and other signal-processing applications. In 1998 it merged with Microsensor Systems, a company developing chemical sensing technology using the same SAWs. Using an advanced, polymer-coated SAW array, a broad spectrum of chemical vapors can be accurately identified. SAW sensors have excellent long-term stability and are effective sensors for higher molecular weight, semi-volatile organic compounds not readily detected by other sensor technologies. VaporLab (Fig. 7.6), a handheld, battery-pow- ered chemical vapor identification system costing $10 000, goes where you need it, providing on-the-spot information on the current status of your process, product, or environment so that immediate action can be taken as required. Typical foremost applications include environmental, food and beverage, fragrance and cosmetics, safety exposure and personal monitoring, and medical and dental.

7.2.3 Enthusiastic Sensor Developers

7.2.3.1 Bloodhound Sensors Ltd. CP sensor research work for Bloodhound Sensors began at the University of Leeds, where the company is currently based. The rather compact BH114 is an instrument comprising an array of 14 CP sensors, and the data processing is performed using Microsoft Excel add-ins and specialized add-ins such as Neuralyst from Palisade Corp. The sensor technology is based on CPs and discotic liquid crystals. These devices are also available individually or in an array.

7.2.3.2 HKR Sensorsysteme GmbH HKR Sensorsysteme was founded in 1993 by three researchers from the Technical University of Munich. An array of six QMBs forms the heart of their benchtop EN

Fig. 7.6 Microsensor systems’ handheld device VaporlabTM 172 7 Commercial Electronic Nose Instruments

consisting of an automated Perkin-Elmer HS 40XL or Dani HSS86.50 headspace-sampler and the proprietary QMB6 array. Optionally, a thermal desorption trap (Markes Int.) can solve the problem of too low concentration by trapping the analytes of interest and purging the high volatile compounds, allowing analytes in the lower parts per billion range to be detected. MS – Sensor is a sensor system on a mass spectrometric basis, using a quadrupole mass spectrometer, TurboMass. Qmbsoft for Windows NT controls the automated measurements and acquires and evaluates the data using PCA, GDF, and RBF neural network pattern recognition techniques.

7.2.3.3 OligoSense n.v. Research on sensors for an EN has been conducted since 1993 at Antwerp University. The original focus was on electrically conducting polymers, however it was noted that short fragments of these polymers, oligomers, have better sensory properties than their polymeric analogs [18], and the investigation was then concentrated on this area. This new focus of investigation will hopefully lead to a steady stream of new sensor materials and sensor modules. As a consequence, OligoSense n.v. has been formed to produce and market the oligomeric technology.

7.2.3.4 Quality Sensor Systems Ltd. Q-Sensor developed a chemical sensor array instrument dedicated to applications in the food and food packaging industry. The QMBA8000, based on eight QMB sensors, has been developed on a generic platform, and it is this modular approach to design which allows chemical sensor systems to be developed for a diverse range of application areas by offering the appropriate sampling system and chemical sensor array.

7.2.3.5 Quartz Technology Ltd. Started in March 1996, Quartz Technology’s main objective is to commercialize QMB- based sensor technology. Nowadays, they market their standard balanced eight-sensor array instrument QTS-1, and in addition a range of separate QMB sensors and even blank quartz crystals are available. Focusing on applications, the QTS-1 can be equipped with a custom array, or even dedicated systems with more sensors can be designed. This company also provides custom solutions to specific measurement problems. The compact system accepts at its inlet (no carousel) sample air from jars or vials or introduced from an external sampling system. The sensor signals are processed and compared to an online library for rapid identification. The software is written for a Windows98/NT platform. Although Quartz Technology would never claim that QTS-1 is an EN, it is capable of diagnosing many aroma problems based on differing chemical fingerprints. 7.2 Commercial Availability 173

7.2.3.6 Technobiochip Ever since 1995, research at the Tor Vergata University in Rome has been carried on porphyrins and related compounds for coating mass transducers for chemical sensors. The main feature of such sensors is the dependence of the sensing properties (in terms of selectivity and sensitivity) on the nature of the central metal and on the peripheral substituents. Technobiochip produces this instrument, named LibraNose, and ships it with a number of data processing algorithms based on PCA, CA and ANNs that are used for information extraction.

7.2.4 Non-Electronic Noses

This section deals with systems that don’t meet the strict definition as explained earlier this chapter (Section 1.3). Note that Alpha M.O.S. (see Section 2.1.1) and HKR Sen- sorsysteme (see Section 2.3.2) also offer a mass spectrometry-based system.

7.2.4.1 Laboratory of Dr. Zesiger Using mass spectrometry should overcome the typical chemical sensor problems such as their sensitivity towards sample and environmental moisture. The advantages of this technique are its sensitivity and robustness. Contrary to GC/MS, there is no preceding separation of the volatile constituents allowing measurements every 5 minutes. However, the basic operation of this kind of equipment needs an adapted gas supply (helium) and requires high vacuum pumps inextricably bound up with a high system price. The price of a nose on the other hand is predominantly determined by the research and development contribution and could eventually go down substantially when the market increases. SMart Nose (Fig. 7.7) is a fully automated combination of a Balzers Instrument Inc. quadrupole mass spectrometer with an autosampler for 2 ml or 20 ml vials. The system is entirely software controlled: the Quadstar from Balzer Instruments controls

Fig. 7.7 SMart Nose mass spectrometer with autosampler 174 7 Commercial Electronic Nose Instruments

operation of the mass spec, channel selection and sample measurement. The SMart Nose software processes the raw mass spectrometric data using statistical algorithms such as PCA or discriminant function analysis (DFA), to yield a more user friendly representation of the results.

7.2.4.2 Agilent Technologies, Inc. Agilent Technolgies (formerly known as Hewlett-Packard) is a well-known manufacturer of all kinds of analytical instruments, in fact a scientific instrument giant. By combining a recent in-house mass spectrometer 5973N MSD with an in-house modified headspace autosampler 7694, the Agilent system uses quadrupole technology as a mass sensor to provide qualitative information about sample attributes. The Infome- trix software allows the raw data to be processed for classification purposes using multivariate techniques and pattern recognition. The Agilent 4440A Chemical Sensor (Fig. 7.8) is priced at $80 000, and is currently available through Gerstel GmbH. Agilent and Cyrano signed a pact to jointly develop new versions of their ENs, con- ceivably expanding the mass spectrometer with a classical composite polymer-based sensorarray, and to collaborate on marketing.

7.2.4.3 Illumina, Inc. Optical sensing technology [12] has been reported by Dickinson et al. of Tufts Uni- versity. Illumina licensed this technology, and has recently started to market an EN. Until now, their main focus was on the large-scale analysis of genetic variation and function. Illumina’s technology is also suited for chemical detection applications, because their BeadArray fiber optic bundles can be designed to ‘house’ cross-reactive, nonspecific sensors capable of responding to a wide variety of solvent vapors.

7.2.4.4 Electronic Sensor Technology, Inc. Electronic Sensor Technology produces the zNose, which consists of only a single patented sensor based on SAW technology and a directly heated 1 m length of capillary chromatography column. Visualization software for making radar plots, EST System

Fig. 7.8 Agilent 4440A Chemi- cal Sensor 7.2 Commercial Availability 175

Software for Windows95, is used. A benchtop and a handheld version of the zNose, priced at $25 000, are offered commercially.

7.2.5 Specific Driven Applications

7.2.5.1 Astrium Astrium is a subsidiary of RST Rostock Raumfahrt und Umweltschutz GmbH, which belongs to the DaimlerChrysler Aerospace division. The EN technology Sam is a measurement technique suitable for an objective, quick, and low cost analysis of odor, aromas, and volatile compounds. They offer a range of three sensor systems based on a modular concept using MOS, QMB, and SAW technology.

7.2.5.2 Element Ltd. Element started developing gas detectors in co-operation with the Science Institute at the University of Iceland, in 1992, then under the name RKS Sensor Systems. The relationship with the university is still maintained. The gas detector systems form the main product line of the company together with Medistor, a data acquisition system. In a project in co-operation with the Icelandic Fisheries Laboratories, Element has developed an instrument called FreshSense to detect fish freshness. FreshSense detects components that are produced in fish during storage and gives comparable results to traditional methods to evaluate freshness such as sensory analysis. FreshSense is built on an array of six commercially available electrochemical sensors (Dra¨ger) with PCA and PLS algorithms to classify samples.

7.2.5.3 Environics Industry Oy The Environics company targets chemical detection applications for the military. De- tection is based on a proprietary ion mobility cell (IMCELL), where sample molecules are first ionized using some radioactive source (e.g., Am241) and then flow towards an array of six detector electrodes. The MGD-1 Industrial Multi-Gas Detector can be used as a portable or a fixed version. VisualNose for Windows is software designed to present the data that has been collected with MGD-1 in 2D format.

7.2.5.4 WMA Airsense Analysentechnik GmbH WMA Airsense’s portable EN PEN2 consists of an array of 10 MOS sensors with adapted software. It is designed for laboratory measurements as well as for online process monitoring. Focusing on air pollution and air quality control measurements, an optional enrichment unit (EDU2 – absorbent trapping on Tenax) can be valuable for this Q 14 900 unit. For operation in hostile industrial environments, an industrial process control EN, i-PEN, is offered. Different configurations of the i-PEN are available: a basic module i-PEN-MOD (based on 10 MOS devices) has an 176 7 Commercial Electronic Nose Instruments

on-board microcontroller; a process control nose i-PEN-PCN consists of a sensor array, gas pumps, and a patented sampling system. The i-PEN-ET has an additional enrichment and desorption unit with A3-technology (automatic ranging, automatic calibration, and automatic enrichment). The software provided for Windows NT4 incorporate PCA and LDA algorithms for visualization of the data and DFA and ANNs for classification and online evaluation. Recommended prices are Q 4900 for the i-PEN-MOD module to Q 14 900 for the PEN-2.

7.3 Some Market Considerations

Prudence is called for when assessing the size of the EN market. The estimates range from a modest Q 10 million to a dazzling Q 4.5 billion globally a year predicted by the Economist [45] based on a world market of 100 000 units sold annually in the first years of the 21st century. The best we can do is to give below the most recent results of a short list of market evaluation studies:

1. David Walt of Illumina and Tufts University estimated that 200 units were sold in the last five years (1994–1998) [46]. 2. In April 1998, the Wall Street Journal published an estimate of the market at that time to the amount of Q 10–15 million [47]. A document by Greenberg [48] estimated the market value at Q 15 million, and these figures seem to be acceptable. 3. According to Bartlett and Gardner [2], the market is estimated at about Q 145 mil- for the year 2000. This estimate is corroborated by a Technical Insights report [50] that states the sales of 2500 units.

Fig. 7.9 Incomplete Overview of trade volume in 1000’s Q for five leading companies (Alpha M.O.S., Bloodhound, Lennartz, Osmetech, and Neotronics) for the period 1994–1998 [51] 7.3 Some Market Considerations 177

4. The German Intotech Consulting Group foresees a market potential of Q 1.2 billion by the year 2004. The British/American journal The Economist even gives a potential of Q 4.5 billion annually [45].

By examining the turnover figures coming from Graydon reports of 5 leading companies at that time (Alpha M.O.S., Bloodhound, Lennartz, Osmetech, and Neotronics) for the period 1994–1998 [51], our conclusions are somewhat modest. The figures for 1997 can be extrapolated for all players active in those days. If one considers a market penetration of NST (with an estimated penetration of 15 % at that time), a rough estimate of the market at Q 10 million seems to be acceptable. Unfortunately, we don’t have the most recent figures, but it would seem there was no explosion of the market. Taking into account an annual growth rate of 7.5 % (which is commonly used for the market of ‘classical’ analytical devices for the period 1995– 2000), the market would be worth around Q 13.4 million for 2000. If we take the annual growth rate of the US market for ‘new generation’ analytical instruments and components [52] of 19 %, the market would be worth some Q 20.0 million for 2000. The newcomers with handheld devices have added an additional new market, which falls beyond the scope of this deduction. The producers express optimistic views with regard to the trade volume. The total market for electronic noses was $ 140 million for 1998 and is projected to be $ 200 million by the year 2003 [49]. There is of course much space for other interpretations. These figures demonstrate the large uncertainties in evaluating the young EN market. It is clearly an emerging high-tech market with enormous potential as well as high risks. Therefore, it is of interest to look at the amount of venture capital that is invested in EN technology up to now (May 2001). A conservative estimate says that well over Q 350 million has been invested in this technology throughout the years, of which Osmetech alone accounts for some Q 70 million. This reveals that investors have an optimistic outlook on the growth potential of this emerging technology, however a loss of Q 6.4 million was reported for Osmetech for the year 2001.

References

1 G. H. Dodd, K. C. Persaud. Nature, 1982, 6 E. Vanneste. Review on the commercial 299, 352–355. availability and research efforts on electronic 2 J. W. Gardner, P. Bartlett. –8.3 Commercial noses. http://nose.uia.ac.be/review. Instruments, in Electronic Noses: Principles 7 H. T. Nagle, R. Gutierrez-Osuna, and Applications. 1999, Oxford University S. S. Schiffman. IEEE Spectrum, 1998, 35(9), Press: Oxford. p. 194. 22–34. 3 D. J. Strike, M. G. H. Meijerink, 8 E. Zubritsky. Analytical Chemistry, 2000, M. Koudelka-Hep. Fresenius Journal of 72(11), 421A–426A. Analytical Chemistry, 1999, 364, 499–505. 9 J. W. Gardner, P. N. Bartlett, eds. Sensors and 4 M. A. Craven, J. W. Gardner, P. N. Bartlett. Sensory Systems for an Electronic Nose. NATO Trends in Analytical Chemistry, 1996, 15(9), ASI Series: Applied Science. Vol. 212. 1992, 486–493. Kluwer Academic Publishers: Dordrecht, 5 J. W. Gardner, P. N. Bartlett. Sensors and the Netherlands. p. 327. Actuators B: Chemical, 1994, 18(1–3), 10 J. Lerchner, D. Caspary, G. Wolf. Sensors 211–220. and Actuators B: Chemical, 2000, 70(1–3), 57–66. 178 7 Commercial Electronic Nose Instruments

11 J. White, J. S. Kauer, T. A. Dickinson, D. R. 29 Y. Vlasov, A. Legin. Fresenius Journal Walt. Analytical Chemistry, 1996, 68(13), of Analytical Chemistry, 1998, 361(3), 2191–2202. 255–260. 12 T.A. Dickinson, J. White, J. S. Kauer, 30 P. Wide, F. Winquist, P. Bergsten, D. R. Walt. Nature, 1996, 382(6593), E. M. Petriu. IEEE Transactions on Instru- 697–700. mentation and Measurement, 1998, 47(5), 13 D. R. Walt, T. Dickinson, J. White, J. Kauer, 1072–1077. et al.. Biosensors and Bioelectronics, 1998, 31 A. V. Legin, A. M. Rudnitskaya, Y. G. Vlasov, 13(6), 697–698. C. Di Natale, et al.. Sensors and Actuators B: 14 J. Stetter, P. Jurs, S. Rose. Analytical Chemical, 1999, 58(1–3), 464–468. Chemistry, 1986, 58(4), 860–866. 32 F. Winquist, I. Lundstro¨m, P. Wide. Sensors 15 B. J. Doleman, M. C. Lonergan, E. J. Severin, and Actuators B: Chemical, 1999, 58(1–3), T. P. Vaid, et al.. Analytical Chemistry, 1998, 512–517. 70(19), 4177–4190. 33 C. Di Natale, R. Paolesse, A. Macagnano, 16 N. S. Lewis, M. S. Freund. US5571401: A. Mantini, et al.. Sensors and Actuators B: Sensor arrays for detecting analytes in fluids, Chemical, 2000, 64(1–3), 15–21. 1996. 34 A. Legin, A. Rudnitskaya, Y. Vlasov, 17 M. Lonergan, E. Severin, B. Doleman, C. Di Natale, et al.. Sensors and Actuators B: S. Beaber, et al.. Chemistry of Materials, 1996, Chemical, 2000, 65(1–3), 232–234. 8(9), 2298–2312. 35 L. Rong, W. Ping, W. L. Hu. Sensors and 18 M. De Wit, E. Vanneste, F. Blockhuys, Actuators B: Chemical, 2000, 66(1–3), L. J. Nagels, et al.. Chemically sensitive sensor 246–250. comprising arylene alkenylene oligomers, 36 F. Winquist, S. Holmin, C. Krantz Rulcker, EP0878711; JP11072474; US6042788, 1997. P. Wide, et al.. Analytica Chimica Acta, 2000, 19 H. Ulmer, J. Mitrovics, U. Weimar, 406(2), 147–157. W. Go¨pel. Sensors and Actuators B: Chemical, 37 K. Toko. Measurement Science & Technology, 2000, 65(1–3), 79–81. 1998, 9(12), 1919–1936. 20 H. Ulmer, J. Mitrovics, G. Noetzel, 38 C. Krantz-Rulcker, M. Stenberg, U. Weimar, et al.. Sensors and Actuators B: F. Winquist, I. Lundstro¨m. Analytica Chemical, 1997, 43(1–3), 24–33. Chimica Acta, 2001, 426(2), 217–226. 21 M. Holmberg, F. Winquist, I. Lundstro¨m, 39 I. Lundstro¨m, S. Shivamaran, C. Svensson, J. Gardner, et al.. Sensors and Actuators B: L. Lundqvist. Applied Physics Letters, 1975, Chemical, 1995, 26–27, 246–248. 26(2), 55. 22 P. Wide, F. Winquist. WO9913325 Electronic 40 T. Bachinger, P. Martensson, Tongue, 1993. C.F. Mandenius. Journal of Biotechnology, 23 Y. G. Vlasov, A. V. Legin, A. M. Rudnitskaya, 1998, 60(1–2), 55–66. C. DiNatale, et al.. Russian Journal of Applied 41 K. C. Persaud, P. J. Wells. Pattern Recognition Chemistry, 1996, 69(6), 848–853. With Combination Of Mappings, EP0909426; 24 C. Di Natale, A. Macagnano, F. Davide, WO9801818, 1998. A. D’Amico et al.. Sensors and Actuators B: 42 K. C. Persaud, P. Pelosi. Semiconducting Chemical, 1997, 44(1–3), 423–428. Organic Polymers, EP0766819; WO9600384, 25 A. Legin, A. Rudnitskaya, Y. Vlasov, 1996. C. DiNatale, et al.. Sensors and Actuators B: 43 K. C. Persaud, P. Pelosi. Semiconducting Chemical, 1997, 44(1–3), 291–296. organic polymers for gas sensors, EP0766818, 26 Y. G. Vlasov, A. V. Legin, A. M. Rudnitskaya, US5882497, 1999. A. D’Amico et al.. Journal of Analytical 44 P. A. Payne, K. C. Persaud. Method and Chemistry, 1997, 52(11), 1087–1092. apparatus for detecting microorganisms, 27 Y. Vlasov, A. Legin, A. Rudnitskaya. Sensors EP0765399, US5807701, 1998. and Actuators B Chemical, 1997, 44, 45 The Economist, Artificial Noses. Now to sniff 532–537. at., Sept 5 1998. 28 F. Winquist, P. Wide, I. Lundstro¨m. Analy- 46 Electronic Noses Grow Up: Versatile Sensors on tica Chimica Acta, 1997, 357(1–2), 21–31. their Way to Market, Technical Insights, John Wiley, 1998. 7.3 Some Market Considerations 179

47 Wall Street Journal, Electronic-nose firm seeks 51 OligoSense. Vooronderzoek met betrekking sweet smell of success, April 20 1998. tot ontwerp van een protoype oligomeer 48 I.Greenberg.TechnologyReview,August1998. sensorenmodule voor implementatie 49 M. Bourne, Intelligent Sensing: Micro Noses, in bestaande toestellen, te omschrijven als Eyes and Tongue, G236, Business Communi- elektronische neuzen, 1999, 12., Antwerp (in cations Company, 1999. Flemish). 50 Electronic Noses: Detection Revolution for Food, 52 C. Wrotnowski. Business Communications Chemical and Healthcare Industries, Market Company, G171, 1998. (The New Generation for electronic noses. 1998, New York, NY, USA: of Analytical Instruments). Technical Insights/Frost & Sullivan. 181

8 Optical Electronic Noses

Todd A. Dickinson, David R. Walt

8.1 Introduction

A tremendous amount of technical infrastructure and scientific development has taken place in the area of optics, optical communications, and optical hardware over the last several decades. These developments have led to new light sources, such as solid- state lasers, laser diodes, and light-emitting diodes (LEDs). Improved materials for conducting light, such as optical fibers and optical fiber arrays, have been developed. Revolutions in detector technology have also taken place; high sensitivity detectors, such as avalanche photodiodes, have been developed with the ability to detect single photons. Array detectors, such as charge coupled device (CCD) cameras, inten- sified CCD cameras (ICCD), and CMOS (complementary metal oxide semi-conductor) detectors are in widespread use for such applications as digital photography and as- tronomy. Color versions of these array detectors are also being introduced commercially. In addition to these components, significant advances in materials science have led to new types of filters, dichroics, light-directing components such as micromirror arrays, and infinity optics. Most of these devices and components have been developed to advance the telecommunications, entertainment, and computer industries for such applications as fiber-optic communications, digital music, projection devices, and optical information storage. With the advent of these new capabilities, a parallel development has been taking place in the field of optical sensing.

8.1.1 Optical Sensors

Optical sensors are devices that measure the modulation of a light property. Examples include changes in absorbance, fluorescence, polarization, refractive index, interference, scattering, and reflectance. Optical sensors are comprised of four basic components: 1) a light source to interrogate the sensor; 2) suitable optics for directing light to and from the sensor; 3) a detector for detecting the light signal coming from the sensor;

and 4) the sensor itself. In the simplest type of sensor, referred to as an intrinsic sensor, the chemical species being measured carries its own signal. For example, some organic molecules absorb light at specific wavelengths, or fluoresce and thereby emit light at particular wavelengths. These molecules can be detected directly by measuring changes in absorbance or fluorescence at their absorption or emission wavelengths, respectively. In these systems, the ‘sensors’ are the molecules themselves. Thus, only the three instrument components are required as the sensor transduction mechanism is intrinsic to the molecule or molecules being detected. In the more common type of optical sensor, an indicating species is employed. These types of sensors are referred to as extrinsic sensors. Indicators can be dyes, polymers, or other materials that interact with the chemical species of interest, the analyte, to produce signal modulation. For example, an optical sensing material can be prepared by attaching a chemically sensitive dye to a substrate. When an analyte interacts with the sensing material, an absorbance or fluorescence change occurs, which is monitored by the optical instrumentation. A variety of substrates can be employed for optical sensors. Polymeric films can be used as supports to attach indicators. Glass slides can be used both as vehicles for attaching materials to their surface as well as for coupling light to the detection system. Optical fibers, also called fiber optics, can be used to carry light both to and from a sensing material attached to its surface, either at its tip or surrounding the fiber along its annulus.

8.1.2 Advantages and Disadvantages of Optical Transduction

Optical sensors have a number of advantages over other sensor transduction mechanisms. As described above, most of the supporting optical instrumentation has been developed for other applications and can be brought to bear on the optical sensing field. The ready availability of inexpensive instrumentation, and the promise of improved performance with new developments in light sources, optics, and detectors, will continue to enable major advances in optical sensing technologies. The continued movement toward fully integrated optical communication and computation bodes well for the field. In addition to the ready availability of instrumentation, there is a large knowledge base, as well as commercial accessibility to a multitude of indicators that are suitable for optical sensing. Optical signals are not susceptible to electromagnetic interferences. Light is fast. Light attenuation is extremely low through modern fiber optics, which enables remote sensing over long distances with no need for repeaters or amplifiers. Optical measurements, in particular fluorescence, are extremely sensitive and can be used to detect single molecules. Optical sensing can be readily multiplexed because different optical signals can be carried and detected simultaneously. There are also several disadvantages of optical sensing compared to other sensing methods. In general, optical instrumentation tends to be more expensive, materials intensive, and more complex than sensors based on mass or electrical transduction. These latter two methods employ instrumentation that can be largely designed as integrated circuits, making them simpler and less expensive. In addition, optical methods are sometimes 8.2 Optical Vapor Sensing 183 susceptible to interference by stray light. Finally, optical approaches that utilize fluorescent indicators suffer from eventual photodegradation of the dye molecules. A variety of electronic noses have been developed using a diversity of optical transduction mechanisms [1]. In most cases, these systems employ cross-reactive sensors (discussed below) combined with smart signal processing (described in other chapters). Optical sensor arrays have a much shorter history than electronic noses. Con- sequently, there is the hope that these systems will develop rapidly over the next few years.

8.2 Optical Vapor Sensing

Given the wonderfully diverse nature of optical signals, the past 25 years have borne witness to the development of a wide range of light-based chemical vapor sensors. Although the ‘artificial nose’ approach to designing sensing systems was first conceived in the early 1980s [2], only in the last few years has this concept been extended to the optical arena. An increasing number of research groups are now beginning to explore the utility of employing optical sensors in cross-reactive arrays for improving sensing capacity and performance. This section provides a general overview of some of the key approaches to building optical vapor sensors that have been developed over the past two decades, and the transition of some of these approaches into ‘optical electronic noses’.

8.2.1 Waveguides

Central to many optical chemical sensors is the use of waveguides in one of several different formats. Fiber optics, capillary tubes, and planar waveguides all exploit the phenomenon of total internal reflection. Optical fibers, for example, are strands of glass or plastic in which a central ‘core’ is surrounded by a ‘clad’ with a slightly lower refractive index. Light introduced into the fiber core is reflected at the clad/core interface and is thereby conducted via total internal reflection to the distal tip of the fiber. Hollow capillary tubes or planar substrates comprised of two or more materials with differing refractive indices can also be made to guide light extremely efficiently from one end to the other. A wide range of creative ways to exploit the properties of waveguides for chemical sensing have been explored.

8.2.2 Luminescent Methods

Fluorescence methods continue to be among the most popular optical sensing and general spectroscopy approaches for a wide range of applications, usually because of high quantum yields, well-separated excitation and emission spectra, and intrinsic 184 8 Optical Electronic Noses

sensitivity. For a detailed review of fluorescence spectroscopy the reader is referred to Lakowicz [3]. Briefly, fluorophores are molecules that absorb light at one wavelength and emit light at a longer wavelength. This difference in wavelength, and thus energy, is referred to as the Stoke’s shift and represents vibrational relaxation and other energy losses experienced by the molecule following light absorption. How well a fluorophore converts absorbed photons to emitted photons is called its quantum yield or quantum efficiency. Walt [4] and co-workers first combined fiber-optic waveguides with fluorescent dyes for the measurement of organic vapors in 1991 using the polarity-sensitive, solvatochromic dye, Nile Red. Following this initial work, the approach was extended to higher-level arrays of solvatochromic sensors and, finally, to its current configuration as high-density microsphere arrays. This work and its evolution are described in more detail in the final section of this chapter. A number of other groups have also begun to explore fluorescence-based methods for vapor sensing. Fluorescent dyes can exhibit spectral changes based on several mechanisms. One such mechanism is the twisted intramolecular charge transfer (TICT) excited state. Molecules such as the one designed and synthesized by Orellana et al. [5], shown in Fig. 8.1, can assume a number of different, highly polar configurations in their excited state. These excited states will be stabilized when solvated in polar environments such as alcohol vapors and lead to red-shifts in their emission spectra. The degree of these shifts will depend on the particular solvation environment and thus can be used to detect specific vapors. By adsorbing these dyes to silica gel and immobilizing the resulting gel at the tip of an optical fiber, Orellana has been able to demonstrate the reversible measurement of various alcohols. Reichardt’s dye, a betaine fluorophore, is another example of a solvatochromic dye

that exhibits high sensitivity to polarity changes, and has been used to create the ET(30) polarity measurement scale for solvents. An increasing number of groups have begun to incorporate betaine dyes onto the ends of optical fibers in various ways to prepare chemical sensors. One group modified the dye molecule and covalently attached it to a Merrifield peptide resin via a five-step synthesis. Following immobilization to a fiber, the resulting sensor was successfully used to measure polar octane improvers in ga- solines [6]. In a similar study, Rose-Pehrrson et al. [7] entrapped Reichardt’s dye within a series of different polymer films and studied the responses resulting from the varying absorption of analytes. A number of groups have begun to explore the potential for exploiting host-guest supramolecular chemistry for sensing. For example, host compounds that form crystalline inclusions, or clathrates, by temporarily trapping guest molecules within their lattice structures have been utilized for detecting solvent vapors [8]. By incorporating a

Fig. 8.1 A polyaromatic-substituted 1,3-oxazole (or 1,3-thiazole) fluorescent indicator that displays polarity-sensitive TICT excited states [5] 8.2 Optical Vapor Sensing 185 fluorescent anthracene moiety as well as a few key functional groups to impart selectivity for vapors, the authors created a class of compounds they call ‘fluoroclathrands’. When vapors are introduced into a hydrogel layer containing these compounds, the host molecules surround the guest vapor molecules and form inclusion complexes with specific crystal structures and characteristic fluorescence behavior. Depending on the guest molecule, the complexes exhibit both wavelength shifts and quantum efficiency (intensity) changes in their emission spectra. The authors speculate that the bathochromic shifts are due to energy losses associated with increased packing density in the inclusion compound, while the intensity changes are most likely a result of self-quenching that varies as a function of the distance between the fluorophores in the crystal. Unlike fluorophores, which require an excitation source to generate the emission signals, chemiluminescence-based sensors employ chemically reactive species capable of directly emitting photons following oxidation. This approach offers the advantage of simplified instrumentation, by circumventing the need for excitation light sources, as well as high sensitivity since signals arise from initially dark backgrounds. While chemiluminescence has frequently been employed for oxygen and metal-ion sensors, the method has recently been extended to detecting organic vapors such as chlorinated hydrocarbons, hydrazine, and ammonia [9]. The commonly-used reagent luminol 3þ was used to detect oxidants while a Ru(bpy)3 complex was used for reductants. Lu- minol sensing capacity was expanded to halogenated hydrocarbons by the addition of an inline heated platinum filament used as a pre-oxidative step.

8.2.3 Colorimetric Methods

Sensors that measure changes in absorbance (i.e., color), or local refractive index changes resulting from indicator color changes, have also been developed for vapor sensing. Some of the earliest work in this area was done by Wohltjen and colleagues [10], who developed a reversible capillary tube-based sensor for ammonia, hydrazine, and pyridine by coating a glass capillary with an oxazine perchlorate dye film. Color changes experienced by the dye upon exposure to these vapors from 60 to 1000 ppm caused proportional changes in transmission through the tube and were detected by a simple phototransistor. Similarly, Stetter, Maclay and Ballantine [11] used a coating of bromothymol blue suspended in a Nafion polymer layer to detect and quantify H2S and HCl acid vapors down to 10 ppb levels. Even commercially available thermal printer papers have been shown to exhibit reversible interactions with solvent vapors and may be useful in solvent vapor sensing. Wolfbeis and colleagues [12] demonstrated that thermal papers could be immersed in an ether atmosphere to produce a dark blue or black color. The treated paper was found to decolorize to varying extents upon exposure to different polar solvent vapors. By incorporating these papers into various optical devices and monitoring light absorption at 605 nm, sensors were prepared that were capable of mid- ppm to high-ppm detection levels for typical laboratory solvents such as alcohols and 186 8 Optical Electronic Noses

acetates. Response times of these sensors ranged from 30 seconds to 3 minutes, with recovery times of up to 7 minutes for certain analytes. Polymers are frequently employed in a large number of optical sensor constructs for their differential vapor sorption or binding properties as well as their emissive properties. For example, the color changes exhibited by amine-containing poly(vinylchloride) membranes when interacting with polynitroaromatics have been used to detect 2,4- dinitrotoluene (DNT), a compound commonly present in landmines [13, 14]. Absorp- tion into the polymer generates a complex with an absorbance at 430 nm that can be monitored over time to characterize DNT levels in an area of interest. Sensor materials play a central role in all of these various optical approaches, and their study and development has become a major field of exploration in its own right [15]. All of the above vapor-sensing techniques rely on changes in color of an organic sensing material. Inorganic compounds that exhibit environmental sensitivity in both their absorptive and emissive properties are another exciting class of sensing materials. At the University of Minnesota, Mann et al. [16] have shown substantial shifts in maximum absorption and emission wavelengths of platinum and palladium isocyanide complexes resulting from exposure to volatile organic compounds (VOCs). The

Pt-Pt compound [Pt(p-C10H21PhNC)4][Pt(CN)4], for example, was found to exhibit absorption and emission maxima shifts as large as 91 nm and 74 nm, respectively, when

exposed to vapor environments ranging from air to CHCl3. The researchers believe that the incorporation of VOCs into the lattice (which appears to be fully reversible) causes a perturbation in the stacking of the anion and cation complexes that leads to the observed color changes. In the case of polar VOCs, dipole-dipole and/or H-bond- 2 ing interactions with the Pt(CN)4 anion are thought to be involved; for nonpolar compounds, however, the ‘vapochromism’ is explained by lypophilic interactions with the isocyanide complexes. Photostability and an insensitivity to water vapor make these materials particularly attractive for incorporation into an opto-electronic nose sensing device. Metalloporphyrins (Fig. 8.2) represent another class of inorganic materials that are particularly good indicators for sensing as they are stable, well characterized, and easily modified with a wide range of substituents.

Fig. 8.2 General structure of a metalloporphyrin. Modifications can occur at each R and R’ position, and a wide range of metals can be incorporated at the core of the complex 8.2 Optical Vapor Sensing 187

These compounds can both form coordination complexes with analytes as well as adsorb them via van der Waal’s and H-bonding interactions, giving rise to broad selectivity particularly suitable for electronic-nose applications. As a result of their aromatic p-systems, porphyrins exhibit unique absorption and luminescence properties depending on the metal centers and peripheral substituents involved. D’Amico and co- workers [17] were able to distinguish between six different liquors by monitoring absorbance changes with a simple LED and photodetector system. The researchers reasoned that the optical changes were caused by competitive interaction of the VOCs with aggregated porphyrin complexes that lead to broadening and shifting of spectral bands. Rakow and Suslick [18] used metalloporphyrins to construct a colorimetric array detector for vapor-phase ligands. An array was assembled by spotting a series of differentially metalated porphyrins onto silica thin-layer chromatography plates. Imaging the array with a common office scanner before and after vapor exposure revealed a unique pattern of response for each of the various analytes (Fig. 8.3). The degree of spectral shift is thought to be a function of the degree of polarizability of the ligand. Thus, by incorporating a range of metal centers of varying ligand-binding affinity, an array can be made to discriminate between several different analytes. The authors report good reversibility as well as linearity of the sensors. A cobalt-based sensor, for example, responded linearly to binary mixtures of trimethylphosphite and 2-methylpyridine, and could therefore be used to predict the composition of these solvent mixtures. Typically, 15-minute exposures were used with the arrays to ensure maximum array response, although the authors showed that these times could be reduced to 30 seconds for at least one of the sensors. The work employed hydrophobic substrates for the array such as reverse phase silica or Teflon films, which had the advantage of limiting interference from water vapor (one of the most formidable challenges that plague electronic noses). Colorimetric techniques, such as these porphyrin arrays, generally employ simple instrumentation. Sensor reproducibility with sensitivity below the ppm level are presumably among the areas targeted for further work with this approach.

8.2.4 Surface Plasmon Resonance (SPR)

In other work, coordination polymers were used as sensing layers in a SPR setup to detect benzene, ethanol, toluene, acetonitrile, and water [19]. Langmuir-Blodgett films were created using poly(CuMBSH) (MBSH þ 5,5’-methylenebis (N-hexadecylsalicyli- deneamine), which were found to be excellent sensing materials due to their rapid and reversible interaction with vapor-phase analyte molecules. The SPR technique exploits the delocalized conducting electron clouds found at the surface of metal films such as silver and gold. The electron clouds maintain a collective wave vector parallel to the interface. Light of a particular wavelength and polarization incident at the interface at a precise, ‘resonant’ angle will couple to these electromagnetic modes, resulting in a sharp decrease in the measured reflected intensity of the excitation beam. The mo- 188 pia lcrncNoses Electronic Optical 8

Fig. 8.3 Scanned images of an array of different metal-containing indicated, for 15 minutes. Reprinted by permission from Nature [17], porphyrins before and after exposure to various chemical vapors, as copyright 2000, Macmillan Magazines Ltd 8.2 Optical Vapor Sensing 189 mentum matching condition, and thus the resonant angle, is dependent upon the refractive index of the dielectric medium. Therefore any changes in refractive index at the surface, such as that caused by the sorption of vapor molecules into a polymer network at the surface, can be closely measured in real time by monitoring the illu- mination angle needed to give a minimum in the measured reflected light. Alterna- tively, since the resonant angle is also a function of the wavelength of the incident light, a white light source can be used in place of a laser to monitor the wavelength at which the surface plasmon resonance occurs [20, 21]. Although the sensitivity was relatively low in this study, responses to high ppm levels of benzene were demonstrated. The SPR signals are thought to be directly related to refractive index changes at the surface due to swelling of the polymer and/or increased density upon absorption of the analyte vapor. In related work, Abdelghani et al. [22] have applied the SPR technique to optical fibers by coating a 50 nm thick layer of silver onto the core of a silica fiber. To protect against oxidation, alkanethiol layers were assembled onto the silver layers. A fluori- nated siloxane was selected to serve as the final cladding layer due to its appropriate refractive index, surface tension, and gas permeability properties. Although the resulting sensor responses appear to have improved reproducibility and signal-to-noise ratios, the detection limits reported were in the high ppm level for both the aromatic and chlorinated compounds tested, and the cumulative response and recovery times were of the order of several minutes.

8.2.5 Interference and Reflection-Based Methods

Another area of recent activity for sensor development has been the use of interference spectroscopy. Having demonstrated that analyte-swelled polymer films experience much larger changes in optical thickness than refractive index [23], Gauglitz and others have pursued reflectometric interference spectroscopy (RIfS) methods for optical vapor sensing. In this approach, light incident at the interface between two planar optical layers can be reflected from both the top and bottom of a polymer sensing film, setting up an interference pattern that is very sensitive to changes in the optical thickness of the polymer layer. Gauglitz suggested that the method offers two primary advantages over non-optical techniques: 1) the ability to use strictly inert materials (glass and siloxane polymer films) in contact with the vapor samples; and 2) a built-in control for checking the condition of the sensing layer. One of the challenges associated with measuring changes in the interference spectrum has been the requirement for relatively bulky and expensive light delivery and detection equipment. Im- provements to this approach have been pursued through simpler and less expensive optical components [24]. Recent work using four inexpensive LEDs and a single photodiode demonstrated that despite the lower-resolution, four-point spectrum, the simplified RIfS system yields comparable sensitivity and linearity to its more costly pre- cursor [25]. The RIfS technique has also been extended to enantiomer discrimination. By de- positing polymer solutions containing chiral peptide residues from the ‘Chirasil- 190 8 Optical Electronic Noses

Val’ chromatographic stationary phase, Go¨pel and colleagues [26] studied the responses of their sensors to several mixtures of (R) and (S)-methyl lactate in varying proportions. A direct correlation was found: as the concentration of the (S) enantiomer rose in the analyte mixture, the amplitude of the (S)-Octyl-Chirasil-Val sensor rose while the (R)-sensor fell. Interference measurements have also been applied to porous silicon chips (PSi). Sailor and coworkers [27] have developed simple chemical etching methods for generating porous silicon films that display both interferometric and photoluminescence properties. In the case of photoluminescence, the group proposed that quenching can be induced via energy transfer by the adsorption of analyte molecules in the pores of the silicon. Thus, by monitoring emission at a specific wavelength (670 nm in this case), one can observe sharp decreases in intensity as the interaction with analyte vapors takes place. Likewise, adsorption events give rise to refractive index changes that lead to shifts in Fabry-Perot interference fringes, measured as changes in reflectivity. Both of these optical attributes were recently used to measure a range of perfumes and solvent vapors. When compared side-by-side to a commercial electronic nose containing metal-oxide sensors, the PSi chips displayed comparable discrimination ability for a few standard solvents, ethyl esters, and perfumes. At the saturated vapor conditions used, the silicon sensors showed significantly faster recovery times than their metal-oxide counterparts (30 s versus 15 min). The ability to create a diverse array with high sensitivity and broad selectivity with this approach, however, remains to be proven. Another absorbance type of vapor sensor is based on simple transmission attenuation through a fiber. Microbending caused by the vapor-induced swelling of siloxane layers adjacent to the fiber results in transmission attenuation [28]. Yet another creative reflection-based approach to chemical sensing has been the use of resonating microcantilevers such as those used in atomic force microscopy (AFM) for atomic-level imaging. Based on the mass-sensing concepts of resonating piezoelectric crystals (e.g., quartz crystal microbalances), the approach uses 180 lm long cantilevers micromachined into silicon that are sensitive to changes in mass occurring at their surfaces. Several groups have explored coating polymer films onto these cantilevers and measuring small changes in mass loading. The technique uses optical detection by measuring the deflection of an incident laser beam as analyte vapors are adsorbed to the surface. In one study, Thundat et al. [29] showed that such sensors could be modified to possess desired selectivities, for example by employing hygro- scopic coatings to improve sensitivity to water vapor. A group in Switzerland recently proposed that arrays of differentially coated cantilevers could be used as a new form of chemical nose [30]. Working in their own microfabrication facility, the group constructed an eight-cantilever sensor array from silicon. The individual cantilever coatings included platinum thin films, alkythiol self-assembled monolayers (SAMs), zeolites, and poly(methylmethacrylate). The authors studied detection of water vapor, alcohols, and several natural flavors. Although the array was read out sequentially due to the use of a single laser and photo-sensitive device, one can envision ways of multiplexing through beam-splitters and larger, higher-resolution two-dimensional detector arrays. Detection limits were 8.3 The Tufts Artificial Nose 191 not calculated in this study, making it difficult to compare the sensitivity of the approach to other methods. In addition, despite the small size of the devices, reported cycle times were of the order of several minutes. Other challenges with the cantilever approach include interference from pressure changes during sampling, loss of signal due to severe bending of the cantilever, laser heating of the cantilever, and limited dynamic range [29]. Nevertheless, as they continue to be developed and improved, cantilever arrays may prove to be a promising opto-electronic nose format capable of simple integration into silicon-based microelectronic devices.

8.2.6 Scanning Light-Pulse Technique

Lundstro¨m and coworkers have taken an innovative optical approach by employing a method called the scanning light-pulse technique [31–34]. In this approach, light im- pinges on the surface of a metal-oxide semiconductor field effect transistor (MOSFET) coated with a thin metal film and penetrates the metal to induce a photocapacitive current. To maintain a constant current, the applied gate voltage (V) must be varied to sustain a constant surface potential. Changes in the gate voltage are monitored and result in a map of the change in voltage (DV) as a function of position on the sensing surface. In one demonstration, a MOSFET array was prepared with three continuous strips of Pt, Pd, and Ir. The sensor surface was divided into a grid, and a temperature gradient (110–180 8C) was established down the length of the sensor surface. This temperature gradient provided a different sensitivity and selectivity at each point of the sensor grid. The sensor grid was exposed to hydrogen, ammonia, and ethanol, and DV was determined. In this manner, image maps of the gases were created. These sensor grids can be applied to identifying gas mixtures, rapid and simultaneous screening of new sensing materials, and mapping spatially inhomogeneous reactions. Light-pulsed sensing combines many types of information, including the catalytic activity of the gate metals, gas flow turbulence, edge effects, etc. While not an optical detection technique, the method demonstrates the utility of employing light combined with electrochemical detection.

8.3 The Tufts Artificial Nose

Optical fibers can be used to create fluorescent-based optical sensors. In this approach, a fluorescent indicating species is attached to the fiber’s distal tip using a variety of immobilization techniques. Excitation light is introduced into the fiber, which carries light efficiently to the fiber’s distal tip. The fluorescent indicator is excited and some of the resulting isotropically emitted light is captured by the same fiber, directed through suitable optics, filtered and sent to a detector. The modulated light signal returning to the detector corresponds to the presence and amount of an analyte. In order to design a cross-reactive optical sensing array, it is necessary to find an appropriate array of sensing materials to respond to a wide variety of analytes. Our 192 8 Optical Electronic Noses

laboratory has developed a series of fluorescence-based optical sensors. It was our goal to create a fluorescent-based cross-reactive array. In 1991, we published a paper in which we used a solvatochromic indicator, Nile Red, to create a generic optical vapor sensor [4]. The sensor was based on immobilizing Nile Red within a polymer matrix and attaching the resulting material to the distal tip of an optical fiber. As discussed above, solvatochromic indicators report on the polarity of their local environment, also called the microenvironment. When solvatochromic dyes, such as Nile Red, are embedded in polymers, they report on the polarity of the polymer’s microenvironment indicated by their color, in particular, their absorption and/or emission spectra. For example, Nile Red has an emission spectrum that is relatively blue in nonpolar, hydrophobic environments, and is red in polar, hydrophilic environments. When an organic vapor sensor containing Nile Red, immobilized within a polymer, is in contact with air, it has an emission spectrum that represents the polarity of the polymer. When such a polymer is exposed to an organic vapor, the organic vapor diffuses into the polymer and modifies the microenvironmental polarity, which is signaled by a change in the emission spectrum of Nile Red. The emission spectrum shift is highly predictable. A vapor that is more polar than the polymer will shift the spectrum to a higher wavelength, whereas a less polar polymer will shift the spectrum to a lower wavelength (Fig. 8.4). The extent of the shift depends both on the polarity difference as well as the partition coefficient of the vapor into the polymer. In this manner, a generic organic vapor sensor was created by simply immobilizing a single solvatochromic dye within a dimethylsiloxane polymer. The sensor was used to detect leaks of hydrocarbon liquids from underground storage tanks by detecting the vapors that preceded the liquid leak. The same sensing principle was used to design a cross-reactive vapor-sensing array [35–37]. In this system, Nile Red was immobilized within a series of polymers. Hun- dreds of polymers were screened empirically. Each polymer defined the initial polarity

Fig. 8.4 The spectra of four sensors made by incorporating Nile Red into four polymers of differing polarity. The emission max shifts to the red with increasing polarity of the polymer matrix 8.3 The Tufts Artificial Nose 193 of the microenvironment as reported by Nile Red. These polymers were dip-coated onto the ends of individual optical fibers. Nineteen sensors were bundled into an array format. Upon exposure to an organic vapor, each polymer sensor absorbed vapor according to its partition coefficient for that vapor. The change in each sensor’s fluorescence spectrum depended on how much vapor partitioned into that sensor as well as the difference between the vapor’s and polymer’s polarities. There are several aspects of the optical sensor array’s operating mechanisms that require elaboration. First, we decided that, unlike most electronic noses, we would not look at static headspace measurements but rather would mimic a sniff by observing the kinetics of the response upon vapor exposure. To this end, we employed a vapor delivery system that was originally designed for delivering odors to animals in olfactory research [38]. The vapor delivery was accomplished by presenting square-wave vapor pulses for a defined period of time to the distal face of the bundled fiber array. Fluor- escence detection was accomplished by using a two-dimensional detector, a CCD camera, so that we could acquire fluorescent signals from all the sensors in the array simultaneously. To simplify signal detection, the fluorescence was collected at a single wavelength by interposing an emission filter between the fiber and the CCD chip. The resulting measured fluorescence signals coming from each sensor, upon exposure to organic vapors, were simply the intensity changes relative to their starting intensity at that particular emission wavelength. An intensity increase simply meant that the emission spectrum of the dye in a particular polymer upon exposure to a particular vapor was shifting closer to the wavelength range defined by the emission filter. Conversely, a decrease in fluorescence intensity indicated that the emission spectrum of the dye was moving further away from the emission filter range. A final aspect of the response mechanism resulted from the interaction of the vapor with the polymer. Some of the polymers exhibited a swelling effect in which the polymer volume increased as vapor partitioned into it. Polymer swelling causes a dye molecule to increase its average distance relative to the fiber surface. As described above, the isotropically emitted light is captured by the optical fiber. When a molecule moves further from the fiber surface, the capture efficiency for the light decreases because the sine of the half angle of the returning light is reduced. Therefore, the response of each sensor is due to a combination of vapor partitioning into the polymer, polarity differences between the polymer and the vapor, and polymer swelling. Because the solvatochromic and swelling effects operate under different kinetic regimes (i.e., swelling at the bulk polymer surface occurs rapidly while the solvatochromicity requires an intimate slower redistribution of vapor molecules within the polymer matrix), nonlinear effects can be observed. The fluorescence images are collected before, during, and after a vapor pulse to provide a characteristic response profile for each sensor in the array. A video image of an array of 19 sensors exposed to a three second pulse of benzene is shown in Fig. 8.5. The digitized responses of each sensor in the array are shown in the graph in Fig. 8.6. These complex temporal responses are characteristic of a benzene pulse at a particular concentration and can be used to train a computational classification program. Both parametric (e.g., intensities, slopes) and nonparametric methods can be used to train the responses. One of the major challenges in the field of electronic noses/cross-reactive arrays is array-to-array variability. This lack of reproducibility 194 8 Optical Electronic Noses

results from the inability to prepare polymeric materials identically. When polymers are put onto optical substrates or other surfaces by dip coating, liquid dispensing, photopolymerization, or electropolymerization, slight volume differences, initiator conditions, or minor heterogeneities can cause significant differences in material composition. These differences, even if the variation is only a few percent, can lead to loss in training fidelity. To address this problem, we have switched to a different array platform. Instead of using individual single-core optical fibers we now employ optical-imaging fiber arrays. These arrays are comprised of thousands of individual optical fibers, each of which is surrounded by a clad material (Fig. 8.7). The arrays are fabricated such that they are coherent in nature meaning that the position of an individual optical fiber within the array retains its position from one end to the other. In this manner, such arrays can be used to carry images, an application that is being pursued for medical endoscopy. These arrays are fused unitary bundles rather than mechanically fixed strands of individual fibers. Thus, they maintain their flexibility and can be handled similarly to single core fibers. A typical optical array contains between 10 000 and 50 000 individual fibers in a diameter of a few hundred microns with the individual fibers having diameters on the order of 3–5 microns each. The difference in materials composition between clads and cores provides a method for selectively etching the cores. When the polished distal tip of a custom optical imaging fiber array is placed into an acid etchant, the cores etch at a faster rate than the clads leading to an array of wells. At the bottom of each well is the distal face of an optical fiber (Fig. 8.8A). In this manner, each well is ‘optically wired’ to its own individual optical fiber. We discovered that latex or silica beads, matched in size to the dimensions of the individual wells, would spontaneously assemble into each well

Fig. 8.5 A sequence of images depicting the fluorescence response of a 19-fiber sensor array to a pulse of benzene vapor 8.3 The Tufts Artificial Nose 195

Fig. 8.6 Temporal plots from 19-fiber array response to benzene vapor pulse in a highly efficient self-organizing fashion. This approach could be used to create sensor arrays based on polymeric microspheres. Microsphere sensors can be created by taking monodisperse polymeric microspheres and swelling them in a suitable organic solvent containing dissolved Nile Red [39]. Upon removal from the solvent, evaporation of residual solvent occurs resulting in Nile Red being trapped within the polymeric matrix. Another class of bead sensors uses surface modified silica beads to which Nile Red is adsorbed (Fig. 8.9). Many different bead types can be prepared out of a variety of polymers and surface

Fig. 8.7 Components of a fiber-optic imaging bundle 196 8 Optical Electronic Noses

Fig. 8.8 A) wells formed by etching an imaging bundle, and B) beads immobilized in the wells

functional groups. As discussed above, in each of these sensors, the Nile Red reports on the polarity of its local environment. A library of bead types is created containing a diversity of responses to vapors. To create a sensing array, the desired individual bead types are mixed. 100 milligrams of beads contains approximately 10 billion beads. The beads are randomly distributed onto the distal face of an etched imaging fiber such that one bead occupies each well (Fig. 8.8B). In order to register the position of each bead in the array after fabrication, the fiber is connected to the optical imaging system and a vapor is pulsed onto the fiber’s sensor end. Because each different type of bead produces a unique and characteristic response profile when exposed to a particular vapor, the responses to the vapor pulse enable the image-processing program to register the bead type occupying each well. We refer to this registration protocol as ‘self-encoding’; that is, the sensor is identified by its response profile to a particular vapor [36]. In this manner, a library of beads can be used to create hundreds to thousands of individual sensing arrays with each array having the same bead types but located in different positions. The bead registration task involves exposing each array to a particular re-

Fig. 8.9 Silica beads can be modified in a variety of ways before being dyed in order to generate a diverse library of sensors 8.3 The Tufts Artificial Nose 197

Fig. 8.10 Signal-to-noise ratios can be dramatically improved by averaging over multiple copies of the same bead type within an array

gistration vapor and using an image-processing program to automatically register the position of each bead in the array using a lookup scheme. A key advantage of the self-encoding array sensors is that the training can be transferred from one sensor array to another. All the sensor beads of a particular type give virtually identical responses because they are all prepared at the same time. Thus, when mixed in a library, each bead type maintains its particular response profile. An- other important feature of these cross-reactive optical arrays is the built-in redundancy of each of the sensors. The small size of the fibers combined with the random distribution of the different microspheres in the array dictates that there will be replicates of each sensor in every array. The numbers of each sensor type will distribute themselves according to Poisson statistics. Replicates provide significant advantages in terms of signal-to-noise. The signal-to-noise ratio scales as 1/5 n, where n is equal to the number of sensors of each type. By summing or averaging sensor replicates, significant signal-to-noise enhancements can be achieved resulting in improved detection limits due to the ability to make more precise measurements at lower concentrations (Fig. 8.10) [36]. The microsphere arrays also have several other advantages such as flexibility of array types, scalability, and simple manufacturing. The major limitation with fluorescent dyes for optical sensor arrays is photobleaching. Upon exposure to light, any indicating material loses its intensity because of photooxidation. Over long periods of exposure, the light intensity degrades considerably. In order to avoid this problem, we employ autoscaled response profiles so that training is not dependent on absolute signal intensities. Despite this autoscaling procedure, photobleaching eventually degrades the signal-to-noise ratios. At this point, the array must be replaced. Since each array has the identical sensing elements, the training performed on one array is transferable to a second array. We have recently demonstrated training transfer of a classifier over a nine-month period with robustness of classification. 198 8 Optical Electronic Noses

8.4 Conclusion

Optical electronic noses have a relatively short history relative to conducting polymer or metal-oxide-based approaches. In the roughly five years since they were first reported, there have been a variety of advances in the types of optical signals employed as well as the materials used to perform the recognition [18]. The area of molecular recognition is burgeoning. Many of these receptors have built-in optical transduction. New polymers [14] and nanostructured materials [27] with recognition and optical signaling are being developed. In addition, the data richness of optical sensor arrays should make them attractive as analytical systems. With continued emphasis on new optical materials and devices development for the telecommunications and computer industries, combined with advances in molecular recognition and advanced materials, optical approaches to sensing should continue to improve in sensitivity, selectivity, and performance.

Acknowledgments The authors wish to thank the ONR and DARPA for research funding, and Keith Albert and Shannon Stitzel for assistance with figures.

Tab. 8.1 Summary table of optical electronic nose approaches.

Transduction Description References Mechanism

Luminescence Fiber-optic sensors using polarity sensitive fluorophores such 4–7, 35–37 as solvatochromic or TICT dyes. Randomly assembled solvatochromic bead arrays. 39, 40 Host-guest supramolecular chemistry: shifts in wavelength and 8 intensity of ‘fluoroclathrands’ based on packing density changes caused by vapors. Chemiluminescence-based detection, using luminol and Rubpy 9 dyes. Colorimetric Color changes of an oxazine perchlorate dye coated on glass 10 capillaries. Bromothymol blue in Nafion polymer layers. 11 Thermal printer paper as vapor sensors. 12 Inorganic sensing materials (e. g. Pt-Pt compounds): color 16 changes caused by perturbation of stacking in charged complexes. Metalloporphyrins: formation of coordination complexes with 17, 18 analytes, and use of different metals for changing sensing properties. Surface plasmon Method for detecting changes in refractive index at a surface. 20–22 resonance Interference, Reflective interferometric ipectroscopy for detecting changes 23–26 reflection in optical thickness of polymer layers. Interference measurements using chemically etched porous 27 silicon chips. Mass loading Detecting mass changes on resonating atomic force microscope 29, 30 microcantilevers. 8.4 Conclusion 199

References

1 K. J. Albert, N. S. Lewis, C. L. Schauer, 22 A. Abdelghani et al.. Anal. Chim. Acta., 1997, G. A Sotzing, S. E. Stitzel, T. P. Vaid, 337, 225–232. D. R. Walt. Chem. Rev., 2000, 100, 23 K. Spaeth, G. Kraus, G. Gauglitz. Fresenius’ 2595–2626. J.Anal. Chem., 1997, 357, 292. 2 K. Persaud, G. Dodd. Nature, 1982, 299, 24 Y. Liu et al.. Optical Sensor Apparatus 352–355. for Detecting Vapor of Organic Solvent, 3 J. Lakowicz. Principles of Fluorescence Spec- EU95203669.7. 1995. troscopy, Plenum Press, New York, 1982. 25 R. Reichl, R. Krage, C. Krummel, 4 S. Barnard, D. R. Walt. Environ. Sci. Technol, G. Gauglitz. Appl. Spectrosc., 2000, 54, 1991, 25, 1301–1304. 583–586. 5 G. Orellana et al.. Anal. Chem., 1995, 67, 26 K. Bodenhofer et al.. Nature, 1997, 387, 2231–2238. 577–580. 6 C. Hubert, D. Fichou, F. Garnier. Adv. 27 S. Letant, S. Content, T. Tan, F. Zenhausern, Mater., 1995, 11, 914–917. M. Sailor. Sens. Actuators B, 2000, 69, 7 H. Krech, S. L. Rose-Pehrsson. Anal. Chim. 193–198. Acta., 1997, 341, 53–62. 28 A. Yasser, B. Lawrence. Sensors, 1996, April, 8 T. H. Brehmer, P. P. Korkas, E. Weber. Sens. 76–77. Actuators B, 1997, 44, 595–600. 29 T. Thundat et al.. Anal. Chem., 1995, 67, 9 G. E. Collins, S. L. Rose-Pehrsson. Sens. 519–521. Actuators B, 1996, 34, 317–322. 30 H. P. Lang, et al.. Appl. Phys. Lett, 1998, 72, 10 Giuliani et al.. Reversible Optical Waveguide 383–385. Vapor Sensor, US4513087. April 23, 1985. 31 M. Lofdahl, M. Eriksson, I. Lundstrom. Sens. 11 J. R. Stetter, J. Maclay, D. S. Ballantine. Actuators B, 2000, 70, 77–82. Optical Waveguide Vapor Sensor, US5315673. 32 F. Winquist, H. Sundgren, E. Hedborg, May 24, 1994. A. Spetz, I. Lundstro¨m. Sens. Actuators B, 12 H. E. Posch, O. S. Wolfbeis, J. Pusterhofer. 1992, B6, 157–168. Talanta, 1988, 35, 89–94. 33 I. Lundstro¨m, R. Erlandsson, U. Frykman, 13 L. Gheorghiu, W. R. Seitz, D. Arbuthnot, E. Hedborg, A. Spetz, H. Sundgren, J. L. Elkind. SPIE Conference on Environ- S. Welin, F. Winquist. Nature, 1991, 352, 47. mental Monitoring and Remediation Techno- 34 I. Lundstro¨m, H. Sundgren, F. Winquist. logies II, 1999, 3853, 296–302. J. Appl. Phys., 1993, 74, 6953–6962. 14 D. T. McQuade, A. E. Pullen, T. M. Swager. 35 J. White, J. S. Kauer, T. A. Dickinson, Chem. Rev., 2000, 7, 2537–2574. D. R. Walt. Anal. Chem., 1996, 68, 15 F. L. Dickert, A. Haunschild, Adv. Mater., 2191–2202. 1993, 12, 887–895. 36 T. A. Dickinson, D. R. Walt, J. White, 16 C. A. Daws, C. L. Exstrom, J. R. Sowa, Jr., J. S. Kauer. Anal. Chem., 1997, 69, K. R. Mann. Chem. Mater., 1997,9, 3413–3418. 363–368. 37 T. A. Dickinson, J. S. White, J. S. Kauer, 17 A. D’Amico et al.. Sens. Actuators B, 1999, 65, D. R. Walt. Nature, 1996, 382, 697–700. 209–215. 38 J. Kauer, G. Shepherd. J. Physiol., 1977, 272, 18 N. A. Rakow, K. S. Suslick. Nature, 2000, 495–516. 406, 710–713. 39 T. A. Dickinson, K. L Michael, J. S. Kauer, 19 R. Casalini et al.. Sens. Actuators B, 1999, 57, D. R. Walt. Anal. Chem., 1999, 71, 28–34. 2192–2198. 20 R. W. Nelson, J. R. Krone, O. Jansson. Anal. 40 K. J. Albert, D. R. Walt. Anal. Chem., 2000, Chem., 1997, 69, 4369–4374. 72, 1947–1955. 21 BIAcore Probe literature, Pharmacia Bio- sensor. http://www.biacore.com. 201

9 Hand-held and Palm-Top Chemical Microsensor Systems for Gas Analysis

A. Hierlemann, U. Weimar, and H. Baltes

Abstract The characteristics and fundamentals of hand-held chemical sensor units for gas analysis are described, commercially available systems based on conventional sensor technology are briefly portrayed, and the emerging field of microsensors and microsensor systems based on planar integrated circuit (IC) technology and their use in hand-held instruments is detailed. Conventional sensor technology is at the base of most hand-held instruments in research and on the market to date. Systems based on mass-sensitive sensors and on electrochemical sensors (chemoresistors) are presented. They are used to detect organic volatiles and rely on changes of physical properties of polymeric layers upon volatile absorption. The same polymers can be used with microsensors based on silicon or IC technology. These microsensors offer substantial advantages such as low power consumption, a very crucial issue in battery-operated systems, small size, rapid response, and batch fabrication at industrial standards and low costs. The present state of the art in IC-based microsensors is summarized and the inclusion of such sensors into hand-held systems is shown.

9.1 Introduction

The first hand-held systems, which are still available on the market [1, 2], were tubes or badges. They are lightweight, inexpensive, disposable devices based on diffusion exposure. They provide an immediate visual indication when a specific chemical hazard is present. They mostly include an indicator layer or impregnated paper, which provides homogeneous and stable color formation or color change upon presence of the target compound. These devices are not continuously operating, exhibit irreversible characteristics, are disposable, and, therefore, are usually referred to as dosimeters [3, 4] rather than as chemical sensors. At present, there are two different categories of sensor-based (i.e., non-disposable), continuously operating hand-held instruments. The first category includes personal

warning and safety systems as an advancement of the already mentioned tubes and badges, which have been on the market for quite some time [5], second there are recently developed multisensor-systems with onboard pattern recognition and/or multicomponent analysis algorithms sometimes denoted as “electronic noses” [6]. The distinction between those two types of devices is mainly due to differences in the instrument architecture or complexity, and in the target applications.

Key requirements for both types of hand-held instruments include:

* Ease of use * Ruggedness * Low power consumption * Low cost and low maintenance * Short recovery and response times * Long-term stability (low drift) and reliability (self-calibration)

For the personal safety devices of the first category, it is desirable that the system also exhibits

* High sensitivity and low limit of detection (LOD) * High selectivity to target analyte and low cross-sensitivity to interferants

The hand-held personal safety devices include in most cases only one or two sensors specifically engineered to detect selected individual gaseous compounds at trace level [1, 5]. Upon reaching a threshold value, the devices issue a warning or an alarm. The device calibration is univariate, i.e., the devices are calibrated using pure gases. Their applications include the detection of toxic or explosive gases in all branches of industry, the measurement of hazardous substances during firefighting operations, and the detection of airborne contaminants such as carbon monoxide, hydrazine, ammo-

Fig. 9.1 Typical hand-held gas warning system (PAC III by Draeger, Luubeck,€ Germany [5]) detecting carbon monoxide (CO). By exchanging the sensor, hydrogen sulfide or oxygen can be detected. Reprinted with kind permission of Draeger 9.2. Conventional Hand-held Systems 203 nia, hydrogen sulfide or hydrides. Figure 9.1 displays a carbon monoxide monitor based on a single electrochemical cell [5]. Electrochemical cells are predominantly used since their sensitivity to anything other than the desired compounds are in most cases negligible. The second category of hand-held instruments includes a sensor array with different coatings on the same type of transducer or even different types of transducers. The target compounds are individual gases or a multitude of gaseous and volatile compounds generating a characteristic fingerprint response pattern. Pattern recognition and multi-component analysis algorithms rely on multivariate calibration. Training and gas phase analysis are restricted to a defined sample set, which has to be calibrated prior to instrument use. Such multisensor systems are the main topic of this chapter and book. Target applications of these hand-held instruments include quality and process control in industrial settings (food processing, packaging, raw material inspection), aroma and odor identification, environmental monitoring, hazardous material identification, and some medical pilot applications. Conventional sensor technology is at the base of most hand-held instruments in research and on the market to date. Recently, microsensors based on silicon or integrated circuit (IC) technology have been developed [7–11], which offer substantial advantages such as low power consumption, a very crucial issue in battery-operated systems, small size, rapid response, and batch fabrication at industrial standards ensuring a high level of sensor-to-sensor reproducibility, quality, and inferring low costs. Additional features include the possibility of on-chip signal conditioning or data preprocessing [12, 13]. In the following, we will describe the characteristics and fundamentals of hand-held instruments, then detail the approach using conventional sensor technology by briefly portraying commercially available systems, and finally we will describe the emerging field of hand-held instruments relying on microsensors and microsensor systems based on planar IC technology.

9.2 Conventional Hand-held Systems

9.2.1 Hardware Setup

A schematic of a hand-held instrument comprising all vital components used in commercial and research-type instruments is shown in Fig. 9.2. The hand-held instrument consists of two major blocks. The upper part represents the gas intake unit with pumps, valves, filter, and the measurement chamber. The bottom “electronic” part includes the sensors, sensor electronics, power pack, the data processing unit with display, and the communication interface. In some cases it is sufficient to rely on diffusion of the analyte molecules to the sensors. This “passive” sampling does not require any of the components in the 204 9 Hand-held and Palm-Top Chemical Microsensor Systems for Gas Analysis

Fig. 9.2 Typical setup of a research-type or commercial hand-held instrument. The upper part represents the gas intake unit with pumps, valves, filter, and the measurement chamber. The bottom electronic part includes the sensors, sensor electronics, power pack, the data processing unit with display, and the communication interface

gas intake unit, only an opening of the sensor chamber to the ambient allowing for fast in-diffusion of the analyte. For “active” sampling, pumps, valves, and filters are required. An active sampling stage is realized in all commercially available systems since the gas phase composition in the sensor chamber is less subject to fluctuations and can be much better controlled. Different operation modes for an active sampling unit have been implemented:

* Pumping only: test gas is pumped into the measurement chamber and pumped out through an exhaust (flow-through), or through the inlet by reversing the pump direction. * Pumping and valving: test gas is pumped to the sensors either from separate inlets for reference gas and analyte gas or by routing a fraction of the analyte gas through an on-board filter unit.

Pumping and valving require a more sophisticated intake unit design but offers the advantage of re-establishing the baseline of the sensors using a filtered or pure purge or reference gas. This leads to better recognition of drift effects and sensor malfunctioning. The bottom part of the schematic represents the electronic part of the instrument. The sensor array is mounted in a measurement chamber and connected to a microcontroller. The AD/DA and I/O channels of the microcontroller can be directly used, or dedicated electronic components can be added. The sensor-microcontroller connection depends on the number of channels and the targeted sensor resolution. Novel research-type instruments exhibit digital communication between sensor array and 9.2. Conventional Hand-held Systems 205 microcontroller, which requires additional electronics (A/D conversion, bus interface) on the sensor side. The microcontroller usually hosts pattern recognition algorithms (KNN, PCA, see Chapter 6), allows for storing calibration data or analyte pattern libraries, and enables logging a limited amount of acquired measurement data. Batteries or accumulators usually power the hand-held instrument. The capacity of a typical battery (or accumulator) is of the order of 5 to 15 Wh. Therefore, the average power needed to operate the instrument should be below 1 W to ensure a decent operation time. Different types of displays are used to communicate the desired information to the user. The simplest display can be realized by a red and green LED indicating a binary decision, e.g., the membership to a class such as “sample o.k.” or “not o.k.”. More information is provided by alphanumeric displays (one or several lines) presenting qualitative (e.g. classification) and quantitative (e.g. analyte concentration) results. Gra- phic displays are very versatile and comfortable and can show, for example, complete PCA plots. All currently available systems are also equipped with a computer interface to con- nect to an external PC for downloading measurement data to the computer and trans- ferring, for example, calibration data to the hand-held instrument. The most common interfaces are RS-232 (serial) and, more recently, universal serial bus (USB). In the near future infrared interfaces and BluetoothTM systems enabling wireless communication will be introduced.

9.2.2 Fundamentals of the Sensing Process

All commercially available hand-held units, and the CMOS chemical microsensors (detailed in Section 9.2) rely on polymeric coatings as sensitive films for the detection of volatile compounds in air. The predominant sensing mechanism is hence physisorption and bulk dissolution of the analyte molecules within the polymer volume. Upon absorption of analyte by the coating, the physical properties of the polymer film, such as its mass, volume or dielectric constant, change. Considering bulk dissolution in polymers, all effects are based on thermodynamics and/or kinetics. High sensor selectivity (strong interaction) and perfect reversibility (weak interaction) impose conflicting constraints on the design of the sensitive layer. For ensuring reversibility, polymers showing partial selectivity to some of the detected species are commonly used. The desired identification of the compounds is then achieved by using an array of different partially selective sensors and applying numerical methods of data evaluation (see Chapter 6) [14–18]. At “infinite dilution”, that is an analyte partial pressure below 3 % of its saturation vapor pressure at the sensor operation temperature, one usually observes a linear correlation between the analyte concentrations and the sensor signals. In this low-concentration range, Henry’s law still holds [19]. Therefore, it is possible to calculate partition coefficients at “infinite dilution”, K, as characteristic thermodynamic equilibrium constants for a certain organic volatile dissolved in a polymer: 206 9 Hand-held and Palm-Top Chemical Microsensor Systems for Gas Analysis

K ¼ cpoly=cA ð9:1Þ

Here, cpoly and cA denote the analyte concentration in the polymer and the gas phase. The partition coefficient, K, is hence a dimensionless “enrichment factor” relating the concentration of a species in the sensing layer to that in the probed gas phase. K is usually of the order of 200 to 5000 for most analytes sorbed into standard polymers. Ideally, an experimentally determined partition coefficient should be constant for the chosen analyte/polymer combination, and independent of the transducer principle [20].

9.2.3. Commercially Available Instruments Based on Conventional Technology

In this section, a brief overview will be given on commercially available hand-held instruments, which are all based on conventional sensor technology (see Chap- ter 4). These hand-held systems weigh between 0.5 and 1 kg and are battery or accumulator operated. All hand-held units feature basic pattern recognition software (PCA, KNN, etc., for details, see Chapter 6) and have some on-board data storage possibility as well as a RS 232 serial interface to communicate with external equipment such as laptops or computers. The devices are specified to operate in a temperature range between 263 and 323 K. The devices feature a LCD display (some include even a graphic display) to show the result of the sensor analysis or pattern recognition. The keyboard is in all cases very simple and allows for using the hand-held device with only a few commands. All instrument producers are providing additional data logging and storage software for external PCs. The individual configuration of this software depends on the user needs. The gas bus and sensor array are different for the various systems and will be described in more detail below. Fig. 9.3 shows three of the currently commercially available systems, the VOCcheck by AppliedSensor [21], the VaporLab by Microsensor Systems [22], and the Cyranose 320 by Cyrano Sciences [23]. A summary of their characteristic features is given in Table 9.1.

Fig. 9.3 Currently commercially available systems: the VOC- check by AppliedSensor [21], the VaporLab by Microsensor Systems [22], and the Cyranose 320 by Cyrano Sciences [23]. Reprinted with kind permission of AppliedSensor, Microsensor Systems, and Cyrano Sciences 9.2. Conventional Hand-held Systems 207

Tab. 9.1 Characteristic features of three commercially available hand- held units: VOCcheck by AppliedSensor [21], VaporLab by Micro- sensor Systems [22], and Cyranose 320 by Cyrano Sciences [23].

Features VOCcheck [21] VaporLab [22] Cyranose 320 [23]

Number of Sensors 4 4–6 32 Transducer Type Thickness Shear Mode Surface Acoustic Wave Chemoresistor Array Resonator Device Sensitive Layer Polymer Polymer Carbon-Loaded Polymer Target Analytes Organic Volatiles Organic Volatiles Organic Volatiles Response Time < 15 s < 1 s 10 s Operating Temperature 10 8C–408C58C–408C08C–408C Weight 400 g 570 g 910 g Dimensions 180 82 53 mm 180 85 56 mm 220 100 50 mm On-board Software Application-Specific Not Specified KNN, Kmeans, PCA, CDA Sampling Stage Pump Pumps and Valves Pump and Valves Display 12 4 Alphanumeric 119 73 Pixel LCD 320 200 Graphic Backlight Backlight Battery Life 6–12 Hours 4–12 Hours 3 Hours Data Log Capacity 100 Measurements 100 Measurements 100 Measurements Interface RS 232 RS 232 RS 232, USB

Since there are severe constraints on overall size, weight and power consumption, the performance of hand-held devices is generally inferior to that of bench-top instruments. Precision and the LOD are degraded by a factor of approximately 10 as compared to a bench-top setup due to the less effective temperature stabilization and due to the use of less bulky measurement and recording electronics such as no precision voltage sources, less-sophisticated counter modules, and off-the-shelf electronic parts. The hand-held units are less pricy and intentionally designed to suit only a few specific applications, in consequence their versatility is rather limited. The unit configuration usually has to be optimized with regard to the target application. In the following we will briefly describe the characteristics of the commercially available systems as well as the underlying transducer principle, for more details we refer to Chapter 19, where some more information on these systems can be found.

9.2.3.1 Hand-held Units Based on Mass-Sensitive Sensors Mass-sensitive sensors are in the simplest case gravimetric sensors responding to the mass of species accumulated in a sensing layer. Some of the sensor devices are additionally capable of detecting changes in a variety of other properties of solid or liquid media in contact with their surface such as polymer moduli, liquid density and visc- osity [24–26], which will not be discussed here. The high sensitivity of mass-sensitive sensors provides good chemical sensitivity: mass changes in the picogram range or lower can be detected and ppm (parts per million) to ppb (parts per billion) detection levels have been reported, for example, for gas and vapor sensors [24, 26]. Most of the mass-sensitive sensors rely on piezoelectric materials such as quartz, lithium tantalate or niobate, aluminum nitride, zinc oxide and others. Piezoelectricity results in general 208 9 Hand-held and Palm-Top Chemical Microsensor Systems for Gas Analysis

from coupling of electrical and mechanical effects. The prerequisite is an anisotropic, noncentrosymmetric crystal lattice. Upon mechanical stress, charged particles are dis- placed thus generating a measurable electric charge in the crystal. In turn, mechanical deformations can be achieved by applying a voltage to such a crystal (for details, see [27]). Using an alternating current (AC), the crystals can be electrically excited into a fundamental mechanical resonance mode. The resonance frequency, which is the recorded sensor output in most cases, changes in proportion to the mass loading on the crystal or device. The more mass (analyte molecules) that is absorbed, for example in a polymer coated onto a piezoelectric substrate or transducer, the lower is the resonance frequency of the device [28, 29]:

2 Df ¼Cf0 Dm=A ð9:2Þ

Df here denotes the frequency shift due to the added mass (in Hz), C is a constant, f0 is the fundamental frequency of the quartz crystal (in Hz) and Dm/A the surface mass loading (in g cm2). The two most common mass-sensitive sensors are the thickness shear mode resonator (TSMR) and the Rayleigh surface acoustic wave (SAW) device (Fig. 9.4), which will be detailed below.

VOCcheck of Applied Sensor [21] This instrument relies on four discrete polymer-coated quartz TSMRs operating at a fundamental frequency of 30 MHz. TSMRs typically consist of circular quartz plates with thin metal (gold) films on both sides. By applying an AC voltage to the electrodes, bulk waves are generated that travel perpendicular to the plate surfaces, the wavelength of which is determined by the plate thickness. Both faces of the quartz plate execute a shear motion (see Fig. 9.4a). For more details on TSMRs, see [20, 24–26]. The system is equipped with an active sampling stage driven by a pump. There is no reference gas or analyte filter on board. Different sampling probes can be attached to the input port of the instrument. A single analyte identification cycle takes about 15 s. The battery pack is designed to enable more than 1000 measurement cycles. This

Fig. 9.4 Schematics of mass-sensitive devices. (a) Quartz plate with electrodes on both sides excited into a shear mode by applying AC. (b) Launching, propagation and detection of a Rayleigh-type surface acoustic wave by applying AC to interdigitated transducers 9.2. Conventional Hand-held Systems 209 allows operation, for example, over 14 hours performing one measurement per minute. Four different data evaluation methods can be stored in the system. Each method is designed to identify up to 15 analyte classes [21]. An important consequence of the simple thermodynamic bulk physisorption mechanism of polymer-coated TSMRs (the same holds for SAWs detailed below) is the good sensor-to-sensor reproducibility as can be taken from Table 9.2 [30]. A signal intensity value, which is composed of the weighted signals of the four TSMRs is displayed for three different VOCcheck systems. The TSMRs in the three systems are coated with the same polymers at identical layer thickness. All three systems repeatedly recorded signals upon exposure to 1000 ppm toluene, 500 ppm anisole, and 2000 ppm propan-1-ol. The intensity value, its standard deviation in three subsequent measurements and the similarity (based on normalized feature vectors, max. 100) in reference to system 1 are given. The standard deviation is less than 6 %, the similarity 93 and better in all cases, which ensures a good system-to-system transferability of results and calibration methods. The VOCcheck is targeted at rapidly identifying organic volatiles in air using its internal library, which can be customized for each application. The applications include inspecting incoming chemicals or containers, identifying solvents and chemicals, verifying the shelf life of products, detecting leaks and monitoring waste and emissions [21].

Tab. 9.2 System-to-system-reproducibility: TSMRs belonging to three different instruments were coated with the same polymers at identical layer thickness. All three systems repeatedly recorded signals upon exposure to 1000 pm toluene, 50 pm anisole, and 2000 pm propan-1- ol. The intensity value, its standard deviation in three subsequent measurements and the similarity (max. 100) in reference to instrument 1 are given (Table courtesy of AppliedSensor GmbH, Reutlingen, Germany).

Analyte Instrument 1 Instrument 2 Similarity Instrument 3 Similarity (reference) (vers. ref.) (vers. ref.)

Toluene 1000 ppm 329 301 97 317 96 Toluene 1000 ppm 341 280 96 316 97 Toluene 1000 ppm 339 289 98 328 97 Standard deviation 5.2 8.6 5.4 Anisole 500 ppm 1020 978 98 917 93 Anisole 500 ppm 1027 1056 98 1011 93 Anisole 500 ppm 1031 1104 98 1008 93 Standard deviation 4.5 52.0 44.0 Propan-1-ol 2000 ppm 172 154 96 166 95 Propan-1-ol 2000 ppm 165 139 97 158 98 Propan-1-ol 2000 ppm 171 135 97 154 98 Standard deviation 3.1 8.2 4.9 210 9 Hand-held and Palm-Top Chemical Microsensor Systems for Gas Analysis

VaporLab of Microsensor Systems [22] The VaporLab instrument is based on an array of typically four (up to 6) Rayleigh- SAW sensors. The schematic of a SAW device is shown in Fig. 9.4b. By applying an AC voltage to a set of interdigital transducers patterned on a piezoelectric substrate with appropriate orientation of the crystal axes, one set of the fingers moves downwards, the other upwards, thereby creating a mechanical surface deformation. This surface deformation generates an acoustic wave, which propagates along the surface and is converted back into an electrical signal by deforming the surface in the region of the receiving transducer. The electrical signal of the receiving transducer is recorded and represents the sensor signal. The SAW devices are coated with thin polymeric films. For more details on SAW devices, see refs [20, 24–26]. The instrument includes a sophisticated sampling system with up to two pumps, two three-way valves and two preconcentrators, which can be thermally desorbed. The measurement cycle time is less than 30 s with a sensor response time of less than 1 s. The battery pack allows for operation between 4 and 12 hours depending on the device configuration. Up to 200 vapor patterns can be stored in the on-board library and 100 measurements can be logged on the instrument [22]. Applications of VaporLab include quality control of packaging, food and beverage freshness, identification of hazardous chemicals, inspection of raw materials, process control in the food and petrochemical industries, aroma identification of products, monitoring flavor formulation in the food industry, and environmental monitoring of VOCs [22]. A variant of the VaporLab system is the HAZMATCAD (hazardous material chemical agent detector) system, which relies on both SAW devices and up to 3 electrochemical sensors (for details on electrochemical sensors, see Chapter 4). The rugged system features fast-mode (20 s response time) and sensitive-mode (120 s) operation. The mission life is 8 hours in the fast and 12 hours in the sensitive mode. The data including alarm level, time and date can be logged for 8 hours. The system exhibits an additional infrared data port and is targeted at detecting chemical warfare agents such as nerve and blister agents, blood and choking agents, as well as toxic industrial chemicals such as hydrides (arsine, silane), halogens (chlorine, fluorine) and acidic gases (sulphur dioxide) at trace levels [22].

9.2.3.2 Hand-held Units Based on Chemoresistors Chemoresistors rely on changes in the electric conductivity of a film or bulk material upon interaction with an analyte. Conductance, G, is defined as the current, I (A), divided by the applied potential, U, (V). The unit of conductance is X1 or S (Sie- mens). The reciprocal of conductance is the resistance, R,(X). The resistance of a sample increases with its length and decreases with its cross-sectional area. Conductometric sensors are usually arranged in a metal-electrode-1/sensitive layer/ metal-electrode-2 configuration [27]. The conductance measurement is done either via a Wheatstone bridge arrangement or by recording the current at an applied voltage in a DC mode or in a low-amplitude, low-frequency AC mode to avoid electrode polarization (for more details on chemoresistors, see Chapter 4). 9.3 Silicon-Based Microsensors 211

Several classes of predominantly organic materials are used for application with chemoresistors at room temperature such as conducting polymers or carbon black- loaded polymers. The chemically sensitive layer is applied over interdigitated electrodes on an insulating substrate. Electrode spacing is typically 5 to 100 lm, and the total electrode area is a few mm2. The applied voltage ranges between 1 and 5 V. Carbon-Black-Loaded Polymers exhibit conducting carbon black particles dispersed in non-conducting polymers. The conductivity is by particle-to-particle charge percolation so that if the polymer absorbs vapor molecules and swells, the particles are, on average, further apart and the conductivity of the film is reduced [31–33]. When the sensor is purged with clean air, the analyte desorbs from the polymer volume, the film shrinks and the conductive pathways are re-established [34]. The conductivity of the sensors depends critically on the morphology of the sensitive layer, i.e., the average distance of the dispersed particles, which involves high demands on sensor-to-sensor reproducibility.

Cyranose 320 of Cyrano Sciences [23] The Cyranose 320 relies on a 32-channel carbon-black polymer composite chemiresistor array [31–33]. A sampling pump and an inlet probe is provided with the system. The system response time is approximately 10 s. The battery pack allows for 3 hours of operation. Two classification methods with 6 classes per method can be stored on board. A maximum of 100 identifications can be saved on the instrument. A universal serial bus (USB) will be available in future software upgrades [23]. Typical applications of the Cyranose 320 include spot testing or continuous monitoring of batch-to-batch consistency and spoilage in raw food materials, solvent verification, identification of organic acids in waste water streams, recognition of gasoline, diesel and crude oil contamination in recycled containers, or enabling quick assessment of the chemical status of industrial processes in the food (coffee roasting and fermentation), petrochemical (plastics manufacture and gasoline blending) and consumer products sector (detergents and deodorants). Possible medical applications include obtaining information on the identity of certain chemical compounds in exhaled air and excreted urine or body fluids related to specific metabolic conditions, certain skin diseases or bacterial infections, such as those common to leg or burn wounds [23].

9.3 Silicon-Based Microsensors

Semiconductor technology provides excellent means to meet some of the key criteria of chemical sensors such as rapid response, low cost, batch fabrication, and offers additional features such as small size, and on-chip signal processing. The rapid development of integrated circuit technology during the past few decades has initiated many initiatives to fabricate chemical sensors consisting of a chemically sensitive layer on a signal-transducing silicon chip [35, 36]. Multi-chip solutions with electronics and sensors on separate chips have been proposed [37, 38]. 212 9 Hand-held and Palm-Top Chemical Microsensor Systems for Gas Analysis

The largely two-dimensional integrated circuit (IC) and chemical sensor structures processed by combining lithographic, thin film, etching, diffusive and oxidative steps have been recently extended into the third dimension using micromachining technologies – a combination of special etchants, etch stops and sacrificial layers [7–11]. A variety of micromechanical structures including cantilever beams, suspended membranes, freestanding bridges, etc. have been produced using micromachining technology (MicroElectroMechanical Systems, MEMS) [7–11]. MEMS technology thus provides a number of key features, which can serve to enhance the functionality of chemical sensor systems. In a further step, microelectronics and micromechanics (MEMS-structures) have been realized on a single chip allowing for on-chip control and monitoring of the mechanical functions as well as for data preprocessing such as signal amplification, signal conditioning, and data reduction [7–11, 39–42]. In the following, a short introduction into micromachining technology (Sec- tion 9.3.1) will be given, and three different types of CMOS-based (CMOS, complementary metal-oxide semiconductor, is a standard IC fabrication process) transducers will be described (Section 9.3.2 to 9.3.4), which serve as components of a single-chip system. The multisensor-chip, which forms integral part of a hand-held microsensor- based gas detection unit, will be detailed in Section 9.3.5.

9.3.1 Micromachining Techniques

9.3.1.1 Bulk Micromachining One approach to enhance the functionality of IC-based substrates includes micromachining the bulk substrate, which in most cases consists of silicon. Silicon can be dry or wet etched by various techniques [9–11, 43]. Some wet etchants such as nitric acid/ hydrofluoric acid lead to isotropic etching – the same etch rate in all directions, others such as potassium hydroxide lead to anisotropic etching, that is they preferentially etch away the silicon along certain crystal planes while preserving it in other directions (Fig. 9.5a). Typical structures obtained by, for example, anisotropic wet etching through the complete bulk silicon of a CMOS wafer include membranes consisting of the remaining dielectric CMOS layers. The thermal oxide serves as an etch-stop layer. The resulting membrane structures can be used for sensors requiring excellent thermal insulation, such as calorimetric chemical sensors or semiconductor-oxide-covered microhotplates. Polysilicon and metal structures sandwiched in-between the dielectric layers can be used to create, e.g., thermopiles and heating resistors [13]. Another technique is dry etching. Again, there is isotropic etching performed by using, for example, xenon difluoride or anisotropic etching by reactive-ion-etching (RIE). RIE is used, for example, to release cantilevers or create bridge structures from preformed membranes [44]. 9.3 Silicon-Based Microsensors 213

Fig. 9.5 Micromachining techniques: (a) bulk micromachining, anisotropic and isotropic etching, (b) surface micromachining with sacrificial layer, structural layer and a subsequent etch step

9.3.1.2 Surface Micromachining Surface micromachining comprises a number of techniques to produce microstructures from thin films previously deposited onto a substrate and is based on a sacrificial layer method (Fig. 9.5b). In contrast to bulk micromachining, surface micromachining leaves the substrate intact. A sacrificial layer is deposited and patterned on a substrate. After that, a structural thin film, in most cases polysilicon, is deposited and patterned, which will perform the mechanical or electrical functions in the final device. A selective etchant then removes exclusively the sacrificial layer material. The thickness of the sacrificial layer determines the distance of the structural parts from the substrate surface. Common sacrificial layer materials include silicon oxide etched by hydrogen fluoride and aluminum etched by a mixture of phosphoric, nitric and acetic acid. Clamped beams, microbridges, or microchannels can be fabricated this way, micro- rotors and even microgears can be realized by repeated layer deposition and etching [9–11, 43].

9.3.2 Microstructured Chemocapacitors

Chemocapacitors (dielectrometers) rely on changes in the dielectric properties of a sensing material upon analyte exposure (chemical modulation of the capacitance by changes in the dielectric constant of the sensitive layer). Interdigitated structures rather than plate capacitors are predominantly used [45–47]. In some cases, plate- capacitor-type structures with the sensitive layer sandwiched between a porous thin metal film (permeable to the analyte) and an electrode patterned on a silicon support are used [48, 49]. The capacitances are usually measured at an AC frequency of a few kHz up to 500 kHz. The size of the capacitor shown in Fig. 9.6 is 800 800 lm2, its electrode width and spacing are 1.6 lm. Since the nominal capacitance of this interdigitated capacitor is in the range of few picoFarad, and the expected capacitance changes upon analyte absorption are in the attoFarad range, a dedicated measurement configuration and specific signal conditioning circuitry had to be developed. The sensor response is read out as the differential signal between a polymer-coated sensing and a nitride-passivated reference capacitor. Both sensor and reference capacitors are split into two parts to improve the charge transfer efficiency. The sensing capacitor, (CS), and the reference 214 9 Hand-held and Palm-Top Chemical Microsensor Systems for Gas Analysis

Fig. 9.6 Micrograph of a capacitive sensor system including a polymer-coated sensing capacitor (S), a passivated reference capacitor (R), four interdigitated feedback capacitors (FB), and the Sigma-Delta circuitry (RD)as detailed in the text. Reprinted from [50] with permission

capacitor, (CR), are incorporated in the first stage of a fully differential second-order Sigma-Delta-modulator (Fig. 9.7) with two switched-capacitor-integrators and a subsequent comparator [50, 51]. A second-order Sigma-Delta-modulator is used to achieve shorter analog-to-digital conversion time. Since the output bit stream of the Sigma-

Delta-modulator is proportional to the ratio (CS-CR)/(Cfb), the four feedback capacitors (FB in Fig. 9.6, Cfb in Fig. 9.7) are realized as interdigital capacitors with the same materials as the sensing and reference capacitors in order to eliminate differences in temperature behavior and ageing. Due to the small signal bandwidth, the output bit stream of the Sigma-Delta-modulator is decimated using a frequency counter. For more details on circuitry see [50, 52]. Two effects change the capacitance of a polymeric sensitive layer upon absorption of an analyte: (i) swelling and (ii) change of the dielectric constant due to incorporation of the analyte molecules into the polymer matrix [52, 53]. For a simple interdigitated

Fig. 9.7 Schematic of the fully differential second-order Sigma-Delta- modulator exhibiting two switched-capacitor-integrators and a subse-

quent comparator. Four feedback capacitors (Cfb) are realized as interdigital capacitors. The Sigma-Delta-modulator provides a pulse- density-modulated digital output that is decimated using the frequency counter 9.3 Silicon-Based Microsensors 215 structure, the space containing 95 % of the field lines includes the polymer volume within a layer thickness of half the periodicity of the electrodes [54]. For a layer thickness less than half the periodicity, swelling of the polymer upon analyte absorption always results in an increase of the measured capacitance regardless of the dielectric constant of the analyte (Eq. 4). This results from the increased polymer/analyte volume within the field line region exhibiting a larger dielectric constant than that of the substituted air. The capacitance change for a polymer layer thicker than half the periodicity of the electrodes is determined by the ratio of the dielectric constants of analyte, eA (the analyte is assumed to be in the liquid state) and polymer, epoly (Eq. 3). If the dielectric constant of the polymer is lower than that of the analyte, the capacitance will be increased. Conversely, if the polymer dielectric constant is larger, the capacitance will be decreased (Fig. 9.8). These effects have been discussed and supported by simulations in [53], where the following formulae have been used to describe the change of the sensor capacitance.

eeff ¼ epolyð1 VFAcAÞþeA VFAcA ð9:3Þ

heff ¼ hð1 þ SAcAÞð9:4Þ

Here eeff denotes the resulting effective dielectric constant of the polymer/analyte system. epoly is the dielectric constant of the polymer, cA the concentration of the analyte in the gas phase, VFA the volume fraction of the analyte in the polymer per unit gas phase concentration, heff the resulting effective polymer thickness after analyte absorption and SA the experimental swelling coefficient of the polymer per unit gas phase concentration for the respective analyte. VFA and SA are constants (Henry’s law is assumed to be valid) and have to be determined experimentally for every polymer/ analyte combination by mass-sensitive or optical measurements. Typical sensor signals for a polymer layer (poly(etherurethane), PEUT), which is thicker (4.3 lm) than the surface-normal extension of the field lines, are shown in Fig. 9.7. The capacitor is alternately exposed to various concentrations of toluene and ethanol at 301 K and the pure carrier gas. The ratio of the dielectric constants of polymer (2.9) and analytes (toluene: 2.4, ethanol: 24.5) controls the signs of the signals. Ethanol with

Fig. 9.8 Frequency responses of a switched-capacitor device upon exposure to different analytes at 301 K. For a thick layer of PEUT (4.3 lm), toluene (dielectric constant lower than that of PEUT) causes positive frequency shifts, ethanol (dielectric constant higher than that of PEUT) leads to negative frequency shifts. Reprinted from [13] with permission 216 9 Hand-held and Palm-Top Chemical Microsensor Systems for Gas Analysis

a dielectric constant larger than that of PEUT causes positive frequency shifts, whereas toluene with a smaller dielectric constant causes negative frequency shifts. The limit of detection of the capacitive microsystem at 301 K is approximately 8 ppm for toluene and 5 ppm for ethanol [51]. For conducting measurements at defined temperatures, sensor and reference capacitors can be placed on thermally isolated membrane structures [50]. The fabrication of capacitors integrated with CMOS circuitry components is described in [50–53, 55, 56]. The main application of microstructured capacitors includes humidity sensing with polyimide films [45–48, 55, 56], since water has a high dielectric constant of 78.5 (liquid state) at 298 K and, therefore, generates large capacitance changes. Capacitive humidity sensors are commercially available from, for example, Sensirion (SUI), Vai- sala (FIN), and Humirel (F) [57–59].

9.3.3 Micromachined Resonating Cantilevers

Micromachined cantilevers commonly employed in atomic force microscopy (AFM) constitute a promising type of mass-sensitive transducer for chemical sensors [60–64]. The sensing principle is quite simple. The cantilever is a layered structure (Fig. 9.9) composed of the dielectric layers of a standard CMOS process, silicon, metallizations, and, eventually, zinc oxide. The cantilever base is firmly attached to the silicon support. The free-standing cantilever end is coated with a sensitive layer and is deflected depending on the added mass. The first step to achieve cantilever structures is the fabrication of membranes using anisotropic bulk etching with an electrochemical etch- stop technique (Fig. 9.10). The cantilevers are then released with additional front-side reactive-ion-etching steps [12, 13, 65]. After micromachining, the cantilevers are spray- coated with the chemically sensitive polymer film. There are two fundamentally different operation methods: (a) static mode: measurement of the cantilever deflection upon stress changes or mass loading by means of

Fig. 9.9 Schematic representation (a) and micrograph (b) of an integrated CMOS cantilever with on-chip circuitry. The cantilever is thermally actuated, its vibration is detected by piezoresistors. Reprinted from [13] with permission 9.3 Silicon-Based Microsensors 217

Fig. 9.10 Cantilever fabrication: (a) thinned CMOS wafer with Si-nitride layer on backside, (b) backside KOH wet etching (bulk micromachining) with electrochemical etch stop, and (c) frontside RIE to release the cantilever [67] laser light reflection [61, 63, 64]; (b) dynamic mode: excitation of the cantilever in its fundamental mode and measurement of the change in resonance frequency upon mass loading [12, 13, 62, 65] in analogy to other mass-sensitive devices. The two methods impose completely different constraints on the cantilever design for optimum sensitivity. Method (a) requires long and soft cantilevers to achieve large deflections, whereas method (b) requires short and stiff cantilevers to achieve high operation frequencies. Method (b) is preferable with regard to integration of electronics and simplicity of the setup (feedback loop) [12, 13, 62, 65, 66]. Method (a) can be applied in liquids as well [63], which is rather difficult using the dynamic mode. The excitation of the cantilever in the resonant mode is usually performed by applying piezoelectric materials (ZnO) [66] or by making use of the bimorph effect, which is the different temperature coefficients or mechanical stress coefficients of the various cantilever materials [12, 13, 60–65]. This difference in material properties gives rise to a cantilever deflection upon heating or applying mechanical forces. Periodic heating pulses in the cantilever base thus can be used to thermally excite the cantilever in its resonance mode at 10–500 kHz [12, 13, 65]. The cantilever vibration is detected by embedded piezoresistors in a Wheatstone-bridge configuration (Fig. 9.11). The cantilever exhibits a quality factor of approximately 1000 in air at a resonance frequency of 380 kHz [67], and acts as the frequency-determining element in an oscillation circuit (Fig. 9.11), which is entirely integrated on the chip. The oscillation frequency is measured with an on-chip counter. The first stage of the feedback circuitry includes a low- noise fully differential difference amplifier (DDA) with a gain of 30 dB, which amplifies the output signal of the Wheatstone bridge. The feedback additionally includes a high-pass filter followed by another amplifier, a limiter, a programmable digital delay line, and a driving stage. The delay line is used to adjust the phase to achieve positive

Fig. 9.11 Schematic of the cantilever feedback circuitry, which includes two cascaded amplifiers, a high-pass filter, a limiter, a programmable digital delay line, and a driving stage 218 9 Hand-held and Palm-Top Chemical Microsensor Systems for Gas Analysis

feedback at the fundamental resonance frequency. For more details on the electronics, see [65, 67]. In a first approximation, the change in the resonance frequency of the cantilever, f 3 (Hz), is proportional to the change in the analyte gas concentration, cA (mol/m ) (linear slope in Fig. 9.12):

f ¼ G MA K cA ð9:5Þ

Here, G (Hz/(kg/m3)) denotes the gravimetric sensitivity of the cantilever. G is cantilever-specific and depends on the geometric dimensions (thickness, length) and the material of the cantilever [67]. Here it is on the order of 10–20 Hz/(kg/m3). A frequency shift of 1 Hz hence corresponds to a change in the vibrating mass of approximately

5 pg. MA [kg/mol] is the analyte molecular weight and K the thermodynamic partition coefficient (Eq. 1), which is dimensionless and characteristic for the particular polymer-analyte combination. The mass resolution of the cantilevers is in the range of a few picograms [60–67]. This high mass sensitivity does not necessarily imply an exceptionally high sensitivity to analytes since the area coated with the sensitive layer usually is very small (on the order of 100 150 lm). The sensing layer is deformed upon motion of the cantilever, therefore, modulus effects are expected to contribute to the overall signal, especially since the coating thickness may exceed that of the cantilever. IC process-compatible fabrication sequences for monolithic integration of the cantilevers with electronics are detailed in [12, 13, 62, 65, 67]. Typical application areas are environmental monitoring such as the detection of different kinds of organic volatiles, such as hydrocarbons, chlorinated hydrocarbons, alcohols in the gas phase by using polymeric layers [60–67]. Figure 9.12 shows the measured frequency shift of a cantilever coated with 3.7 lm PEUT upon exposure to different concentrations of n-octane. The concentrations are ramped up and down to test for reproducibility [67]. The measurement chamber is purged with synthetic air

Fig. 9.12 Measured frequency shifts of a cantilever coated with 3.7 lm PEUT upon exposure to different concentrations of n- octane (250–1500 ppm). The mass-sensitivity is approximately 5 picograms/Hz [67] 9.3 Silicon-Based Microsensors 219 after each analyte exposure. A continuous directional sensor drift of 0.15 Hz/min has been subtracted from the results shown in Fig. 9.12.

9.3.4 Micromachined Calorimetric Sensors

This type of sensor relies on the thermoelectric or Seebeck-effect. If two different semiconductors or metals are connected at a hot junction and a temperature difference is maintained between this hot junction and a colder point, then an open circuit voltage is developed between the different leads at the cold point. This thermovoltage is proportional to the difference of the Galvani potentials (the Galvani potential is defined as the difference of the Fermi levels of the two materials) at the two temperatures and thus proportional to the temperature difference itself [68]. This effect can be used to develop a thermal sensor by placing the hot junction on a thermally isolated structure like a membrane, bridge etc. and the cold junction on the bulk chip with the thermally well-conducting silicon underneath [13, 69–71]. The membrane structure (hot junction) is covered with a sensitive or chemically active layer liberating or abstracting heat upon interaction with an analyte. The resulting temperature gradient between hot and cold junctions then generates a thermovoltage, which can be measured. Figure 9.13 displays the schematic of a CMOS thermopile. The overall sensor system includes two 700 lm by 1500 lm dielectric membranes with 300 polysilicon/aluminum thermocouples each (Seebeck coefficient: 111 lV/K) and an on-chip amplifier. One of the membranes is coated with a gas-sensitive polymer, the other one is passivated and serves as a reference [13, 71, 72]. The sensing and reference thermopiles are connected in parallel to the input stage of an on-chip amplifier for monitoring the

Fig. 9.13 Schematic of a thermoelectric sensor. Polysilicon/aluminum thermopiles (hot junctions on the membrane, cold junctions on the bulk chip) are used to record temperature variations upon analyte sorption in the polymer. The overall system includes a sensor membrane coated with gas-sensitive polymer, a passivated (Si-nitride) reference membrane and an on-chip amplifier [13, 71] 220 9 Hand-held and Palm-Top Chemical Microsensor Systems for Gas Analysis

Fig. 9.14 Schematic of the calorimeter circuitry. Sensing and reference thermopiles are connected in parallel to the input stage of a low- noise chopper-stabilized instrumentation amplifier followed by an anti- aliasing filter, a Sigma-Delta A/D converter and a decimation filter

temperature differences between the two membranes (Fig. 9.14). The low-noise chopper-stabilized instrumentation amplifier features a tunable gain of up to 8000 and a bandwidth of 500 Hz with an equivalent input noise of 15 nV/Hz1/2 [73]. The anti- aliasing filter prevents downsampling of high-frequency noise into the low-frequency signal band by the A/D converter. After Sigma-Delta A/D conversion and after passing a decimation filter (13 bit word length at 800/s), the data are read out or transferred to a digital interface [81]. Calorimetric detection includes four principal steps: (i) absorption and partitioning, or chemical reaction; (ii) generation of heat, which causes (iii) temperature changes to be transformed in (iv) thermovoltage changes (see examples [13, 70]). Each of the four steps contributes to the overall sensor signal. The thermovoltage change U (V) is pro-

portional to the derivative of the analyte concentration as a function of time dcA/dt (mol/m3s):

U ¼ A B Vpoly H K dcA=dt ð9:6Þ

Here A (K s/J) and B (V/K) are device- and coating-specific constants describing the translation of a generated molar absorption enthalpy H (J/mol) via a temperature chan-

ge into a thermovoltage change. Vpoly denotes the sensitive polymer volume, and K is the partition coefficient (Eq. 1) or reaction equilibrium constant. The calorimetric sensor only detects changes in the heat budget at nonequilibrium state (transients) upon changes in the analyte concentration. Thus, the sensors provide a signal upon absorption (condensation heat) and desorption (vaporization heat) of gaseous analytes into the polymer [70–72, 74, 75] or during chemical reaction of an analyte with the sensing material [75–79]. Processing sequences for the integration of thermoelectric sensors with circuitry in a CMOS standard process are detailed in [71, 79]. Sensors are commercially available from Xensor Integration (NL) [80]. Typical applications include the detection of different kinds of organic volatiles in the gas phase, for example, hydrocarbons, chlorinated hydrocarbons, and alcohols, by using polymeric layers [13, 70–75]. Figures 9.15 (a) and (b) show the output voltage of the microsystem while switching from synthetic air (nitrogen/oxygen mixture without humidity) to toluene (4000 ppm) and back to air at a temperature of 301 K [13, 71]. Enthalpy changes can be roughly approximated by integration over the peak area of the sensor signals [71]. The peak maximum and signal characteristics differ for 9.3 Silicon-Based Microsensors 221

Fig. 9.15 Thermovoltage of the microsystem while (a) switching from synthetic air (nitrogen/oxygen mixture without humidity) to toluene (4000 ppm) and (b) back to air at a temperature of 301 K [13]. Enthalpy changes can be roughly approximated by integration over the peak area of the sensor signals. Reprinted from [13] with permission. For details, see text

switching on and switching off the analyte. This is mainly due to the different time needed for generating the gas mixture (longer time) and just shutting down the analyte gas flow (short time). Optimizing the gas flow and considerable shortening of the gas paths are the most urgent improvements planned for the setup used here.

9.3.5 Single-Chip Multisensor System

The CMOS fabrication approach makes it possible to realize all three different transducers with their signal conditioning circuitry on the same chip. The orthogonality of the sensor signals is ensured by the fundamentally different transduction principles and is system-inherent. Additionally, several of such microsystem chips can be coated with different sensitive coatings and arranged in an array to further enhance the analyte characterization and discrimination performance. A schematic of the overall microsensor system architecture is shown in Fig. 9.16, a micrograph is displayed in Fig. 9.17. On the left-hand side of Fig. 9.16, are the four different sensors: capacitive, mass-sensitive, calorimetric and temperature [81]. The chip includes a temperature sensor in addition to the three different chemical sensor types since bulk physisorption of volatiles in polymers is strongly temperature-dependent: A temperature increase by 10 8C decreases the fraction of analyte molecules absorbed in the polymer by approximately 50 % and consequently leads to a drastic sensor signal reduction. 222 9 Hand-held and Palm-Top Chemical Microsensor Systems for Gas Analysis

Fig. 9.16 Schematic of the overall microsystem architecture comprising sensors, driving and signal conditioning circuitry (sensor front end), analog/digital converters, sensor control and power management unit, and a digital interface. Reprinted from [81] with permission

The sensor front end represents all the sensor-specific driving circuitry and signal- conditioning circuitry. The analog/digital conversion is done on chip as well, which achieves a unique signal-to-noise ratio since noisy connections are avoided and a robust digital signal is generated on chip and afterwards transmitted to an off-chip data port via an I2C interface [82]. The I2C bus interface offers the additional advantage of having only very few signal lines (essentially two) for bi-directional communication and also allows for operating multiple chips on the same bus system. An on-chip digital controller manages the sensor timing and the chip power budget [81]. Figure 9.17 shows the processed and tested single-chip gas sensor system [81]. The chip is 7 by 7 mm in size. The chip processing steps include an unaltered 15-mask commercial CMOS process [83] followed by applying a backside mask and anisotropic sodium hydroxide etching of the membrane structures for the calorimetric sensors and the cantilever. Front side reactive ion etching (RIE) is then executed to release the cantilever from the respective membrane. Finally, the sensors (cantilever, calorimeter, capacitor) have to be coated with the polymer using an airbrush method. The sensors are located in the center within a metal frame, which is used to apply a flip-chip packaging technique [84]. The signals of all three transducers correlate linearly with the analyte concentration in the low-concentration range (below 3 % of saturation vapour pressure at operating

Fig. 9.17 Micrograph of the single-chip gas microsensor system. Reprinted from [81] with permission 9.3 Silicon-Based Microsensors 223

Fig. 9.18 Six multi-sensor chips arranged on a ceramic board as sensing unit in a hand-held device. Each of the chips is coated with a different polymer

temperature) [85]. Each transducer provides different information on the target analytes. Alcohols, for example, provide comparably low signals on mass-sensitive transducers due to their high saturation vapor pressure and low molecular mass. On the other hand, alcohols exhibit a dielectric constant of 24.5 and provide large signals on capacitors. Drastic changes in thermovoltages on the thermopiles are measured upon exposure to chlorinated hydrocarbons (not shown here) used in cooling sprays, for example, which in turn have a low dielectric constant and provide only small signals on capacitors. Thus, comprehensive and complementary information is acquired with the multi-transducer system, and analyte characterization and identification are significantly improved. Six multi-sensor chips have been arranged on a ceramic board (Fig. 9.18) as sensing unit for a hand-held instrument. Each of the chips is coated with a different polymer. The unit thus provides a total of 18 different (6 capacitive, 6 mass-sensitive and 6 calorimetric) chemical sensor signals. The hand-held unit comprises in addition to the sensors a gas intake unit with pumps, valves and filters, a signal processing unit (microcontroller) and a power pack allowing for more than 24 hours of continuous operation. This is the first hand-held unit benefitting from microsystem technology components as sensor elements. The average power consumption of one of the multisensor chips is approximately 100 mW.

9.3.6 Operation Modes for CMOS Microsystems

Since hand-held systems are designed to by small and thus have severe constraints imposed on the use of calibration gases or filter units as well as on the overall power consumption, one has to come up with dedicated operation modes, which still enable reliable qualitative and quantitative measurements. One possibility includes the so- called “reverse mode of operation” (RMO) [86], which has been developed for the hand-held microsensor unit and adapted to the needs of the different transducing principles as will be described in the following. 224 9 Hand-held and Palm-Top Chemical Microsensor Systems for Gas Analysis

9.3.6.1 Reverse Mode of Operation (RMO) [86] In classical systems, the baseline is established by purging with pure carrier gas (typically nitrogen or synthetic air) or filtered ambient air. Purging is in most cases the basic state of the system and the sensors are exposed to the target analytes only for a short time. The sensor signal upon analyte exposure is then related to the carrier gas or ambient air signal, which serves as the “zero-signal”. Since neither carrying large reference gas cylinders nor using high-capacity filters is feasible with portable hand-held units, another solution had to be found. The new concept is based on the idea to invert operation conditions. Instead of purging as standard state and short gas exposure times, the system is now continuously exposed to analyte gas, and only short pulses of filtered or reference gas are used to re-establish the baseline. This operation mode induces higher demands on the sensor stability, since large drift deteriorates the sensor signal much more in using RMO than in using “classical” operation conditions. For the microsensor system, the reference baseline is established by 5 seconds of purging with filtered ambient air. This technique allows performing some hundred measurement cycles using a very small filter element. Figure 9.19 illustrates the operation states of valves and pumps in the RMO and the corresponding gas concentrations in the chamber. Figure 9.19 additionally shows the strategy and timing of the signal recording for the different transducers and the resulting sensor signals. Line 1 is indicating the valve status. “0” represents the basic state of the valve, when ambient analyte-loaded gas is directly transferred to the measurement chamber. In state “1”, the ambient gas passes a filter unit, and analyte molecules are removed from the gas stream: cleaned “reference” gas is flowing over the sensors. Line 2 represents the pump status. “0” means “pump off”, “1” means that the pump is operational. In line 3, the corresponding analyte gas phase concentrations are displayed. In the beginning of a measurement sequence the gas composition in the measurement chamber is not defined. The pump is then switched on for three seconds and analyte gas is pumped into the measurement chamber. The gas remains in the chamber for two seconds. Equilibrium signals of the capacitive and mass sensitive transducers are recorded, the measurement timing of which is displayed in line 4. The resulting sensor signals are schematically shown in line 5. The pump is then switched on again for three seconds, and the valve is set to make the analyte gas pass the filter. A concentration step between analyte-loaded and filtered gas is thus generated. Equilibrium state capacitive and mass-sensitive reference signals in filtered air are recorded over two seconds, after which the pump is switched on once more with the valve set to analyte gas. The last pump operation would not be necessary for the equilibrium-based sensors but it is necessary to get the second calorimetric transient as shown in line 7. As already described in Section 9.3.4, the calorimetric sensor relies on transients and provides signals exclusively upon concentration changes. Therefore, the calorimetric recording has to be performed at high time resolution (1 kHz) in two short intervals covering both flanks of the concentration signal (line 3), i.e., at the maximum gradient of the analyte concentration. The two transient signals of the calorimetric transducer (negative upon analyte desorption, positive upon analyte absorption) are displayed in line 7. 9.3 Silicon-Based Microsensors 225

Fig. 9.19 Reverse Mode of Operation (RMO) [86], as developed for a hand-held unit including micromachined multisensor chips: Operation states of valves and pumps and corresponding gas concentrations in the chamber (lines 1–3), and timing of the signal recording for the different transducers as well as resulting sensor signals (lines 4–7). For details see text 226 9 Hand-held and Palm-Top Chemical Microsensor Systems for Gas Analysis

9.4 Summary and Outlook

The field of hand-held sensor systems has been developing rather fast during the past five years. Most of the commercially available systems rely on polymers (pure or carbon-loaded) as sensitive layers, and are hence targeted at detecting organic volatiles at room temperature. Metal-oxide-based sensors operating at higher temperatures are not yet included into hand-held units due to their high energy consumption. The applications of hand-held instruments presently include, for example, quality and process control in industry, medical and environmental monitoring or hazardous material identification in industrial and military settings. The trend in commercial hand-held devices is for dedicated sensor solutions serving defined target applications. This trend may trigger the application-specific development of individual hand-held systems for certain key applications rather than following the “nose”-concept of using one universal hand-held system for a wealth of applications. Extensive customer support will be a crucial issue in this context. CMOS-based chemical microsensors with polymeric layers constitute a promising approach and offer a number of substantial advantages such as full microelectronics compatibility, extremely small size, low power consumption, and production at industrial standards. Further research and the fast progress of microelectronics development will help to significantly improve the system performance and reduce its size. Palm-size or even credit-card-size detection units based on CMOS technology are conceivable in the near future.

Acknowledgments The authors are greatly indebted to current and former staff of the Physical Electronics Laboratory at ETH Zurich and of the sensor laboratory at University of Tu¨bingen involved in the chemical microsensor development, in particular Christoph Hagleitner, Dirk Lange, Andreas Krauss, and Michael Frank. The authors like to express their gratitude to Dr. Heiko Ulmer, AppliedSensor GmbH, Reutlingen, Germany, for providing scientific material. This work has been financially supported by the Ko¨rber Foundation, Hamburg, Germany.

References

1 http://www.draeger.com 6 J. W. Gardner, P. N. Bartlett. Sens. Actuators 2 http://www.afcintl.com/badges.htm B, 1994, 18–19, 211–220. 3 R. A. Durst, R. W.Murray, K. Izutsu, 7 (a) S. Middelhoek, S. A. Audet. Silicon K. M. Kadish, L. R. Faulkner. Draft IUPAC Sensors, Academic Press Inc., London, Report Commission V.5. 1989; (b) Middelhoek, S., Sens, Actuators A, 4 A. Hulanicki, S. Glab, F. Ingman. Pure & 1994 41–42, 1–8. Appl. Chem., 1991, 63, 1247. 8 S. M. Sze, (ed.). Semiconductor Sensors, 5 http://www.draeger.com/com/ST/Prod/ Wiley, New York, 1994. Detection/Mobile_messtechnik/Gasmess- 9 G. T. A. Kovacs. Micromachined Transducers, geraete/PacIII/PacIII.jsp WCB McGraw-Hill, Boston, 1998. 9.4 Summary and Outlook 227

10 M. Elwenspoek, H. Hansen. Silicon Micro- 29 D. S. Ballantine, H.Wohltjen. Anal. Chem., machining, Cambridge University Press, 1989, 61, 705–712. Cambridge, 1998. 30 Table reprinted with permission of Applied 11 M. Madou. Fundamentals of Microfabrication, Sensor GmbH, Reutlingen, Germany. CRC Press, Boca Raton, FL, 1997. 31 M. C. Lonergan, E. J. Severin, B. J. Doleman, 12 H. Baltes, D. Lange, A. Koll. IEEE Spectrum, S. A. Beaber, R. H. Grubbs, N. S. Lewis. 1998, 9, 35–38. Chem. Mater., 1996, 8, 2298–2312. 13 A. Hierlemann, D. Lange, C. Hagleitner, 32 E. J. Severin, B. J. Doleman, N. S. Lewis. N. Kerness, A. Koll, O. Brand, H. Baltes. Anal. Chem., 2000, 72, 658–668. Sens. Actuators B, 2000, 70, 2–11. 33 B. J. Doleman, M. Lonergan, E. J. Severin, 14 W. P. Carey, K. R. Beebe. B. R. Kowalski. T. P. Vaid, N. S. Lewis. Anal. Chem., 1998, Anal. Chem. 1986, 58, 149–153. 70, 4177–4190. 15 D. S. Ballantine, S. L Rose, J. W. Grate, 34 D. Stauffer, A. Aharony. Introduction into H. Wohltjen. Anal. Chem. 1986 58, percolation theory, Taylor and Francis, Bristol, 3058–3066. PA, 1994. 16 A. Hierlemann, M. Schweizer-Berberich, 35 A. Van den Berg, P. D. van der Waal, U. Weimar, G. Kraus, A. Pfau, W. Go¨pel. B. B. van der Schoot, N. F. de Rooij. Sens. Pattern Recognition and Multicomponent Mat. 1994, 6, 23–43. Analysis, In: Sensors Update Vol. 2, Baltes , 36 G. Mu¨ller, P. P. Deimel, W. Hellmich, W. Go¨pel H., Hesse J., Eds.; VCH: Wein- C. Wagner. Thin Solid Films, 1997, 296, heim, Germany, 1996. 157–163. 17 D. L. Massart, B. G. M.Vandeginste, 37 J. V. Hatfield, P. I. Neaves, P. J. Hicks, S. N. Deming, Y. Michotte, L. Kaufman. K. Persaud, P. Traves. Sens. Actuators B, Data Handling in Science and Technology 2: 1994, 18–19, 221–228. Chemometrics: a Textbook, Amsterdam, 38 P. I. Neaves, J. V. Hatfield. Sens. Actuators Elsevier 1988. B, 1995, 26–27, 223–231. 18 R. G. Brereton, (ed.). Multivariate pattern 39 J. L. Rodriguez, R. C. Hughes, W. T. Corbett, recognition in chemometrics, Data handling P. J. McWhorter. Technical Digest IEEE in science and technology, Vol. 9, Elsevier, International Electron Devices Meeting Amsterdam, New York, 1992. New York, NY, USA, 1992, 521–524. 19 P. W. Atkins. Physical Chemistry, Oxford 40 M. J. Vellekoop, G. W. Lubking, P. M. Sarro, University Press, Oxford, 5th edition, 1994. A. Venema. Sens. Actuators A, 1994, 44, 20 K. Bodenho¨fer, A. Hierlemann, G. Noetzel, 249–263. U. Weimar, W. Go¨pel. Anal. Chem., 1996, 41 H. Baltes, O. Brand. IEEE AES Systems 68, 2210–2218. Magazine, 1999, 14, 29–34; and Proc. SPIE, 21 http://www.appliedsensor.com 1999, 3673, 2–10. 22 http://www.microsensorsystems.com/ 42 J. N. Zemel. Rev. Sci. Instrum. 1990, 61 (6), frameset_products.htm 1579–1606. 23 http://cyranosciences.com/products/ 43 W. Menz, J. Mohr, O. Paul. Microsystem cyranose.html Technology, Wiley-VCH, Weinheim, 24 J. W. Grate, G. C. Frye. In: Sensors Update; New York, 2001, ISBN 352-729-634-4. Vol. 2; Baltes, H.; Go¨pel, W.; Hesse, J., Eds.; 44 O. Brand, H. Baltes. Micromachined resonant VCH: Weinheim, FRG, 1996; pp. 37–83. sensors, In: Sensors Update, Vol. 4, Baltes, 25 J. W. Grate, S. J. Martin, R. M. White. H.; Go¨pel, W.; Hesse, J., (Eds.), VCH, Anal. Chem., 1993, 65, 940A–948A and Weinheim, Germany, 1999. 987A–996A. 45 N. F. Sheppard, D. R. Day, H. L. Lee, 26 D. S. Ballantine, R. M. White, S. J. Martin, S. D. Senturia. Sens. Actuators, 1982,2, A. J. Ricco, G. C. Frye, E. T. Zellers, 263–274. H. Wohltjen. Acoustic Wave Sensors: Theory, 46 M. C. Glenn, J. A. Schuetz. Technical Digest Design, and Physico-Chemical Application, Transducers 1985, 217–219. Academic Press, San Diego, 1997. 47 S. D. Senturia. Technical Digest Trans- 27 J. Janata. Principles of Chemical Sensors, ducers 1985, 198–201. Plenum, New York, 1989. 28 G. Z. Sauerbrey. Phys., 1959, 155, 206–222. 228 9 Hand-held and Palm-Top Chemical Microsensor Systems for Gas Analysis

48 G. Delapierre, H. Grange, B. Chambaz, 65 (a) D. Lange, C. Hagleitner, O. Brand, L. Destannes. Sens. Actuators, 1983,4, H. Baltes. Proc. Transducers ’99, Sendai, 97–104. Japan, 1999 1020–1023, (b) C. Hagleitner, 49 H. Shibata, M. Ito, M. Asakursa, D. Lange, O. Brand, A. Hierlemann, K. Watanabe. IEEE Transactions on H. Baltes. Digest of Technical Papers, Instrumentation and Measurement, 1996, IEEE International Solid-State Circuits 45, 564–569. Conference (ISSCC), San Francisco, USA, 50 (a) C. Hagleitner, A. Koll, R. Vogt, O. Brand, 2001, 246–247, ISBN 0-7803-6608-5. H. Baltes. Proc. Transducers ’99, Sendai, 66 S. S. Lee, R. M. White. Sens. Actuators A, Japan, 1999, 1012–1015, (b) S. Kawahito, 1996, 52, 41–45. A. Koll, C. Hagleitner, H. Baltes, 67 D. Lange. D. Dissertation ETH Zurich, 2000. Y. Tadokoro. Trans. IEE of Japan, 1999, 68 A. W. Van Herwaarden, P. M. Sarro. Sens. 119–E, 3, 138–142. Actuators, &hf,1986, 10, 321–346. 51 A. Koll, A. Kummer, O. Brand, H. Baltes. 69 P. M. Sarro, A. W. van Herwaarden, Proc. of SPIE Smart Structures and W. van der Vlist. Sens. Actuators A, 1994, Materials, Vol. 3673, Newport 1999. 41–42, 666–671. 52 C. Cornila, A. Hierlemann, R. Lenggen- 70 A. W. Van Herwaarden, P. M. Sarro, hager, P. Malcovati, H. Baltes, G. Noetzel, J. W. Gardner, P. Bataillard. Sens. U. Weimar, W. Go¨pel. Sens. Actuators B, Actuators A, 1994, 43, 24–30. 1995, 24–25, 357–361. 71 A. Koll, A. Schaufelbu¨hl, O. Brand, 53 F. P. Steiner, A .Hierlemann, C. Cornila, H. Baltes, C. Menolfi, H. Huang. Proc. G. Noetzel, M. Ba¨chtold, J. G. Korvink, IEEE Workshop on Micro Electro W. Go¨pel, H. Baltes. Technical Digest Mechanical Systems MEMS 99, Orlando, Transducers 1995, Vol. 2, 814–817. 1999 547–551, ISBN 0-7803-5194-0. 54 P. Van Gerwen, W. Laureys, G. Huybe- 72 N. Kerness, A. Koll, A.Schhaufelbuehl, rechts, M. Op De Beeck, K. Baert, J. Suls, C. Hagleitner, A. Hierlemann, O. Brand, A.Varlan, W. Sansen, L. Hermans, H. Baltes. Proc. IEEE Workshop on R. Mertens. Technical Digest Transducers Micro Electro Mechanical Systems 1997, Vol. 2, 907–910. MEMS 2000, Myazaki, Japan, 2000, 96–101, 55 T. Boltshauser, H. Baltes. Sens. Actuators A, ISBN 0-7803-5273-4. 1991, 25–27, 509–512. 73 C. Menolfi, Q. Huang. IEEE Journal Solid 56 T. Boltshauser, L. Chandran, H. Baltes, State Circuits, 1997, 32, 1–9. F. Bose, D. Steiner. Sens. Actuators B, 1991, 74 J. Lerchner, J. Seidel, G. Wolf, E. Weber. 5, 161–164. Sens. Actuators B, 1996, 32, 71–75 57 http://www.sensirion.com 75 D. Caspary, M. Schro¨pfer, J. Lerchner, 58 http://www.vaisala.com G. Wolf. Thermochimica Acta 1999, 337, 59 http://www.humirel.com 19–26. 60 T. Thundat, G. Y. Chen, R. J. Warmack, 76 J. Lerchner, A. Wolf, G. Wolf.. J. Thermal. D. P. Allison, E. A. Wachter. Anal. Chem. Anal., 1999, 55, 212–223. 1995, 67, 519–521. 77 P. Bataillard, E. Steffgen, S. Haemmerli, 61 M. Maute, S. Raible, F. E. Prins, D. P. Kern, A. Manz, H. M. Widmer. Biosensors, H. Ulmer, U. Weimar, W. Go¨pel. Sens. Bioelectronics, 1993, 8, 89–98. Actuators B 1999, 58, 505–511. 78 J. M. Ko¨hler, E. Kessler, G. Steinhage, 62 H. Jesenius, J. Thaysen, A. A. Rasmussen, B. Gru¨ndig, K. Cammann. Mikrochim. L. H. Veje, O. Hansen, A. Boisen. Appl. Acta 1995 120, 309–319. Phys. Letters, 2000, 76, 2615–2617. 79 P. M. Sarro, H. Yashiro, A. M. van Her- 63 J. Fritz, M. K. Baller, H. P. Lang, waarden, S. Middelhoek. Sens. Actuators, H. Rothuizen, P. Vettiger, E. Meyer, 1988 14, 191–201. H. J. Gu¨ntherodt, C. Gerber, 80 A. W. Van Herwaarden. Meas. Sci. Technol., K. J. Gimzewski. Science, 2000, 288, 1992, 3, 935–937. 316–318. 81 C. Hagleitner, A. Hierlemann D. Lange, 64 R. Berger, E. Delamarche, H. P. Lang, A. Kummer, N. Kerness, O. Brand, C. Gerber, J. K. Gimzewski, E. Meyer, H. Baltes. Nature, 2001, 414, 293–296. H. J. Gu¨ntherodt. Science, 1997, 276, 2021. 9.4 Summary and Outlook 229

82 I2C is a communication standard developed 85 A. Hierlemann, A. J. Ricco, K. Bodenho¨fer, by Philips, Eindhoven, The Netherlands. A. Dominik, W. Go¨pel. Anal Chem. 2000, 83 CMOS process as provided by Austria Mikro 72, 3696–3708. Systeme International, Unterpremsta¨tten, 86 A. Krauss, D.Lange, U. Weimar, Austria. A. Hierlemann, O. Brand, H. Baltes. 84 A. Koll, S. Kawahito, F. Mayer, Proc. InternationaI Symposium on C. Hagleitner, D. Scheiwiller, O. Brand, Electronic Noses, ISOEN, Brighton, 2000, H. Baltes. Proc. of SPIE Smart Structures 35–36. 40 and Materials 3328, 1998, 223–232. 231

10 Integrated Electronic Noses and Microsystems for Chemical Analysis

Julian W. Gardner, Marina Cole

Abstract There is considerable interest in the miniaturization and mass production of electronic noses through the exploitation of recent advances in the emergent field of microsystems technology. In this chapter we explore the future outlook for integrated electronic noses by first reviewing the different types of microfluidic components that have been reported in the literature, such as microchannels, microvalves, and micropumps. Next, we describe recent efforts to develop microelectronic noses based upon the integration of sensor arrays and smart interfaces. Finally we report upon work in the related field of micro total analysis systems, in which, for example, a micro gas chromatograph or a micro mass spectrometer are being fabricated; these physically-based microinstruments may be regarded as a type of micronose and thus in competition with integrated electronic micronoses.

10.1 Introduction

The integration of gas microsensors and signal processing circuitry is a subject of ever- increasing importance in the chemical sensor community. It offers lower unit cost through batch production of wafers, smaller device size, better reproducibility, superior signal conditioning by less noise generated in the transmission of the sensor signals to the processing electronics, and an improved limit of detection for the whole sensing system. The full integration of gas microsensors and signal processing circuitry has been brought a step nearer with reports of an increase in sensor reproducibility by the integration of arrays of sensors onto the same substrate [1–4], improvements in sensor sensitivity through advances in individual microsensor technologies [5, 6] and finally the development of novel gas-sensitive materials, for example see Gardner and Bartlett [7], and Attard et al. [8]. Many commercial (non-chemical) sensors have been realized in recent years through the integration of both the electronic signal processing circuitry and the sensing part on the same silicon die. Some examples include pressure sensors, ultrasound sensors and gas flow sensors, proximity and temperature sensors

[9–12]. For a review of silicon sensors, readers are referred to Gardner et al. [13]. The field of gas sensors is covered in Chapter 9 of this book where a detailed description of the design of the single-chip multisensor system comprising four different sensors, as well as driving and signal-conditioning circuitry, can be found. Recent years have also seen substantial efforts in the development of smart and intelligent sensor technology. The main advantages of intelligent sensors are their improved performance and reliability – achieved through the addition of self-testing and self-diagnostic functions [14]. Emphasis has also been given to the development of application-specific integrated circuits (ASICs) for intelligent sensors. Taner and Brignell [15] have studied the advantages of ASIC technology, which enables intelligent devices to deal both with systematic variation in sensor parameters and provides good solutions for sensor communications. Parallel with the integration of microsensors and signal processing electronics, and the realization of smart sensor interfaces, sampling and fluid-handling techniques have been rapidly developing. Micro flow sensors, micropumps and microvalves started emerging in the late 1980s marking the beginning of the field of microfluidics. So far, life sciences and chemistry have been the main application areas of microfluidics in the liquid phase. Considering that sample handling is a critical area, which has an enormous influence on the performance of e-noses (see Chap- ter 3), microfluidics should have a significant impact on the future development of superior, integrated electronic nose (e-nose) systems. Microfluidic technology combined with smart silicon sensor arrays could lead to the development of cheap, and possibly disposable devices, particularly important in medical applications such as chemical and biological assays. In this chapter, a review of different solid-state sensor systems and smart sensor interfaces for e-noses will be given together with an overview of existing microfabricated components for fluid handling, such as microvalves, micropumps, and micro-

Fig. 10.1 Basic components that make up an integrated e-nose or chemical analysis system 10.2 Microcomponents for Fluid Handling 233 channels. A block diagram of the basic components that make up an integrated e-nose system is shown in Fig. 10.1 and will be described below. The integration of different components into microsystems and microinstruments will also be discussed, and the future outlook concerning sensor arrays, biological assay devices, and neuromorphic systems will be briefly outlined. Microsystems for chemical analysis based on gas chromatography, mass spectrometry, and optical spectrometry techniques will also be reviewed. Finally, a future outlook is given of e-noses and microsystems for chemical analysis.

10.2 Microcomponents for Fluid Handling

In the early 1990s, microfluidics was established as a general term for the research discipline dealing with fluid transport phenomena on the micrometer scale and fluidic components, devices, and systems built with microfabrication technologies. The major applications of microfluidics are in the fields of medical diagnostics, genetic se- quencing, drug discovery, and proteomics. This section focuses on microcomponents for fluid handling, such as microchannels, microchambers, microvalves and micropumps that could be applicable to the development of integrated e-noses and microsystems for chemical rather than biological analysis. Advancements in photolithography turned the possibility of miniaturizing analytical systems into reality. Initially, only simple channels and reservoirs could be made by photolithography on glass or silicon wafers, and electro-osmosis was the only way to move liquids. Over the last 10 years, the fabrication of new microfluidic components, such as valves, pressure systems, metering systems, reaction chambers, and detection systems, has led towards the development of more complex manufacturing technologies, e.g. deep reactive ion etching (DRIE), and multiplayer processes such as the five- layer polysilicon Sandia process) and hence the possibility for lab-on-a-chip prototypes [16]. Apart from their use in research, microfluidic devices also have significant commercial potential. In 1999, the Systems Planning Corporation, Arlington, VA, released a market study on microelectromechanical systems (MEMS) that projected a microfluidics market of 3 to 4.5 billion euros by 2003. 30 % of this total is split equally between sensors and lab-on-a-chip applications. Another microsystems market study completed in 1996 by a task force of the European Commission’s Network of Excel- lence in Multi-functional Microsystems (NEXUS) forecasted a market of Q 2.8 billion for microstructure-based disposable assay devices alone by 2002.

10.2.1 Microchannels and Mixing Chambers

Microchannels are essential components of microfluidic systems. They provide the connections between pumps, valves and sensors [17], and they are used as separation columns for different types of gas or liquid chromatographs [18, 19]. They also act as 234 10 Integrated Electronic Noses and Microsystems for Chemical Analysis

Fig. 10.2 Scanning electron microscope (SEM) images of several types of microchannels, fabricated with: (a) bulk micromachining and wafer bonding, (b) surface micromachining, and (c) buried channel technology. From Boer et al. [21] 10.2 Microcomponents for Fluid Handling 235 heat exchangers in the cooling of electronic chips [20]. Common methods used for the fabrication of microchannels include bulk etching with wafer-to-wafer bonding, bulk etching with sealing using low-pressure chemical vapour deposition (LPCVD) materials, conventional surface micromachining of channels, imprinting plastic substrates, X-ray LIGA (German acronym: Lithographie, Galvanoformung, Abformung) technology and DRIE, and channel forming from UV photodefinable SU-8 photoresist using LIGA (sometimes called UV LIGA). Figure 10.2 shows scanning electron micrograph (SEM) images of several types of microchannels fabricated with bulk micromachining and wafer bonding (Fig. 10.2a), surface micromachining (Fig. 10.2b), and buried channel technology (BCT) (Fig. 10.2c). One problem with wafer-to-wafer techniques, such as anodic bonding or direct fusion bonding, is the possible creation of wafer-to-wafer misalignments and the formation of microvoids at the bonding interface, which may affect the functional performance of the device. Another difficulty is that electronic circuitry (e.g. CMOS) cannot be incorporated on the same substrate because of the high process temperatures and voltages needed to perform anodic bonds. The use of surface micromachining obviates the need for accurate wafer alignment. In this approach, structural parts are embedded in layers of a suitable sacrificial material on the surface of a substrate. Dissolving the sacrificial material forms a complete microchannel. By this method microchannels can be fabricated in various different passive materials, e.g. silicon nitride, polysilicon, metal, polymer, and silicon dioxide. One major disadvantage of this technique has been that the vertical dimension of such channels is restricted by the maximum sacrificial layer thickness that can be deposited and etched within an acceptable time period. Researchers at the University of Twente have proposed BCT as an alternative to conventional bulk and surface micromachining [21]. BCT allows the fabrication of complete microchannels in a single wafer with only one lithographic mask, and processing on one side of the wafer. The microstructures are constructed by trench etching, coating of the sidewalls of the trench, removal of the coating at the bottom of the trench, and finally etching into the bulk of the silicon substrate. This method for the fabrication of these devices was derived from the SCREAM (single-crystal reactive etching and metallization) process [22]. The structure can be sealed by the deposition of a suitable layer that closes the trench. Using the above procedure it is possible to construct cavities, reaction chambers or crossing channels. A spiral-shaped channel with a length of 10 m and a diameter of 30 lm was also developed by the same research group for possible application as a separation element in gas chromatography. The method of imprinting plastic substrates involves the low-temperature pattern- ing of plastic substrates using either small diameter wire or a micromachined silicon template. Silicon templates have also been used as a negative master tool for fabrication of polymer microchannels and mixers by hot embossing and microinjection molding [23]. Templates have also been made in metals and, more recently, dia- mond. Figure 10.3 shows a micrograph of a mixer, which has been replicated on a hot embossed polymethylmethacrylate wafer. The advantage of this method is that the resultant channels are robust, easy to fabricate at low cost, and compatible with biological fluids (unlike silicon); it also allows the integration of other microfluidic elements and the sensors but not electronic circuitry. Another technology that 236 10 Integrated Electronic Noses and Microsystems for Chemical Analysis

Fig. 10.3 Micrograph of an embossed mixer (smallest channel dimensions: 50 lm 50 lm). From Greschke et al. [22]

Fig. 10.4 Process steps for fabricating the MWlCs (a) etch channel, (b) form porous silicon (PS), (c) under-etch PS. From Tjerkstra et al. [24] 10.2 Microcomponents for Fluid Handling 237

Fig. 10.5 SEM photograph of a MWlC containing two porous layers. From Tjerkstra et al. [24]

could be used is LIGA technology. The principal advantage of the LIGA process is that microdevices can be fabricated with a height-to-width aspect ratio of up to 200, typically several millimeters in height and 10 lm in width, but the microchannels fabricated using this technology are usually made from plastics, metals, and ceramics rather than silicon. The process is also very expensive because it requires a synchrotron source. Finally, channels can be made using thick UV LIGA. Using this method only the sidewalls of the channel can be formed, so to seal them some method of bonding is required. The advantage is that the thickness of the wall can be easily controlled and a high aspect ratio achieved. Releasing electrochemically formed porous silicon from the bulk silicon substrate by under-etching at increased current density is another technique that can be used for the microfabrication of microchannels, in particular multi-walled microchannels (MWlCs) [24]. Figure 10.4 shows the main processing steps for the formation of a MWlC by etching channels in a p-type silicon wafer using LPCVD silicon-rich silicon nitride as the mask material. Figure 10.5 shows an SEM image of a MWlC containing two porous layers. To create more robust devices, i.e. to increase the strength of the structure, microchannels can be fabricated with a porous silicon membrane suspended halfway across an etched cavity surrounded by silicon nitride walls. Most of the above methods allow for some integration with sensors, but external integration with the electronic circuitry is typically used, e.g. hybrid packaging or multi-chip modules. In order to allow direct integration of sensors, actuators, and other electronics with the microchannels, Rasmussen et al. have proposed two methods for the fabrication of microchannels using the standard CMOS process and simple and inexpensive post-processing steps [25]. In the first method, shallow microchannels of the order of 0.4 lm are realized by removing surface layers incorporated in a standard CMOS integrated circuit process. Larger channels with depth of 30 to 300 lm can be fabricated through the second method that employs bulk micromachining techniques. Both methods offer the possibility to create a complete smart microfluidic sys- 238 10 Integrated Electronic Noses and Microsystems for Chemical Analysis

Fig. 10.6 Microscope photographs of a fabricated chip for surface-micromachining shallow channels before etch. From Rasmussen et al. [25]

tem that possesses integrated microfluidic elements, sensors and electronics as shown in Fig. 10.6.

10.2.2 Microvalves

Microvalves are one of the most important building blocks of microfluidic systems used for fluid flow control. They can be classified in two categories: active valves (with an actuator) and passive check valves (without an actuator).

10.2.2.1 Active Microvalves An active microvalve consists of a device body that contains the fluid under pressure, a valve seat to modify the fluid flow, and an actuator to control the position of the valve seat. The first reported microvalve was designed as an injection valve for use in integrated gas chromatography [26]. It had a silicon valve seat and a nickel diaphragm actuated by an external solenoid. Following this first design, a large number of microvalves have been designed and reported, and they can be classified on the basis of the actuation method employed. These methods include pneumatic, thermopneumatic, thermomechanic, piezoelectric, electrostatic, electromagnetic, electrochemical, and capillary force microvalves. Figure 10.7 shows some of these actuating devices. Pneumatic valves have a membrane structure as the valve seat. Although pneumatic actuation is based upon a very simple principle it requires an external pressure source, which makes the pneumatic valves unsuitable for most compact applications. A low spring constant is also an important parameter and in order to achieve it thin membranes or corrugated membranes have to be designed. Soft elastic materials, such as 10.2 Microcomponents for Fluid Handling 239

Fig. 10.7 Schematics and principles of operation of (a) bimetallic (from Jerman [31]), (b) electrostatically actuated (from Shoji and Esashi [41]), and (c) electromagnetic active microvalves (from Yanagisawa et al. [37]) silicon rubber or Parylene, can be used to realize the low spring constant [27],whereas hard materials such as silicon and glass are problematic. Thermopneumatic valves utilize a sealed pressure cavity filled with a liquid. Actuation is based on the change in volume of the sealed liquid. The phase change from liquid to gas or from solid to liquid can also be used if a larger volume expansion is required. The disadvantage of these types of valves is the incompatibility of the technology, since the liquid has to be primed, filled, and sealed individually. Thin films of solid paraffin material could be used as an alternative [28] since they could be integrated in the batch fabrication. Thermomechanic microvalves utilize the principle of the conversion of thermal energy directly into mechanical stress. There are three types of thermomechanic actuators: solid-expansion [29], shape-memory alloy [30], and bimetallic actuators. A bimetallic valve as shown in Fig. 10.7a, was introduced by Jerman [31] and was one of the first commercialized microfluidic components. The valve was designed for a gas flow controller, and its actuator consists of a central boss, a circular diaphragm made of bimetallic materials, and a circular heater. The temperature, controlled by the heater, of the bimetallic structure on the diaphragm was used to vary the force applied to the boss by the diaphragm so that the gas flow can be adjusted. The valve was designed to control the gas volumetric flow-rate from 0 to 90 mL min1. Piezoelectric microvalves utilize the piezoelectric effect as an actuation principle. The thin-film piezoelectric actuators do not deliver sufficient force/displacement for valve 240 10 Integrated Electronic Noses and Microsystems for Chemical Analysis

application, so all reported piezoelectric microvalves use external stack-type piezoelectric actuators, bimorph piezocantilevers, or bimorph piezodisks. Bimorph actuators are used if larger deflections are required. They either consist of two bonded piezoelectric layers, with anti-parallel polarization [32], or of a piezoelectric layer bonded to a non-piezoelectric layer [33]. Electrostatically actuated microvalves have typically been designed with a valve seat in the form of a cantilever. Figure 10.7b shows a cantilever structure fabricated by surface micromachining. The valve seat could also have a form of flexible metal bridge. Such a design of microvalve was fabricated by Shikida and Sato [34] in order to allow an adequate gas flow-rate under low pressure. The valve is suitable for rarefied gas control systems but it requires a large actuating voltage of 100 V. In order for a large deflection to be achieved with a relatively small drive voltage, a combination of the electrostatic force, buckling effect, and pneumatic force were used. This arrangement was utilized for a bistable valve in an implantable drug delivery system [35]. Vandelli et al. have designed a microvalve array for fluid flow control with the valve membrane and one electrode made out of a polysilicon layer, whereas the other electrode was bulk silicon fabricated using surface micromachining. The valves were arranged in an array so that the flow rate could be controlled digitally [36]. Electromagnetic microvalve with a valve cap made out of a soft magnetic Ni-Fe alloy supported by a spring is shown in Fig. 10.7c [37]. It moves vertically in the magnetic field gradient applied by the external electromagnet. This valve was designed as a flow regulator for a high vacuum application. A microvalve activated by a combination of electromagnetic and electrostatic forces has also been fabricated [38]. The structure consists of a gas flow inlet having a counter electrode, a deflectable membrane coated with a metal conductor and two permanent magnets. Current pulses of 200 mA and a voltage of 30 V were typically applied for electromagnetic and electrostatic actuation. The response time was below 0.4 ms. There are also some reported examples of so-called electrochemical and capillary- force valves. Electrochemical valves are actuated gas bubbles generated by the electrolysis reaction, where the pressure inside the bubble is proportional to the surface tension and the radius of curvature. Capillary-force valves use capillary-force actuation where the surface tension and the capillary force can be controlled actively or passively by different means: electrocapillary, thermocapillary, and passive capillary. The use of some of these effects are described in two published papers [39, 40].

10.2.2.2 Passive Microvalves (Check Valves) This type of microvalve is typically designed for use in micropumps where a very small leakage under reverse applied pressure and a large reverse-to-forward flow resistance ratio is required. The dimensions of check valves are small in comparison with the valves with integrated or external actuators. A review paper by Shoji and Esashi describes a wide variety of check valve structures [41]. A typical cantilever-type structure is shown in Fig. 10.8. Oosterbroek et al. have reported on the design, simulation, and realization of in-plane operating passive microvalves [42]. Figure 10.9 shows functional, art, and SEM impressions of a 2 5 high-density in-plane check valve array. 10.2 Microcomponents for Fluid Handling 241

Fig. 10.8 Schematic of cantilever-type passive (check-valve) microvalve. From Shoji and Esashi [41]

Fig. 10.9 (a) Functional (b) art, and (c) SEM impressions of a 2 5 high-density in-plane check valve array. From Osterbroek et al. [42]

10.2.3 Micropumps

Micropumps can be classified into two categories: mechanical pumps with moving parts and non-mechanical without moving parts. Mechanical pumps can be further divided into three major groups: reciprocating, peristaltic, and valve-less rectification pumps. Non-mechanical pumps are mainly used to move liquids and constituents in liquids, which need to avoid any moving mechanical parts. Electrohydrodynamic, electro-osmotic, or ultrasonic effects are normally employed in the operation of non-mechanical micropumps.

10.2.3.1 Mechanical Micropumps A reciprocating-type micropump consists of a pump pressure chamber with a flexible membrane driven by an actuator unit and passive microvalves (check valves). Two main conditions have to be satisfied in order for reciprocating micropumps to function correctly. First, the minimum compression ratio of micropumps (the ratio between the stroke volume and the dead volume that causes serious constraints in most micropump designs) has to be determined and secondly the pump pressure has to be high. For the gas micropumps that are of interest in integrated e-nose ap- 242 10 Integrated Electronic Noses and Microsystems for Chemical Analysis

plications, the criterion for the minimum compression ratio e is given by Richter et al. [43]: p 1=c e > 0 1 p0 jDpcritj

where p0 is atmospheric pressure, Dpcrit is the critical pressure required for the opening of the check valve, and c is the adiabatic coefficient (for air c takes a value of 1.4). (The above equation could be simplified at small critical pressures and low pump frequencies where there is isothermal behaviour.) The maximum output pressure of the micropump depends directly on the available force of the actuator used. The different types of actuators that have been employed to date are piezoelectric, pneumatic, electrostatic, and thermopneumatic. The first piezoelectric micropump (Fig. 10.10), developed and reported in 1988 by van Lintel et al. [44], was made in silicon and used a piezoelectric disk. Since then, various different types of micropumps have been designed in order to satisfy the two requirements mentioned above. Over the years the dead volume of the pump chamber has became smaller, and the check valves and the pump membranes have been made out of softer materials with low spring constants. An example of a self-priming micropump [45] with a small dead volume able to pump gas is shown in Fig. 10.11. Here the dead volume was minimised by minimising the dead volume of the valve unit down to 500 nL, hence increasing the compression ratio to 0.111. In another development [46] van Lintel’s original design was improved to achieve a compression ratio of 1.16. Another way to achieve improvements in the micropump design is through the design of flexible check-valves. The valves, in the forms of cantilever, have been used with integrated electrostatic [47] or bimorph piezoelectric disks [48] as actuators. Materials that are more flexible than silicon, such as polyimide, polyester, and parylene, have also been used in the design of flexible check-valves. A thermo-pneumatically driven micropump was fabricated using the LIGA process [49]. In this the pump case is made by injection moulding of polysulfone (PSU) and the pump chamber is covered by a polyimide membrane. A similar design was fabricated by plastic injection moulding and uses a polyester valve [50] whereas the design reported in [51] has a pump membrane made out of silicone rubber, and the disk valve from parylene

Fig. 10.10 Piezoelectric disk reciprocating micropump. From van Lintel et al. [44] 10.2 Microcomponents for Fluid Handling 243

Fig. 10.11 Check-valve self-priming micropump. From Linneman et al. [45]

deposited through CVD process. Maximum flow rates reported for reciprocating type micropumps range from 4 to 13 000 lL min1. Functionality of peristaltic micropumps is based on the peristaltic motion of the pump chambers, and theoretically, peristaltic pumps should have three or more pump chambers with reciprocating membranes. These types of pumps do not require passive valves for the flow rectification, nor do they require a high chamber pressure, so the two main conditions that have to be satisfied are a large stroke volume and a large compression ratio. Three types of actuation principles have been employed in reported peristaltic micropumps: piezoelectric, electrostatic and thermopneumatic along with several types of fabrication processes, such as bulk micromachining, surface micromachining, plastic moulding, and a combination of bulk micromachining and anodic bonding. Maximum reported flow rates range from 3 to 30 000 lL min1 (for air). Examples include a surface micromachined pump with electrostatic actuators [52], a thermopneumatically driven micropump having three active pressure chambers with flexible membranes [53], a micropump with curved pump chambers and a flexible plastic membrane with electrostatic actuation [54]. A new pumping principle called dual-diaphragm pump, which consists of two actuating membranes in the pump chamber, is reported by Cabuz et al. [55]. This type of pump is able to pump up to the maximum reported flow-rate 30 mL min1 of air. Valveless rectification micropumps are similar to check-valve pumps except that, instead of check-valves, diffusers/nozzles are used for the flow rectification. In order to optimize the valveless pump designs, the stroke volume has to be maximized while the dead volume has to be minimized. The first piezoelectric micropump using nozzle/diffuser elements instead of check-valves [56] was presented in 1993. The original valve was fabricated in brass using precision machining. The same research group in 1997 presented the first valveless diffuser pump [57] fabricated using DRIE fabrication process shown in Fig. 10.12. The maximum pump pressure was 74 kPa and the maximum pump volumetric flow-rate was 2.3 mL min1 for water. Two different thermoplastic replication methods for the fabrication of valveless pumps: hot embossing and injection moulding have also been tested [58]. Deep precision-milled brass mould inserts and deep microelectroformed nickel mould inserts defined from DRIE silicon wafers were used for these designs. Figure 10.13 shows the design of the precision- 244 10 Integrated Electronic Noses and Microsystems for Chemical Analysis

Fig. 10.12 Fabrication process of the DRIE diffuser pump. From Olsson et al. [57]

milled brass mould insert pump. Tested pumps had a maximum volumetric flow-rate of 1.2 mL min1. Nguyen and Huang [59] have demonstrated the design of miniature valveless pumps based on a printed circuit board technique. The pump could be operated as a single diffuser/nozzle pump (or a peristaltic pump) and has a maximum flow-rate of 3 mL min1.

Fig. 10.13 (a) Top layout view of single-chamber diffuser pump, and cross-sectional views of (b) a single-chamber pump unit, and (c) two pump units stacked in parallel arrangement for high pump flow. From Olsson et al. [58] 10.3 Integrated E-Nose Systems 245

10.2.3.2 Nonmechanical Micropumps Nonmechanical pumps are based on non-mechanical pumping principles and can driven by capillary-force, thermal, chemical, electrical or magnetic means. Micro- pumps using electrohydrodynamic and electrokinetic effects have been reported. Electrohydrodynamic (EHD) actuation is based on electrostatic forces acting on dielectric fluids, such as organic solutions. Two main types of electrohydrodynamic micropumps have been reported: the EHD induction [60] and the EHD injection pumps [61]. Electrokinetic micropumps use the electrical field for pumping conductive fluid. They can be divided into two categories: pumps based on the electrophoresis effect and pumps based on the electro-osmosis effect. Electrophoresis can be described as the motion of charged particles under an electric field in a fluid relative to the uncharged fluid molecules. The velocity of the charged species is proportional to the applied electrical field. Electrophoresis pumps have their application in processes such as the separation of large molecules such as DNA or proteins [62]. The separation, performed in microchannels, is called capillary electrophoresis. Electro-osmosis is the effect of pumping fluid in a channel under an applied electric field. Changing the applied electric field or the pH of the solution that affects the potential arising from the charge on the channel wall can control the electro-osmotic flow velocity. In microanalysis systems, the electro-osmosis effect is used for delivering buffer solutions [63] and, when combined with the electrophoretic effect, for separating out different molecules. The drawbacks of electro-osmosis are that it cannot be used when several interconnected channels are required for sample processing, and it is not compatible with high-ionic-strength buffers. Information on various other types of micropumps, such as surface-tension driven pumps, magnetohydrodynamic pumps, and ferrofluidic magnetic pumps can also be found in the literature.

10.3 Integrated E-Nose Systems

10.3.1 Monotype Sensor Arrays

The performance of integrated e-nose systems largely depends upon the performance of the sensor array used. The integration of the sensor array on to the same substrate offers a reduction in sensor variation and also improves device reliability. Other advantages include reduced fabrication cost, smaller dimensions and lower power consumption. The majority of sensor arrays reported to date are monotype and can be divided into several categories based upon either the sensor material or type employed, such as conducting polymer, tin oxide, quartz resonator, surface acoustic wave (SAW) and FET sensor arrays. Readers are referred to Chapter 4 for a full description of the different types of chemical sensors. 246 10 Integrated Electronic Noses and Microsystems for Chemical Analysis

Conducting polymers are very attractive for gas-sensing applications because of their ease of deposition, the large variety of available polymer combinations, and their ability to operate at room temperature [64, 65]. Neaves and Hatfield [66, 67] reported on one of the first ASIC chips for conducting polymer sensor arrays. In this, the integrated sensor array consists of 64 parallel gold electrodes (forming 32 resistive sensors) and a ground plane fabricated on a ceramic substrate. Each conducting polymer sensor consists of a two-layer composite: the poly(pyrrole) as the base polymer and a second polymer grown electrochemically onto this base. Figure 10.14 shows a micrograph of a resistance-measuring ASIC used in the final design of the integrated nose system for interrogation of 32 conducting polymer resistors. Another sensor array consisting of five conducting polymers in a microbridge configuration has been developed at Warwick University [68]. The microbridge arrangement was used in order to reduce the temperature dependence of the discrete polymer elements. The array, which has been fabricated in a standard 0.8 lm CMOS technology, is shown in Fig. 10.15. P-type silicon (or metal electrodes) is used to form resistors for either electrochemical deposition of polymers or spray coating of carbon-black polymer composite materials, such as those reported by Lewis [69] (Caltech). These polymer materials can also be deposited on co-fired ceramic substrates, glass slides, and silicon. These materials have also been used for other micromachined gas sensor arrays. Zee and Judy [70] reported two types of devices, bulk micromachined and patterned thick-film sensors. The micromachined, so-called ‘wells’, have been designed to contain the liquid volume during deposition. This type of design permits good reproducibility in the deposition and creates larger exposure areas for sensing while minimizing the chip area. It also allows for the integration of on-chip electronics for signal conditioning and processing. Figure 10.16 shows photographs of micromachined gas sensor arrays with polymer carbon black composite materials.

Fig. 10.14 Photomicrograph of a resistance-measuring ASIC used in the design of the integrated nose system for interrogation of 32 conducting polymer resistors. From Neaves and Hatfield [67] 10.3 Integrated E-Nose Systems 247

Fig. 10.15 Photograph of a five-element CMOS gas sensor [68]

A novel ChemFET sensor array reported by Covington et al. [71] also uses carbon- black composite polymer materials. A linear dependence to toluene concentration and sensitivity of up to 2.8 lV ppm1 was observed. The device comprises an array of four n-type enhanced MOSFET sensors, fabricated by the Institute of Microtechnology (IMT, University of Neuchatel, Switzerland). Briand et al. [72] first reported on these devices as catalytic field-effect gas sensors in 2000. In this case, three of the MOSFETs had their gate covered with thin catalytic metals and were used as gas sensors, while the fourth one had a standard gate covered with nitride and acted as a reference. Sen- sitivities to the gases hydrogen and ammonia were tested. Most of the reported integrated gas array sensors are based on tin oxide technology. Gardner et al. [73] reported on an array of six MOS odour sensors on single silicon substrate with six separate integrated heaters in 1995. A sensor array reported by Das et al. [74] consists of four integrated thick-film tin-oxide gas sensors. The array was fabricated on a single substrate and the sensor responses to different concentrations of various alcohols and alcoholic beverages were reported. Another micromachined tin oxide gas sensor array composed of three different devices on the same rectangular membrane and working at different temperatures was used for the detection of NO2, CO, and toluene [75]. A sensor array consisting of 40 monolithic sensor elements with different sensitivities achieved by gradient techniques was used for halitosis analysis [76]. Forty-one parallel platinum strips partitioned the surface of the device into 40 gas-sensitive segments (SnO2 and WO3 were used). Four heating elements were based on the reverse side and the sensitivity of the array to malodour components was tested. Flexural plate wave (FPW) sensors and SAW devices have both been used as the elements for analytical sensor systems. In both sensors acoustic waves are generated within a piezoelectric substrate that has usually been coated with a chemically sensitive film. A pair of interdigital transducers is normally used to generate and receive acoustic waves. The difference between these two types of devices is that the active region of the FPW sensor is the membrane of thickness much smaller than the acoustic wave- 248 10 Integrated Electronic Noses and Microsystems for Chemical Analysis

Fig. 10.16 Photograph of gas sensors with polymer-carbon black films deposited: (a) on a custom built low-temperature co-fired ceramic substrate and (b) on a micromachined chip attached to the ceramic substrate. From Zee and Judy [70]

length. SAW sensor arrays have often been used and reported in gas analysis instruments [77, 78] but rarely as integrated arrays. Baca et al. [79], from the Sandia National Laboratories (USA), have reported on the development of a GaAs monolithic surface acoustic wave integrated circuit (Fig. 10.17) aimed at chemical sensing applications. A prototype instrument describing an integrated array of six polymer-coated FPWs used together with an absorbent pre-concentrator is reported by Cai et al. [80]. Each FPW membrane is a layered composite, 5-lm thick, consisting of a silicon nitride layer, a polished layer of p-doped polysilicon, and a ZnO piezoelectric layer attached periph- erally to the silicon substrate. The whole system is shown in Fig. 10.18. Responses to thermally desorbed samples of individual organic solvent vapours and binary and ternary vapour mixtures are reported. Another example of an integrated FPW sensor array has been recently reported by Cunningham et al. [81] of the Draper Laboratory. They have designed a chemical-vapour detection and biosensor array based on microfabricated silicon resonators (FPW sensors) coated with thin-film polymer sorption layers. The devices were fabricated on silicon-on-insulator (SOI) wafers and the work was an initial step towards the development of a large multi-element FPW array with several hundred devices operating within a single silicon chip. Bulk acoustic wave sensors, in particular the thickness mode quartz-crystal microbalances (QCM), have also been used in e-nose applications but not as integrated microsensor arrays. A monolithic sensor array based on six elements integrated on the same quartz crystal designed for monitoring agricultural emissions was reported 10.3 Integrated E-Nose Systems 249

Fig. 10.17 Micrograph of a monolithic GaAs SAW integrated circuit. From Baca et al. [79]

by Boeker et al. [82]. The dimensions of the quartz substrate are 12 mm 20 mm edge length and 168 lm thickness with resonant frequency of 10 MHz. Optical sensor arrays using image processing are another attractive technique for application in e-nose systems (see Chapter 8). A fibre-optic bead-based sensor array has been designed at Tufts University and employed to discriminate between different odours [83]. The system incorporates high-density arrays of micrometer-scale optical fibres, with polymer beads doped with fluorescent dyes placed at the end of each fibre. The binding of vapor molecules to the polymers changes the light emitted from the dyes, forming a colour signature. A similar technique has been used for the characterization of multicomponent monosaccharide solutions. In this, a chip-based sensor array composed of individually addressable polystyrene-poly(ethylene glycol) and agarose microspheres has been used. The microspheres are arranged in anisotropically etched cavities that are designed to serve as miniaturized reaction vessels and analysis chambers (Fig. 10.19). Identification of analytes takes place through colorimetric and fluorescence changes to receptor and indicator molecules, which are covalently attached to termination sites on the polymeric microspheres [84]. Photomechan-

Fig. 10.18 Photograph showing an integrated array of six polymers coated flexural plate wave sensors (FPWs) with an absorbent pre-concentrator (PCT). From Cai et al. [80] 250 10 Integrated Electronic Noses and Microsystems for Chemical Analysis

Fig. 10.19 Microspheres arranged in anisotropically etched cavities designed as miniaturized reaction vessels and analysis chambers. From Goodey et al. [84]

ical chemical microsensors reported by Datskos et al. [85] should also be mentioned. In this work it was demonstrated that photo-induced bending of microcantilevers depends on the number of absorbed molecules on their surface. The authors claim that by choosing different wavelengths tuneable chemical selectivity could be achieved. Apart from identification, a real time visualisation of gas/odour flow has also been studied (see Chapter 16 on odour tracing). A portable homogeneous gas sensor array was used to visualize the flow of a target gas, and the direction of the gas source was estimated using a real-time image-processing algorithm [86]. Finally, silicon-based microelectrode arrays for chemical analysis have been reported. An array consisting of various electrode shapes and sizes designed and used for a systematic study on some aspects of electrochemical sensing (i.e. influence of electrode geometry) was reported by Schoning et al. [87]. Sensor arrays with different electrode geometries have been studied at Warwick University for organic crystals, metal oxide, and polymer resistive devices [88–90] and offer certain functional improvements, such as faster responses or higher common-mode rejection ratios.

10.3.2 Multi-type Sensor Arrays

A study on the advantage of hybrid modular systems over monosystems, aimed at the possibility of achieving optimum discrimination power of an e-nose system, has been conducted by Ulmer et al. [91] at Tuebingen University. The system used for this comparative study consisted of 16 QCMs and MOSs. The results suggest that when- ever high reliability and a high degree of reproducibility and separation power are required in the analysis of a complex gas matrix hybrid modular systems should be used. Another hybrid instrument, designed by Dyer and Gardner [92] at Warwick University, employs both resistive and piezoelectric sensors in arrays with improved 10.4 Microsystems for Chemical Analysis 251 dynamic characteristics, and agrees with the above findings. High-precision programmable interface circuitry was developed for this system and a resolution of 0.05 % was achieved. Recently, two examples of multi-type sensor arrays have been reported. The first one consisting of four different sensors designed at ETH Zurich has been described in Chapter 9. The second system is a result of collaboration between Cam- bridge and Warwick Universities. In this, an integrated smart sensor has been developed consisting of two types of devices, chemo-resistive gas sensors and microcalor- imeteric devices with active microFET heaters and temperature sensors on an SOI membrane [93]. The smart SOI sensors can operate at temperatures up to 350 8C and offer excellent, uniform thermal distribution over the sensing area. A method for selecting an optimum sensor array has been suggested by Chaudry et al. [94]. A step-wise elimination procedure, which ranks the inclusion of sensors in an array according to their contribution to the overall sensitivity and selectivity values, was adopted in this study. Various other techniques could be used to optimize sensor array response through either smart sensor interfacing [95] or smart signal processing (i.e. adaptive thresholding for improving selectivity or signal processing for improving gas sensor response time using analogue VLSI) [96, 97]. A combination of microfluidic technology, sensor arrays, smart sensor interfacing and signals processing should result in the development of superior e-nose systems and they may, perhaps, be comparable to the conventional chemical analysis microsystems currently being developed. These micro, total analysis systems (lTAS) are described in the next section.

10.4 Microsystems for Chemical Analysis

10.4.1 Gas Chromatographs

Chromatography is a popular analytical tool commonly employed by chemists to analyze liquid and gas mixtures. Figure 10.20 illustrates the basic components of a typical gas chromatograph (GC) [98], namely, a carrier gas bottle, an injection port, a long separation column through which the gas components pass down, a detector, and a data processing system. The components in the gas mixture are separated out because the column is either coated or packed with a stationary-phase film that absorbs the different components to differing degrees. Consequently, the components travel down the tube at different rates depending on their specific sorptive properties, and

Fig. 10.20 Basic set-up of a gas chromatography system used to analyze gas mixtures 252 10 Integrated Electronic Noses and Microsystems for Chemical Analysis

Fig. 10.21 Ideal gas chromatograph in which all the components of a chemical mixture are separated out and appear as distinct peaks in the time trace

hence are partitioned out. Ideally, the components will be totally separated out in time when they emerge from the column and hence can be measured by a single detector. The ideal graph is illustrated in Fig. 10.21 with five major components clearly visible. This technique is widely used and a description can be found of it in most analytical chemistry books. However, GC systems tend to be bulky, fragile and expensive items of equipment with limited sensitivity. Gas chromatography has been used in olfaction to help analyze complex odours with only limited success. They can be used to separate out fairly large concentrations of certain organic components for which specific coatings (stationary phases) exist. GCs are also used as the front-end of an olfactometer with a person sniffing the output, instead of the sensor, and recording the specific notes as they emerge. This so-called GC olfactometer can help organoleptic panels identify the presence of certain notes in complex odors. In fact, there is some evidence that the human olfactory system generates its own spatio-temporal sorption patterns in the olfactory mucosa and so is itself a type of GC [99]. The first attempts to make a micromachined version of a GC were initially reported as long ago as 1975 by Terry at Stanford University (USA). The separation column was made from the isotropic wet etching of a silicon wafer. Figure 10.22 shows a cross- section of the device reported later in 1979 [100] with a pyrex glass lid. The system included a sample injector (silicon valve) and integrated thermal conductivity sensor but not the air supply. From 1975 to 1998 this research group further developed the micro GC and a recent review of the field has been published by Kolesar et al. [101]. Figure 10.23 shows a photograph of a micromachined GC column that is 10 lm deep, 300 lm wide, and 0.9 m to 1.5 m in length. In this case copper phthalocyanine has

Fig. 10.22 Cross-section of a micromachined GC unit showing an integrated thermal conductivity sensor at the end of the silicon column. From Terry et al. [100] 10.4 Microsystems for Chemical Analysis 253

Fig. 10.23 SEM of the cross-section of a silicon micromachined GC column [101] been sputtered down to act as the stationary phase sensitive to reactive gases. The micro column was shown to separate mixtures of ammonia and nitrogen dioxide in air. In 1997 a Japanese group led by Hannoe et al. [102] (Japanese Integrated Informa- tion and Energy Systems Laboratories) reported on the use of an ultrasonic etching technique to produce the micro channel, again with a pyrex lid, but this time a PCTFE- sputtered thin-film coating. The GC micro column has the dimensions of 10 lm deep, 100 lm across and 2 m long. Then Wiranto et al. [103], an Australian group, isotropically etched a GC column again with a pyrex lid; this time the column was 20 lm deep, 200 lm wide and only 125 cm long. The problem associated with the etching of deep channels was solved in the late 1990s with the advent of the DRIE process, and so it is now possible to make micro GC columns more accurately and with superior properties. Perhaps the most sophisticated system is that being developed by Matzke et al. [104] of the Sandia Laboratories using a plasma-etched (DRIE Bosch process) pyrex lid. The GC column is part of what is referred to as the ChemLab and Fig. 10.24a shows the schematic arrangement of this chip. The columns are now 200 to 400 lm deep with width of 10, 40 and 80 lm and lengths of only 10, 30 and 100 cm. The group plans to microfabricate a pre-concentrator and pump thus making the entire instrument on a chip as shown in Fig. 10.24a. It is also possible to try and simplify the integration process through the combination of electrophoresis to pump the mobile phase, and chromatography to separate with stationary phases, a method called micro capillary electrochromatography [105]. How- ever this technique is mainly suitable for a liquid mobile phase, and requires a high voltage supply which are incompatible with standard integrated circuit processes. 254 10 Integrated Electronic Noses and Microsystems for Chemical Analysis

Fig. 10.24 (a) Schematic layout of the ChemLab (Sandia Laborato- ries, USA). The chip is envisaged to be the size of a dime coin, (b) commercial portable ‘micro’ GC called the Chrompack and used widely for environmental gas analysis 10.4 Microsystems for Chemical Analysis 255

GC is a useful analytical tool for chemists and there are a number of companies that make portable micro GC with some micro machined parts in them – but still about the size of a computer tower case. For example, the company AST make a battery-operated unit that contains two micro GC columns with a silicon micromachined thermal conductivity sensor. Similar units are made by MTI (see Fig. 10.24b) and by Chrompack International; these so-called Chrompack units are widely used to analyze the air for organic pollutants [106]; modifications to this basic unit by Tuan et al. [107] have also been reported that seek to enhance its basic performance. However, all of these micro GCs have some major drawbacks as regards analyzing complex odours. Firstly, the time it takes for the odorant components to travel down the columns and partition can be tens or even hundreds of seconds and, secondly, the separation for some important classes of odours is relatively poor. However, there are two other analytical tools used by chemists alongside the GC, namely, the mass spectrometer and the optical spectrometer that may be regarded as complementary techniques. We shall now discuss them in turn.

10.4.2 Mass Spectrometers

The composition of a liquid (or vapour) can be analyzed using a mass spectrometer (MS). Figure 10.25 shows the general layout of an MS in which the sample is injected in to the mobile phase (normally helium gas) and the molecules ionized [108]. The ions are first accelerated in a vacuum by applying a voltage and finally separated by a magnetic field according to the ratio of their mass to charge. The number of ions is counted for each particular mass (the ions are usually singly charged species) using an ion gauge and this is commonly referred to as the abundance. The magnetic sector can be replaced by either a quadrupole electrostatic lens or a time-of-flight element to produce a more compact unit. Indeed a quadrupole mass spectrometer is now mar- keted by Agilent Technologies Inc. (USA) as the Chemical Sensor (Agilent 4440) and

Fig. 10.25 Layout of a magnetic sector mass spectrometer. From Gardner and Bartlett [108] 256 10 Integrated Electronic Noses and Microsystems for Chemical Analysis

Fig. 10.26 Photograph of the Chemical Sensor (a quadrupole mass spectrometer sold by Agi- lent Technologies)

comprises a headspace autosampler connected up to a quadrupole mass spectrometer unit and a PC for data analysis (see Fig. 10.26). This unit has been used to analyze various odorant problems and Fig. 10.27 shows the mass spectra for a complex odour generated from the headspace of a bacterial sample and covers a mass range from 45 to 550 Daltons [109]. As can be seen, the mass spectra for natural odours is complex and a pattern recognition system is needed to analyze the differences. In this example, a linear technique such as discriminant function analysis was able to resolve the differences between the growth phases. The MS, like the GC, is a fairly large, heavy, and expensive instrument. Recent efforts have been made to miniaturize parts of the MS, such as the quadrupole lens and the sampling orifice, using various micromachining techniques. For example, Fig. 10.28a shows a miniature quadrupole lens system produced by Syms et al. [110] in 1996 together with a more recent version reported by Friedhoff [111] in

Fig. 10.27 Mass spectra for the headspace of the bacteria E. coli when in two of its phases of growth. From de Matos et al. [109] 10.4 Microsystems for Chemical Analysis 257

Fig. 10.28 Micro mass spectrometer: (a) schematic parts of a quadrupole electrostatic lens, (b) photograph of a micro quadrupole lens (from Syms et al. [110]), and (c) mass spectrum from a micro mass spectrometer 258 10 Integrated Electronic Noses and Microsystems for Chemical Analysis

1999 (Fig. 10.28b). These relatively crude microsystems are capable of separating out a small number of different light masses as shown in Fig. 10.28c. Further advances are being made in the development of micro-injection ports for micro-MS instruments but the challenges associated with making a miniature ion- source detector and a vacuum system are significant. Nevertheless an integrated microfluidic-tandem MS has been reported by Figeys et al. [112] for the analysis of protein and peptide masses (in solution). This instrument examines the higher masses of 500 to 1,000 Daltons and is aimed at analyzing biological systems at the protein level rather than odours – by definition, odours have lower weights otherwise they are not volatile. Finally, it should also be noted that there are clearly many examples in which molecules of the same mass have quite different odors. For example, the position of a ketone group in undecanone causes the smell to change from fruity to rue-like. Simi- lar changes occur when comparing cis and trans isomers of unsaturated compounds. The other examples of this phenomenon may be found in Chapter 1. Consequently, there is no simple mass-activity relationship for odours. The situation is further complicated by the fact that the ionization of a single fragile odorant molecule can lead to its fragmentation and so the mass spectrum is more complicated. Consequently, the mass spectra should really be considered as a chemical signature rather than an accurate measure of the mass content in the original complex odour, and of course, there may not be a unique mapping between smell and mass signature. The combination of a GC followed by an MS instrument is a powerful and sensitive analytical tool but is clearly an extremely large and expensive unit. Making a micro GC-MS would be the ultimate challenge!

10.4.3 Optical Spectrometers

Molecules have characteristic modes of mechanical vibration and rotation, and these can be detected by looking at the amount of light at different frequencies that is absorbed by the molecules. The technique is called optical spectroscopy and the molecules are usually analyzed using light in the UV to IR range. Although the technique is generally much less sensitive to odorant molecules than GC or MS, micro spectrometer integrated circuits are being developed rapidly for the telecommunications industry. For example, Fig. 10.29 shows the principle in which light from a fibre-optic is split by a deflection grating in to its various frequencies, and these are detected using a 256-element CCD array [113]. The combination of these technologies with a microfluidic system could lead to a low-cost solution for the screening of simple odours. However, that will require improvements in both the sensitivity of the optical sensor array and the width of the frequency spectrum. 10.4 Microsystems for Chemical Analysis 259

Fig. 10.29 Optical micro spectrometer integrated circuit chip: (a) schematic and (b) actual device. From Gardner et al. [113] 260 10 Integrated Electronic Noses and Microsystems for Chemical Analysis

10.5 Future Outlook

A nose in its totality comprises a sampling system (i.e. sniffing mechanism), and a fluid flow system as well as a distributed sensor array and complex signal processing architecture. In this chapter we have described some of the current efforts towards making a micro nose that integrates the sampling and fluidic system with the actual sensing system. It is clear that the integration of a sampling system will improve the reliability and performance of an electronic nose, as it is also evident that the creation of multi-type sensor arrays enhances its dynamic range. The latter may also be further extended through the use of biological materials within the sensing elements; however the lack of stability of most biological materials exposed to the environment suggests that such bio-electronic noses could only really be used for a very short period of time. This lack of stability creates the need for a micro cassette that holds a sequence of biosensors that can be employed at the appropriate time, somewhat analogous to a photographic film cassette. Based on an analogy with the human olfactory system, the cassette would need to be wound on every 20 or so days. The production of a reliable sampling microsystem will also enable the use of the dynamical part of the sensor signal, which has been shown to be very useful [114]. However, miniaturization of the system is essential so that the time-constants associated with the physical transport of the odour around the channels and chambers are much smaller than the response times of the sensors themselves. This permits the different rate kinetics of the chemical sensors for the different analytes to be observed, and thus used to help the discrimination process. The micro channels, micro chambers and micro pumps will permit the delivery of odours extremely quickly and reproducibly to the sensor array (or mass filter), and so this should permit the creation of a new generation of dynamical micronoses. Nevertheless, the technological advances that permit the creation of such a physical embodiment of e-noses will not, in our opinion, be sufficient to solve the more complex odour problems. It is difficult to visualize a mass/optical spectrum from a mass/ optical spectrometer (miniaturized or not) resolving subtle differences in the headspace of such as cheeses and beverages. Instead, the spatio-temporal information generated by GC-based and/or sensor-based micronoses will require quite different types of signal processing algorithms from the customary the principal components analysis, discriminant function analysis and neural networking methods described in earlier chapters. The types of nonlinear dynamical filters that will be required may well be neuromorphic algorithms similar to those used in our human olfactory systems. Con- sequently, the future emphasis will turn from the construction of the miniature hardware towards the identification of suitable dynamical models, which could either be data-driven or parametric. Of course this generic approach is challenging and will lead to integrated noses whose cost may be unacceptable in some application fields. For instance, the most likely competitor to e-noses in the medical domain may be disposable biochips that seek specific proteins or protein sequences. It is unlikely that a generic micronose can compete with such a low-cost screening method. However, there are other biomedical applications in which it is possible to use an e-nose to 10.5 Future Outlook 261 screen for whole viable micro-organisms (see Chapter 18) because a protein biochip lacks such a capability.

References

1 J. W. Gardner. Intelligent gas sensing Conf. on Solid-State Transducers, using an integrated sensor pair, Sensors Southampton, UK, 1998, 575–578. and Actuators B, 26–27 (1995), 261–266. 12 M. Tuthill. A switched-current, switched- 2 A. C. Pike. University of Warwick, UK, capacitor temperature sensor in 0.6 lm 1996, PhD Thesis. CMOS, IEEE Journal of solid-state circuits, 3 P. Althainz, J. Goschnick, S. Ehrmann, 33(7) (1998), 1117–1122. H. J. Ache. Multisensor microsystem for 13 J. W. Gardner, V. Varadan, O. O. Awadel- Contaminant in Air, Int. Conf. on Solid karim. Microsensors MEMS and Smart State Sensor and Actuators, Transducers Devices, J. Wiley and Sons Ltd, Chicester, ’95, Stockholm, Sweden, 1995 pp. 2001, 503. 699–702. 14 S. Middelhoek, A. C. Hoogerwerf. Smart 4 C. L. Johnson, J. W. Schwank, K. D. Wise. sensors: when and where?, Sensors and Integrated ultra-thin-film gas sensors, Actuators B, 8, (1985) 39–48. Sensors and Actuators B, 20 (1994), 15 A. H. Taner, J. E. Brignell. Aspects of 55–62. intelligent sensor reconfiguration, Sensors

5 W. Go¨pel, K. D. Schierbaum. SnO2 sensors: and Actuators A, 47, (1995) 525–529. current status and future prospects, Sen- 16 M. A. Burns, et al.. An integrated nanoliter sors and Actuators B, 26–27, (1995) 1–12. DNA analysis device, Science, 282 (1998), 6 E. Souteyrand, D. Nicolas. E. Queau, 484–487. J. R. Martin. Influence of surface modifi- 17 M. Elwenspoek, T. S. J. Lammerink, cations on semiconductor gas sensor be- R. Miyake, H. J. Fluitman. Toward inte- haviour, Sensors and Actuators B, 26–27, grated micro liquid handling systems, (1995) 174–178. Journal of Micromechanics and Micro- 7 J. W. Gardner, P. N. Bartlett. Device for engineering, 4, (1994), 227–245. sensing volatile materials, Patent Appl. 18 S. T. Terry, J. J. Jerman, J. B. Angell. No. W093/03355, Feb. 1993. A gas-chromatographic air analyser 8 G. S. Attard, P. N. Bartlett, N. R. B. Cole- fabricated on a silicon wafer, IEEE man, J. M. Elliott, J. R. Owen, J. H. Wang. Transactions of Electron Devices, ED-26, Nanostructured platinum films from (1979), 1880–1886. lyotropic liquid crystalline phases, Science, 19 R. R. Reston, E. S. Kolesar. Jr.. Silicon 778 (1997), 838–840. micromachined gas chromatography sy- 9 F. V. Schnatz, U. Schoneberg, stem used to separate and detect ammonia W. Brockherde, P. Kopystynski, and nitrogen dioxide – Part I: Design, T. Mehlhorn, E. Obermeier, H. Bensel. fabrication, and integration of the gas Smart CMOS capacitive pressure trans- chromatography system, IEEE Journal of ducer with on-chip capability, Sensors Microelectromechanical Systems, 3, and Actuators A, 34 (1992) 77–83. (1994), 134–146. 10 C. Kuratly, Q. Huang. A fully integrated 20 A. Weisberg, H. M. Bau, J. N. Zemel. self-calibrating transmitter/receiver IC Analysis of microchannels for integrated for an ultrasound presence detector cooling, International Journal of Heat and microsystem, IEEE Journal of Solid-State Mass Transfer, 35, (1992), 2465–2473. Circuits, 33 (1998), 833–841. 21 M. J. de Boer, R. W. Tjerkstra, 11 P. A. Passeraub, P. A. Besse, A. Bayadroun, J. W. Berenschot, H. V. Jansen, E. Bernascony, R. S. Popovic. First inte- G. J. Burger, J. G. E. Gardeniers. grated inductive proximity sensor with M. Elwenspoek , A. van den Berg. on-chip CMOS readout circuit and elec- Micromachining of buried micro channels trodeposited 1 mm flat coil, The 12 Euro. in silicon, IEEE Journal of Micro- 262 10 Integrated Electronic Noses and Microsystems for Chemical Analysis

electromechanical Systems, 9, (2000), Digest of the IEEE Solid-State Sensor and 94–103. Actuator Workshop, Hilton Head Island, 22 K. A. Shaw, Z. L. Zhang, N. C. MacDonald. SC, (1990) 65–69. SCREAM I: a single mask, single-crystal 32 M. Stehr, et al.. The VAMP – a new device silicon, reactive ion etching process for handling liquids or gases, Sensors and for micromechanical structures, Sensors Actuators A, 57, (1996), 153–157. and Actuators A, 40 (1994) 63–70. 33 N. T. Nguyen, et al.. Hybrid-assembled 23 O. Geschke, W. Rong, P. T. Tang, micro dosing system using silicon-based J. P. Kutter, P. Telleman. Polymer Struc- micropump/valve and mass flow sensor, tures for lTAS, The 12th Euro. Conf. Sensors and Actuators A, 69, (1998), on Solid-State Transducers, Copenhagen, 85–91. Denmark, 2000, 291–293. 34 M. Shikida, K. Sato. Characteristics of an 24 R. W. Tjerkstra, J. G. E. Gardeniers. electrostatically-driven gas valve under J .J. Kelly, A. van den Berg. Multi-Walled high-pressure conditions, Proceedings Microchannels: Free-standing porous of MEMS’94, the 7th IEEE International silicon membranes for use in lTAS, Workshop Micro Electromechanical IEEE Journal of Microelectromechanical Systems, Oiso, Japan, (1994), 235–240. Systems, 9, (2000), 495–501. 35 B. Wagner, et al.. Micromachined bistable 25 A. Rasmussen, M. Gaitan, L .E. Locascio, valves for implantable drug delivery M.E. Zaghloul. Fabrication techniques to system, Proceedings of the 18hth Annual realize CMOS-compatible microfluidic International Conference of the IEEE Eng. microchannels, IEEE Journal of Micro- In Med. and Boi. Soc., Amsterdam, electromechanical Systems, 10, (2001), Netherlands, (1996), 254–255. 286–297. 36 N. Vandelli, D. Wroblewski, M. Velonis, 26 S. C. Terry, J. H. Jerman, J. B. Angell. A gas T. Bifano. Development of a MEMS chromatographic air analyser fabricated microvalve array for fluid flow control, on a silicon wafer, IEEE Transactions IEEE Journal of Microelectromechanical of Electron Devices ED-26, (1979), Systems, 7, (1998), 395–403. 1880–1886. 37 K. Yanagisawa, H. Kuwano, A. Tago. 27 C. A. Rich, K. D. Wise. An 8-bit microflow An electromagnetically driven microvalve, controller using pneumatically-actuated Proceedings of Transducers ’93, the 7hth microvalves, Proceedings of MEMS’99, International Conference on Solid-State the 12th IEEE International Workshop Sensors and Actuators, Yokohama, Japan, Micro Electromechanical Systems, (1993), 102–105. Orlando, FL, (1999), 130–134. 38 D. Bosch, B. Heimhofer, G. Muck, 28 E. T. Claren, C. H. Mastrangelo. Parafin H. Seidel, U. Thumser, W. Wesler. actuated surface micromachined valves, A silicon microvalve with combined Proceedings of MEMS’00, the 13th IEEE electromagnetic/electrostatic actuation, International Workshop Micro Electro- Sensors and Actuators A, 37–38, (1992), mechanical Systems, Miyazaci, Japan, 684–692. (2000), 381–385. 39 A. P. Papavasiliou, D. Liepmann, 29 T. Lisec, et al.. Thermally driven microvalve A. P. Pisano. Electrolysis-bubble actuated with buckling behaviour for pneumatic gate valve, Technical Digest of the IEEE applications, Proceedings of MEMS’94, Solid State sensor and Actuator Workshop, the 7th IEEE International Workshop Micro Hilton Head Island, SC, (2000), 48–51. Electromechanical Systems, Oiso, Japan, 40 P. F. Man, et al.. Microfabricated capillary- (1994), 13–17. driven stop valve and sample injector, 30 H. Kahn, M. A. Huff, A. H. Heuer. Proceedings of MEMS’98, the 11th IEEE The Ti Ni shape-memory alloy and its International Workshop Micro Electro- applications for MEMS, Journal of Micro- mechanical Systems, Heidelberg, mechanics and Microengineering, 8, Germany, (1998), 45–50. (1998), 213–221. 41 S. Shoji, M. Esashi. Microflow devices and 31 H. Jerman. Electrically-activated microsystems, Journal of Micromechanics and machined diaphragm valves, Technical Microengineering, 4, (1994), 157–171. 10.5 Future Outlook 263

42 R. E. Osterbroek, et al.. Designing, simu- MEMS’91, the 3rd IEEE International lation and realization of in-plane operating Workshop Micro Electromechanical microvalves, using new etching techni- Systems, Nara, Japan, (1991), 182–186. ques, Journal of Micromechanics and 53 J. A. Folta, N. F. Raley, E. W. Hee. Design Microengineering, 9, (1999), 194–198. fabrication and testing of miniature peri- 43 M. Richter, R. Linnemann, P. Woias. staltic membrane pump, Technical Digest Robust design of gas and liquid micro- of the IEEE Solid State Sensor and Actuator pumps, Sensors and Actuators A, 68, Workshop, Hilton Head Island, SC, (1992), (1998), 480–486. 186–189. 44 H. T. G. van Lintel, F. C .M. van den Pol, 54 C. Cabuz, et al.. Mesoscopic sampler based S. Bouwstra. A piezoelectric micropump on 3D array of electrostatically activated based on micromachining in silicon, diaphragms, Proceedings of Transducers Sensors and Actuators A, 15, (1988), ’99, the 10th International Conference on 153–167. Solid-State Sensors and Actuators, Sendai, 45 R. Linneman, et al.. A self-priming and Japan, (1999), 1890–1891. bubble tolerant piezoelectric silicon 55 C. Cabuz, et al.. The Dual Diaphragm micropump for liquids and gases, Pump, Proceedings of MEMS’01, the 14th Proceedings of MEMS’98, the 11th IEEE IEEE International Workshop Micro International Workshop Micro Electro- Electromechanical Systems, Interlaken, mechanical Systems, Heidelberg, Switzerland, (2001), 519–522. Germany, (1998), 532–537. 56 E. Stemme, G. Stemme. A valveless dif- 46 D. Maillefer, et al.. A high-performance fuser/nozzle-based fluid pump, Sensors silicon micropump for an implantable and Actuators A, 39 (1993), 159–167. drug delivery system, Proceedings of 57 A. Olsson, P. Enoksson, G. Stemme, MEMS’99, the 12th IEEE International E. Stemme. Micromachined flat-walled Workshop Micro Electromechanical valveless diffuser pump, Journal of Systems, Orlando, FL, (1999), 541–546. Microelectromechanical Systems, 6, 47 R. Zengerle, et al.. A bi-directional silicon (1997), 161–166. micropump, Proceedings of MEMS’95, 58 A. Olsson, O. Larsson, J. Holm, the 8th IEEE International Workshop Micro L. Lundbladh, O. Ohman. Valveless Electromechanical Systems, Amsterdam, diffuser micropumps fabricated using The Netherlands, (1995), 19–24. thermoplastic replication, Sensors and 48 M. Koch, N. Harris, A. G. R. Evans, Actuators A, 64 (1998), 63–68. N. M. White, A. Brunnschweiler. A novel 59 N.-T. Nguyen, X. Huang. Miniature micromachined pump based on thick-film valveless pumps based on printed circuit piezoelectric actuation, Sensors and board technique, Sensors and Actuators A, Actuators A, 70 (1998), 98–103. 88 (2001), 104–111. 49 W. K. Schomburg, et al.. Microfluidic 60 G. Fuhr, et al.. A micromachined electro- components in LIGA technique, Journal of hydrodynamic (EHD) pumps for liquids Micromechanics and Microengineering, 4, of higher conductivity, Journal of Micro- (1994), 186–191. electromechanical Systems, 1, (1992), 50 S. Boehm, W. Olthuis, P. Bergveld. 141–145. A plastic micropump constructed with 61 A. Richter, et al.. A micromachined elec- conventional techniques and materials, trohydrodynamic (EHD) pump, Sensors Sensors and Actuators A, 77 (1999), and Actuators A, 29 (1991), 159–168. 223–228. 62 J. R. Webster, et al.. Electrophoresis system 51 E. Meng, et al.. A check-valved silicone with integrated on-chip fluorescence diaphragm pump, Proceedings of detection, Proceedings of MEMS’00, the MEMS’00, the 13th IEEE International 13th IEEE International Workshop Micro Workshop Micro Electromechanical Electromechanical Systems, Miyazaci, Systems, Miyazaci, Japan, (2000), 62–67. Japan, (2000), 306–310. 52 J. W. Judy, T. Tamagawa, D. L. Polla. 63 O. T. Guenat, et al.. Partial electro-osmotic Surface-machined micromechanical pumping in complex capillary systems. membrane pump, Proceedings of Part 2: Fabrication and application of a 264 10 Integrated Electronic Noses and Microsystems for Chemical Analysis

micro total analysis system (lTAS) suited 75 C. Cane, I. Gracia, A. Gotz, L. Fonseca, for continuous volumetric nanotitrations, E. Lora-Tamayo, M. C. Horrillo, I. Sayago, J. Sensors and Actuators B, 72 (2001) I. Robla, J. Rodrigo, J. Gutierez. Detection 273–282. of gases with arrays of micromachined tin 64 A. C. Pike. University of Warwick, UK, oxide gas sensors, Sensors and Actuators B, 1996, PhD Thesis. 65, (2000) 244–246. 65 J. W. Gardner, P. N. Bartlett. Application of 76 S. Ehrmann, J. Jungst, J. Goschnick, conducting polymers in microsystems, D. Everhard. Application of a gas sensor Sensors and Actuators A, 51 (1995) 57–66. microarray to human breath analysis, 66 P. I. Neaves, J. V. Hatfield. Current-mode Sensors and Actuators B, 65, (2000), multiplexer for integrating resistive array 247–249. sensors, Electronics Letters, 30 (1994), 77 M. Rapp, J. Reibel, A. Voigt, M. Balzer, 942–943. O. Bulow. New miniaturized SAW-sensor 67 P. I. Neaves, J. V. Hatfield. A new gene- array for organic gas detection driven ration of integrated electronic noses, by multiplexed oscillators, Sensors and Sensors and Actuators B, 26–27 (1995) Actuators B, 65, (2000), 169–172. 223–231. 78 J. Reibel, U. Stahl, T. Wessa, M. Rapp. Gas 68 M. Cole, et al.. Active bridge polymeric analysis with SAW sensor systems, Sensors resistive devices for vapor sensing, Euro- and Actuators B, 65, (2000), 173–175. sensors XIV, the 14th European Conference 79 A. G. Baca, E. J. Heller, V. M. Hietala, on Solid-State Transducers, Copenhagen, S. A. Casalnuovo, G. C. Frye-Mason, Denmark, (2000), 895–898. J. F. Klem, T. J. Drummond. Development 69 E. J. Severin, B. J. Doleman, N. S. Lewis. of a GaAs monolitic surface acoustic wave An investigation in the concentration integrated circuit, IEEE Journal of Solid- dependence and response to analyte mix- State Circuits, 34, (1999), 1254–1258. tures of carbon black/insulating organic 80 Q. Y. Cai, J. Park, D. Heldsinger, polymer composite vapor detectors, Ana- M.-D. Hsieh, E. T. Zellers. Vapor lytical Chemistry, 72 (2000), 658–668. recognition with an integrated array of 70 F. Zee, J. W. Judy. Micromachined poly- polymer-coated flexural plate wave sensors, mer-based chemical gas sensor array, Sensors and Actuators B, 62, (2000), Sensors and Actuators A, 72 (2001) 121–130. 120–128. 81 B. Cunningham, et al.. Design, fabrication 71 J. A. Covington, J. W. Gardner, D. Briand, and vapor characterisation of microfabri- N.F. de Rooij. A polymer gate FET sensor cated flexural plate resonator sensor and array for detecting organic vapours, Sen- application to integrated sensor arrays, sors and Actuators A, 77 (2001) 155–162. Sensors and Actuators B, 73, (2001), 72 D. Briand, B. van der Schoot, N. F. de Rooij, 112–123. H. Sundgren, I. Lundstrom. A low-power 82 P. Boeker, G. Horner, S. Rosler. Monolithic micromachined MOSFET gas sensor, sensor array based on a quartz micro- Journal of Microelectromechanical balance transducer with enhanced sensiti- Systems, 9, (2000), 303–308. vity for monitoring agricultural emissions, 73 J. W. Gardner, A. Pike, N. F. de Rooij, Sensors and Actuators B, 70, (2000), M. Koudelka-Hep, P. A. Clerc, A. Hierle- 37–42. mann, W. Go¨pel. Integrated chemical 83 K. J. Albert, D. R. Walt, D. S. Gill, sensor array for detecting organic solvents, T. C. Pearce. Optical multibead arrays Sensors and Actuators B, 26 (1995), for simple and complex odour discrimi- 135–139. nation, Analytical Chemistry, 73, (2001) 74 R. R. Das, K. K. Shukla, R. Dwivedy, 2501–2508. A. R. Srivastava. Discrimination of 84 A. Goodey, et al.. Development of multi- individual gas/odour using responses of analyte sensor arrays composed of integrated thick film tin oxide sensor array chemically derivatized polymeric micro- and fuzzy-neuro concept, Microelectronics spheres localized in micromachined Journal, 30, (1999) 793–800. cavities, Journal of the American Chemical Society, 123, (2001), 2559–2570. 10.5 Future Outlook 265

85 P. G. Datskos, M. J. Sepaniak, C. A. Tripple, 97 D. M. Wilson, S. P. DeWeerth. Signal N. Lavrik. Photomechanical chemical processing for improving gas sensor microsensors, Sensors and Actuators B, 76, response time, Sensors and Actuators B, (2001), 393–402. 41, (1997), 63–70. 86 H. Ishida, T. Yamanaka, N. Kushida, 98 D. H. Desty (ed.). Gas Chromatography, T. Nakamoto, T. Moriizumi. Study of real- Butterworths, London (1958). time visualisation of gas/odour flow image 99 D. E. Hornung, S. L. Youngentob, using gas sensor array, Sensors and M. Mozell. Olfactory mucosa/air partitio- Actuators B, 65, (2000), 14–16. ning of odorants, Brain Research, 413, 87 M. J. Schoning, et al.. A silicon-based (1987), 147–154. microelectrode array for chemical analysis, 100 S. C. Terry, J. H. Jerman, J. B. Angell. A gas Sensors and Actuators B, 65, (2000), chromatographic air analyser fabricated on 284–287. a silicon wafer, IEEE Transactions on 88 J. W. Gardner, M. Iskandarani, B. Bott. Electron Devices, 26, (1979), 147–1886. Effect of electrode geometry on gas sensi- 101 E. S. Kolesar, R. R. Reston. Review and tivity of lead phthalocyanine thin films, summary of a silicon micromachined gas Sensors and Actuators B, 9, (1992), chromatography system, IEEE Compo- 133–142. nents, Packaging and Machine Technology, 89 J. W. Gardner. Intelligent gas sensing 21, (1998), 324–28. using an integrated sensor pair, Sensors 102 S. Hannoe, I. Sugimoto, K. Yanagisawa, and Actuators B, 27, (1995), 261–266. H. Kuwano. Enhanced chromatographic 90 J. W. Gardner, P. N. Bartlett, K. F. Pratt. performance of silicon-micromachined Modelling of gas-sensitive conducting capillary column with clean structure and polymer devices, IEE Proceedings of interactive plasma organic films, Int. Conf. Circuits Devices and Systems, 142, (1995), on Solid-State Sensors and Actuators, 321–333. Transducers ’97, Chicago, USA, 1997, pp. 91 H. Ulmer, J. Mitrovics, U. Weimar, 515–518. W. Go¨pel. Sensor arrays with only one or 103 G. Wiranto, N. D. Samaan, D. E. Mulcahy, several transducer principles? The advan- D. E. Davey. Microfabrication of capillary tage of hybrid modular systems, Sensors columns on silicon, SPIE, 324, (1997), and Actuators B, 65, (2000), 79–81. 59–64. 92 D. C. Dyer, J. W. Gardner. High-precision 104 M. Matzke, et al.. Microfabricated silicon intelligent interface for a hybrid electronic gas chromatographic microchannels: nose, Sensors and Actuators A, 62, (1997), fabrication and performance, SPIE, 3511, 724–728. (1998), 262–268. 93 F. Udrea, D. Setiadi, J. A. Covington, 105 S. Constantin, R. Freitag, D. Solignac, T. Dogaru, C.-C. Lu, W. I. Milne. Design A. Sayah, M. Gijs. Capillary electrochro- and simulations of a new class of matography chip integrated in cartidge, SOI CMOS micro hot-plate gas sensors, Proceedings of Eurosensors XIV, Copen- Sensors and Actuators B, 78, (2001), hagen, Denmark, 2000, 287–290. 180–190. 106 G. Etiope. Evaluation of a micro gas chro- 94 A. N. Chaudry, T. M. Hawkins, matographic technique for environmental

P. J. Travers. A method for selecting an analyses of CO2 and C1 –C6 alkanes, Jour- optimum sensor array, Sensors and nal of Chromatography A, 775 (1997), Actuators B, 69, (2000), 236–242. 243–249. 95 G. C. Cardinali, et al.. A smart sensor 107 H. P. Tuan, et al.. Novel preconcentration system for carbon monoxide detection, technique for on-line coupling to high- Analog Integrated Circuits and Signal speed narrow-bore capillary gas chroma- Processing, 14, (1997), 275–296. tography, Journal of Chromatography A, 96 D. M. Wilson, S. P. DeWeerth. Nonlinear 791, (1997), 187–195. preprocessing for smart chemical sensing 108 J. W. Gardner, P. N. Bartlett. Electronic systems, Int. Conf. On Solid-State Sensors noses: principles and applications, Oxford and Actuators, Transducers ’95, Stock- University Press, Oxford, 1999, p60. holm, Sweden, 1995, pp. 814–817. 266 10 Integrated Electronic Noses and Microsystems for Chemical Analysis

109 R. Esteves de Matos, D. J. Mason, 112 D. Figeys, S. P. Gygi, G. McKinnon, C. S. Dow, J. W. Gardner. Investigation of R. Aebersold. An integrated microfluidic- the growth characteristics of E. coli using tandem mass spectrometry system headspace analysis, in Electronic Noses and for automated protein analysis, Analytical Olfaction, eds. J.W. Gardner and K.C. Chemistry 70 3728–3734. Persaud, IOP Publishing Ltd, Bristol, 2000, 113 J. W. Gardner, V. K. Varadan, O. Awadel- 181–188. karim. Microsensors, MEMS and Smart 110 R. R. Syms, T. J. Tate, M. Ahmad, S. Taylor. Devices, J. Wiley and Sons Ltd, Chichester, Fabrication of a microengineered quadru- 2001, 436. pole electrostatic lens, Electronic Letters, 114 E. L. Hines, E. Llobet, J. W. Gardner. 32, (1996), 2094–2095. Electronic noses: a review of signal pro- 111 C. B Friedhoff, et al.. Chemical sensing cessing techniques, Proceedings IEEE: using non-optical microelectromechanical Circuits, Systems and Devices, 146, (1999), systems, Journal of Vacuum Science and 297–310. Technology, 17, (1999), 2300–2307. 267

11 Electronic Tongues and Combinations of Artificial Senses

F. Winquist, C. Krantz-Ru¨lcker, I. Lundstro¨m

11.1 Introduction

The field of measurement technology is rapidly changing due to the increased use of multivariate data analysis, which has led to a change in the attitude of how to handle information. Instead of using specific sensors for measuring single parameters, it has in many cases become more desirable to get information of quality parameters, such as sample condition, state of a process, or expected human perception of, for example, food. This is done by using arrays of sensors with partially overlapping selectivities and treating the data obtained with multivariate methods. These systems are often referred to as artificial senses, since they function in a similar way as the human senses. One such system, the electronic nose, has attracted much interest [1–3]. This concept is based on the combination of a gas sensor array with different selectivity patterns with pattern recognition software. A large number of different compounds contribute to a measured smell; the chemical sensor array of the electronic nose then provides an output pattern that represents a combination of all the components. Although the specificity of each sensor may be low, the combination of several specificity classes allows a very large number of odors to be detected. Similar concepts, but for use in aqueous surroundings have also recently been developed. These systems are related to the sense of taste in a similar way as the electronic nose to olfaction, thus, for these systems the terms ‘electronic tongue’ or ‘taste sensor’ have been coined [4–6]. In some applications, there are advantages when measuring in the aqueous phase compared to measurements in the gaseous phase; gas analysis is an indirect method that gives the final information about the aqueous phase via measurements in the gaseous phase. Many compounds, such as ions or those having a low vapor pressure, can only be measured in the aqueous phase, also for many online or inline applications it is only possible to use systems that measure directly in the solution. Furthermore, the development of electronic tongues offers an intriguing possibility to study their combinations with other types of artificial senses.

In principle, the electronic tongue or taste sensor functions in a similar way to the electronic nose, in that the sensor array produces signals that are not necessarily specific for any particular species; rather a signal pattern is generated that can be correlated to certain features or qualities of the sample. Electronic noses and tongues are normally used to give qualitative answers about the sample studied, and only in special cases to predict the concentration of individual species in the sample. Different sensing principles can also be used in electronic tongues or taste sensors, such as electrochemical methods like potentiometry or voltammetry, optical methods, or measurements of mass changes based on, for example, quartz crystals. The sense of taste may have two meanings. One aspect denotes the five basic tastes of the tongue; sour, salt, bitter, sweet, and ‘umami’. These originate from different, discrete regions on the tongue containing specific receptors called papillae. This aspect of taste is often referred to as the sensation of basic taste. The other aspect of taste is the impression obtained when food enters the mouth. The basic taste is then merged with the information from the olfactory receptor cells, when aroma from the food enters the nasal cavities via the inner passage. This merged sensory experience is referred to as the descriptive taste by sensory panels. The approach to more specifically mimic the basic taste of the tongue is made by the taste sensor system [4, 7, 8], in which different types of lipid membranes are used to determine qualities of food and liquids in terms of taste variables such as sweetness, sourness, saltiness, bitterness, and ‘umami’. There is thus a difference between the use of a sensor array as electronic tongues or as taste sensors. A taste sensor system is used to classify the different basic taste sensations mentioned above, and the results are compared with human test panels. An electronic tongue classifies a quality of one or another kind in food, such as drinks, water, and process fluids, and the results are not necessarily compared with human sensations, but with other quality properties of the sample. The concept of the electronic tongue and the taste sensor has developed very quickly during the last years due to its large potential. There are already commercial versions on the market [9, 10], and a number of other applications have also been reported, and are described later. Theperformanceofanartificialsensesuchastheelectronictonguecanbeconsiderably enhancedbythecombinationofsensorsbasedondifferenttechnologies.Thereasonis,of course,thatforeachnewmeasurementprincipleadded,anewdimensionofinformation is also added. A natural extension of this fundamental concept is the combination of different artificial senses. This is especially important when estimating the quality of food, since the guide is the impression of the human being using all five senses. A first attempt to measure the elusive parameter ‘mouthfeel’ for crispy products such as potato chips or crispbread was made by the development of an artificial mouth. The intention was to collect information mimicking three human senses: olfaction, auditory, and tactile. The samples were placed in a special ‘crush chamber’, and, while crushed, information corresponding to three senses could be obtained: ‘auditory’ by a microphone, ‘tactile’ by a force sensor, and ‘olfaction’ by a gas sensor array [11, 12]. Furthermore, combinations of electronic noses and tongues have been used for quality estimation of different wines [13, 14]. 11.2 Electronic Tongues 269

A new dimension for the assessment of human-based quality evaluation is thus obtained by using the artificial analogs to all the five human senses. All information obtained from this sensor system is then fused together to form a human-like decision. Such a sensor head has been used for quality estimation of crispy products, such as crispbread and chips [15].

11.2 Electronic Tongues

11.2.1 Measurement Principles

There are several measurement principles that have the potential to be used in electronic tongues. The most important ones are based on electrochemical techniques such as potentiometry, voltammetry, and conductometry, and there are a number of textbooks on the subject [16–18]. The use of electrochemical measurements for analytical purposes has found a vast range of applications. There are two basic electrochemical principles: potentiometric and voltammetric. Both require at least two electrodes and an electrolyte solution. One electrode responds to the target molecule and is called the working electrode, and the second one is of constant potential and is called the reference electrode. Potentiometry is a zero-current-based technique, in which a potential across a surface region on the working electrode is measured. Different types of membrane materials have been developed, having different recognition properties. These types of devices are widely used for the measurement of a large number of ionic species, the most important being the pH electrode, other examples are electrodes for calcium, potassium, sodium, and chloride. In voltammetric techniques, the electrode potential is used to drive an electron transfer reaction, and the resulting current is measured. The size of the electrode potential determines if the target molecules will lose or gain electrons. Voltammetric methods can thus measure any chemical specie that is electroactive. Voltammetric methods provide high sensitivity, a wide linear range, and simple instrumentation. Further- more, these methods also enable measurements of conductivity and the amount of polar compounds in the solution. Almost all electronic tongue or taste sensors developed are based on potentiometry or voltammetry. There are, however, also some other techniques that are interesting to use and which have special features making them useful for electronic tongues, such as optical techniques or techniques based on mass sensitive devices. Optical techniques are based on light absorption at specific wavelengths, in the region from ultraviolet via the visible region to near infrared and infrared. Many compounds have distinct absorption spectra, and by scanning a certain wavelength region, a specific spectrum for the sample tested will be obtained. Optical methods offer advantages of high reproducibility and good long-term stability. Mass sensitive devices, based on piezo electric crystals are also useful. A quartz crystal resonator is operated at a given frequency, and by the absorption of certain 270 11 Electronic Tongues and Combinations of Artificial Senses

compounds on the surface of the crystal, its frequency will be influenced [19]. For a surface acoustic wave (SAW) based device, a surface wave is propagated along the surface of the device [20], and due to adsorption of a compound in its way, the properties of this surface wave will be changed. These types of devices are very general and provide for the possibility to detect a large number of different compounds.

11.2.2 Potentiometric Devices

The equipment necessary for potentiometric studies includes an ion-selective electrode, a reference electrode, and a potential measuring device, as shown schematically in Fig. 11.1. A commonly used reference electrode is the silver-silver chloride electrode based on the half-cell reaction:

AgCl þ e ! Ag þ Cl E0 ¼þ0:22V ð1Þ

The electrode consists of a silver wire coated with silver chloride placed into a solution of chloride ions. A porous plug will serve as a voltage bridge to the outer solution. The ion-selective electrode has a similar configuration, but instead of a voltage bridge, an ion-selective membrane is applied. This membrane should be nonpor- ous, water insoluble and mechanically stable. It should have an affinity for the selected ion that is high and selective. Due to the binding of the ions, a membrane potential will develop. This potential, E, follows the well-known Nernst relation:

E ¼ E0 þðRT=nFÞlna ð2Þ

where E0 is a constant for the system given by the standard electrode potentials, R is the gas constant, T the temperature, n the number of electrons involved in the reaction, F the Faraday constant and finally, a is the activity of the measured specie. The po-

Fig. 11.1 Schematic diagram of an electrochemical cell for potentiometric measurements 11.2 Electronic Tongues 271

Fig. 11.2 Schematic diagram of an ion-sensitive field effect transistor

tential change is thus logarithmic in ionic activity, and ideally, a ten-fold increase of the activity of a monovalent ion should result in a 59.2 mV change in the membrane potential at room temperature. In the early 1970s, ion-selective field effect transistors (ISFETs) were developed, in which the ion-selective material is directly integrated with solid-state electronics [21]. A schematic diagram of an ISFET is shown in Fig. 11.2. The current between the drain and source (IDS) depends on the charge density at the semiconductor surface. This is controlled by the gate potential, which in turn is determined by ions interacting with the ion-selective membrane. In the ISFET, the normal metal gate is replaced with the reference electrode and sample solution. An attractive feature of ISFETs is their small size and ability to be directly integrated with microelectronics, for example, signal processing, furthermore, if mass fabricated, they can be made very cheaply. These features make them especially valuable for use in electronic tongues. Potentiometric devices offer several advantages for use in electronic tongues or taste sensors. There are a large number of different membranes available with different selectivity properties, such as glass membranes and lipid layers. A disadvantage is that the technique is limited to the measurement of charged species only.

11.2.2.1 The Taste Sensor The first concepts of a taste sensor were published in 1990 [22, 23]. It was based on ion- sensitive lipid membranes and developed to respond to the basic tastes of the tongue, that is sour, sweet, bitter, salt, and ‘umami’. The multichannel taste sensor [5, 23] was also based on ion-sensitive lipid membranes, immobilized with the polymer PVC. In this taste sensor, five different lipid analogs were used: n-decyl alcohol, oleic acid, dioctyl phosphate (bis-2-ethylhex- yl)hydrogen phosphate, trioctylmethyl ammonium chloride, and oleylamine, together with mixtures of these. Altogether eight different membranes were fitted on a multichannel electrode, where each electrode consisted of a silver wire with deposited silver chloride inside a potassium chloride solution, with the membrane facing the solution to be tasted. A schematic of the multichannel electrode is shown in Fig. 11.3. The voltage between a given electrode and a Ag/AgCl reference electrode was measured. The setup is shown in Fig. 11.4. This taste sensor has been used to study responses from the five typical ground tastes, HCl (sour), NaCl (salt), quinine (bitter), 272 11 Electronic Tongues and Combinations of Artificial Senses

Fig. 11.3 Schematics of the multichannel electrode with eight lipid/polymer membranes

sucrose (sweet), and monosodium glutamate (umami). The largest responses were obtained from the sour and bitter compounds, thereafter umami and salt, and for sucrose almost no response was obtained. For other sweet tasting substances, such as the amino acids glycine and alanine, larger responses were obtained. It was further shown that similar substances, such as sour substances like HCl, acetic acid, citric acid, or salty substances such as NaCl, KCl, and KBr showed similar response patterns. The system does not respond well to nonelectrolytes, which have little effect on the membrane potential [24]. The multichannel system has been commercialized [9]. The detecting part is an eight-channel multisensor, placed on a robot arm and controlled by a computer.

Fig. 11.4 The measurement setup for the eight-channel electrode system 11.2 Electronic Tongues 273

The samples to be tested are placed in a sample holder together with a cleaning solution as well as reference solutions. The measurements then take place in a special order: first the multisensor is cleaned by dipping into the cleaning solution, thereafter into the sample solution, and the cycle repeats. At certain intervals, the multisensor is placed in the reference solution for calibration purposes. This taste sensing system has been used in a number of different applications. These have mainly dealt with discrimination and estimation of the taste of different drinks. In one investigation, 33 different brands of beers were studied [25]. The samples were analyzed both by using a sensory panel and by the taste sensing system. The sensory panel expressed the taste of the different beers in the parameters sharp-versus- mild and rich-versus-light. The output pattern from the taste sensor was analyzed using principal components analysis (PCA). An interesting observation was that the first principal component corresponded well to the taste parameter rich-versus- light taste, and the second principal component corresponded well to the parameter sharp-versus-mild taste. Mineral water has also been studied using the taste sensing system [24]. A good correlation of the sensor responses to the hardness of the water could be seen in PCA plots, and also the sensor could discriminate between different brands. Other applications involve the monitoring of a fermentation process of soybean paste [26], estimation of the taste of milk [27] or coffee [28], and the development of a monitoring system for water quality [29].

11.2.2.2 Ion-Selective Electrodes The term ‘electronic tongue’ was first coined in 1996 at the Eurosensors X conference [5, 30]. The concept had been developed as a research collaboration between an Italian group (DiNatale, Davide and D’Amico) and a Russian group (Legin, Vlasov and Rud- nitskaya). This device has now been developed further, and a large number of applications have been studied, and are described in the following. The first devices consisted of potentiometric sensor arrays of two general kinds: conventional ones such as pH, sodium and potassium-selective electrodes, and especially designed ones. The latter ones were based mainly on chalcogenide vitreous materials. Altogether the array included 20 potentiometric sensors: glass, crystalline, PVC plasticized sensor, and metal electrodes. The sensor system was used for the recognition of different kinds of drinks such as tea, soft drinks, juices, and beers. Each sample was measured twice, and the information obtained from the sensor array was treated using PCA. The score plots showed good separation between all these samples. The deterioriation of orange juice during storage was also followed, and by using an artificial neural network (ANN) on the data obtained, a model for storage-time prediction could be made. The measurements of compounds of relevance for pollution monitoring in river water using this electronic tongue have also been reported [31]. River water was taken at three locations and artificially polluted with Cu, Cd, Fe, Cr, and Zn, all in ionic form, representing a ‘common’ pollution from the industries. The sensor array consisted of 22 electrodes mainly based on chalcogenide glasses and conventional electrodes. Dif- 274 11 Electronic Tongues and Combinations of Artificial Senses

ferent approaches of data analysis were performed such as multiple linear regression (MLR), projection to latent structure (PLS), nonlinear least square (NLLS), back-propagation ANN, and a self-organizing map (SOM). Two modular models were developed, the first a combination of PCA and PLS, the second a combination of ANN and SOM, and both could predict pollutant ions well. A similar setup of this electronic tongue has been used for qualitative analysis of mineral water and wine [32], and for multicomponent analysis of biological liquids [33]. A flow-injection system based on chalcogenide glass electrodes for the determination of the heavy metals Pb, Cr, Cu, and Cd was also developed [34]. The approach of combining flow injection analysis in combination with a multisensor system and analyzing data using multivariate data analysis appears very advantageous. The flow injection analysis (FIA) technique offers several advantages: since relative measurements are performed, the system is less influenced by sensor baseline drift, calibration samples can be injected within a measurement series, and the system is well adapted for automization. One should also remember that most electronic nose measurements are based on a gas-phase FIA technique, one reason is to compensate for the drift of the gas sensors.

11.2.2.3 Surface Potential Mapping Methods A very interesting technique has been developed, in which the surface potential of a semiconductor structure is measured locally [35–37]. This is a new type of a potentiometric system that provides for a contactless sensing over a surface and is thus a convenient way to analyze a multifunctional surface. It also opens up possibilities to use gradients of different functional groups as the sensing principle. The semiconductor surface acts as the working electrode on to which the test solution is applied. Into this solution a reference electrode and an auxiliary electrode are also applied. On the backside, a light-emitting diode is applied, which can scan the surface in both x and y directions. By illuminating a certain region on the semiconductor (via the backside), a photocurrent will be generated, the size being a measure of the surface potential at that particular region. In one application [35], five lipid membranes (oleic acid, lecithin, cholesterol, phos- phatidyl ethanolamine, and dioctyl phosphate) were deposited at different areas on the semiconductor surface. First, one lipid was coated onto the whole area, the next on two thirds of the area, and the third on the last third of the area. The whole surface was rotated by 908, and the procedure was repeated with the remaining lipids. The sensing area could thus be divided into nine different regions with varying composition and thickness of lipid layer. This sensor surface was then investigated for the basic taste substances, HCl (sour), NaCl (salt), quinine (bitter), sucrose (sweet), and monosodium glutamate (umami). The responses obtained had similar responses to the taste sensor system described earlier, that is the largest responses were obtained from the sour and bitter compounds, thereafter umami and salt, and for sucrose almost no response was obtained. The method has also been further developed [36–38]. 11.2 Electronic Tongues 275

11.2.3 Voltammetric Devices

In voltammetric devices, the current is measured at given potentials. This current is then a measure of the concentration of a target analyte. The reactions taking place at the electrode surface are: O þ ne ! R ð3Þ where O is the oxidized form and R is the reduced form of the analyte. At standard conditions, this redox reaction has the standard potential E0. The potential of the electrode, Ep, can be used to establish a correlation between the concentration of the oxidized (C0) and the reduced form (Cr ) of the analyte, according to the Nernst relation:

p 0 E ¼ E þ RT=nFðlnðC0=Cr ÞÞ ð4Þ

A well-known voltammetric device is the Clark oxygen electrode, which operates at 700 mV, the potential at which oxygen is reduced to hydrogen peroxide on a platinum electrode. By reverting the potential, the electrode will be sensitive to hydrogen peroxide. The use of voltammetry as a sensing principle in an electronic tongue appears to have several advantages: the technique is commonly used in analytical chemistry due to features such as very high sensitivity, versatility, simplicity, and robustness. The technique also offers the possibility to use and combine different analytical principles such as cyclic, stripping, or pulsed voltammetry. Depending on the technique, various aspects of information can be obtained from the measured solution. Normally, redox- active species are being measured at a fixed potential, but by using, for example, pulse voltammetry or studies of transient responses when Helmholtz layers are formed, information concerning diffusion coefficients of charged species can be obtained. Further information is also obtainable by the use of different types of metals for the working electrode. Whenusingvoltammetryincomplexmediacontainingmanyredox-activecompounds and different ions, the selectivity of the system is normally insufficient for specific analysis of single components, since the single steps in the voltammogram are too closetobeindividuallydiscriminated.Rathercomplicatedspectraarethereforeobtained andtheinterpretationofdataisverydifficultduetoitscomplexity.Thesevoltammograms contain a large amount of information, and to extract this there has been an increasing interest and use of multivariate analysis methods in the field [39–42]. Among the various techniques mentioned, pulse voltammetry is of special interest due to the advantages of greater sensitivity and resolution. Two types of pulse voltammetry are commonly used, large amplitude pulse voltammetry (LAPV) and small amplitude pulse voltammetry (SAPV). At the onset of a voltage pulse, charged species and oriented dipoles will arrange next to the surface of the working electrode, forming a Helmholtz double layer. A charging nonfaradic current will then initially flow as the layer builds up. The current flow, i, is equivalent to the charging of a capacitor in series with a resistor, and follows an equation of the form: 276 11 Electronic Tongues and Combinations of Artificial Senses

* * i ¼ E RSexpðt=RSBÞð5Þ

* where RS is the resistance of the circuit (¼ solution), E is the applied potential, t is the time, and B is an electrode related equivalent capacitance. The redox current from electroactive species shows a similar behavior, initially large when compounds close to the electrode surface are oxidized or reduced, but decays with time when the diffusion layer spreads out. The current follows the Cottrell equation [16–18]:

i ¼ nFADCðð1=ðpDtÞ1=2Þþ1=rxÞÞ ð6Þ

where A is the area of the working electrode, D is the diffusion constant, C is the concentration of analyte and 1/rx is an electrode constant. At constant concentration, the equation can be simplified:

1=2 i ¼ K1ð1=tÞ þ K2 ð7Þ

where K1 and K2 are constants. In LAPV, the electrode is held at a base potential at which negligible electrode reactions occur. After a fixed waiting period, the potential is stepped to a final potential. A current will then flow to the electrode, initially sharp when the Helmholtz double-layer is formed. The current will then decay as the double-layer capacitance is charged and electroactive compounds are consumed, until the diffusion-limited faradic current remains, as depicted by Eqs. (5) and (7). The size and shape of the transient response reflect the amount and diffusion coefficients of both electroactive and charged compounds in the solution. When the electrode potential is stepped back to its starting value, similar but opposite reactions occur. The excitation waveform consists of successive pulses of gradually changing amplitude between which the base potential is applied. The instantaneous faradic current at the electrode is related to surface concentrations and charge transfer rate constants, and depends exponentially on the difference of the electrode potential between the start value and the final potential. In SAPV, a slow continuous direct current (DC) scan is applied to the electrode on to which small amplitude voltage pulses are superimposed. This DC scan causes a change in the concentration profile of the electroactive species at the surface. Since only small pulse changes in the electrode potential are considered, this will result in small perturbations in the surface concentration from its original value prior to the application of the small amplitude excitation. Normally for SAPV, the current is sampled twice, one just before the application of the pulse, and one at the end of the pulse, and the difference between these is recorded as the output. This differential measurement gives a peaked output, rather than the wave-like responses that are usually obtained. 11.2 Electronic Tongues 277

11.2.3.1 The Voltammetric Electronic Tongue The first voltammetric electronic tongue described used both LAPV and SAPV applied to a double working electrode, an auxiliary, and a reference electrode [6]. The double working electrode consisted of one wire of platinum and the other of gold, both with a length of 5 mm and a diameter of 1 mm. Current and current transient responses were measured by a potentiostat connected to a personal computer (PC). The PC was also used for onset of pulses and measurement of current transient responses and to store data. Via two relays, the PC was also used to shift the type of working electrode (gold or platinum) used. Current responses from both LAPV and SAPV were collected and used as input data for PCA. In a first study, samples of different orange juices, milk, and phosphate buffer were studied. A PCA plot performed on the data showed good separation of the samples, as shown in Fig. 11.5. This electronic tongue was also used to follow the ageing process of orange juice when stored at room temperature. The voltammetric electronic tongue has been further developed. A recent configuration is shown in Fig. 11.6. It consisted of five working electrodes, a reference electrode and an auxiliary electrode of stainless steel. Metal wires of gold, iridium, palladium, platinum, and rhodium used as working electrodes were embedded in epoxy resin and placed around a reference electrode in such a way that only the ends of the working electrodes and the reference electrodes were exposed. The opposite ends of the working electrodes were connected to electric wires. The arrangement was inserted in a plastic tube ending with a stainless steel tube as an auxiliary electrode. The wires from the working electrode were connected to a relay box, enabling each working electrode to be connected separately in a standard three-electrode configuration. Dif- ferent types of pulsed voltammetry could be applied, LAPV, SAPV and staircase. In Fig. 11.7, typical voltage pulses and the corresponding current responses are shown.

Fig. 11.5 PCA analysis of different samples analyzed with the voltammetric electronic tongue 278 11 Electronic Tongues and Combinations of Artificial Senses

This electronic tongue has been used to follow the deterioration of milk due to microbial growth when stored at room temperature [43]. The data obtained were treated with PCA, and the deterioration process could clearly be followed in the diagrams. To make models for predictions, projections to latent structure and ANNs were used. When trained, both models could satisfactorily predict the proceedings of bacterial growth in the milk samples. A hybrid electronic tongue has also been developed, based on the combination of the measurement techniques potentiometry, voltammetry, and conductivity [44]. The hybrid electronic tongue was used for classification of six different types of fermented milk. Using ion-selective electrodes, the parameters pH, carbon dioxide, and chloride ion concentrations were measured. The voltammetric electronic tongue consisted of six working electrodes of different metals (gold, iridium, palladium, platinum, rhe- nium, and rhodium) and a Ag/AgCl reference electrode. The measurement principle was based on large amplitude pulse voltammetry in which current transients were measured. The data obtained from the measurements were treated with multivariate data processing based on PCA and an ANN. The hybrid tongue could separate all six different types of fermented milks. Also, the composition of the microorganisms of the different fermentations was reflected in the PCA results. A measurement system, based on the FIA technique applied to a voltammetric electronic tongue has also been developed [45]. A reference solution was continuously pumped through a cell with a voltammetric electronic tongue, and test samples

Fig. 11.6 A recent configuration of the voltammetric electronic tongue 11.2 Electronic Tongues 279

Fig. 11.7 Three different pulsed voltammetric techniques used by the voltammetric electronic tongue. The upper part shows applied voltage pulses. The lower part shows the corresponding current responses for four different electrodes (gold, iridium, palladium, and platinum) due to the onset of voltage pulses

were injected into the flow stream. Responses were obtained by measuring the resulting pulse height. The FIA technique offered several advantages: since relative measurements are performed, the system is less influenced by sensor baseline drift, calibration samples can be injected within a measurement series, and the system is well adapted for automization. The system was used to analyze standard solutions of H2O2, KCl, CuNO3,K4[Fe(CN)6], and NaCl, and results obtained were treated with multivariate data analysis. PCA showed that electrode drift was considerably decreased. The setup was also used for classification of different orange juices. The voltammetric electronic tongue has also been used for the monitoring of drinking water quality, and a review has recently been published [46].

11.2.3.2 Feature Extraction To be able to describe correctly the shape of the current pulses during the voltage pulses, a large amount of variables are collected. For each pulse, up to 50 variables can be taken for the multivariate data processing. In a complete measurement series using up to 100 pulses applied to four electrodes, a total number of up to 2000 discrete values can be collected. Most of these are redundant having a low level of information. 280 11 Electronic Tongues and Combinations of Artificial Senses

The shape of the current responses for LAPV follows Eqs. (5) and (7) in principal, which means that constants can be calculated that express the current response. In a first attempt, constants fitting Eq. (5) were calculated, and for a given application for classification of different teas, PCA showed that a better separation was obtained using these constants compared with the original data [47].

11.2.3.3 Industrial Applications using the Voltammetric Electronic Tongue The list of possible industrial applications for voltammetric electronic tongues can be made very long. Electronic tongues are versatile in their applicability since they can give general information as well as specific information, such as pH and conductivity, about a sample [48]. In addition the construction of the voltammetric electronic tongue can be made very robust – another reason that makes it suitable for many different areas of applications. One example where this quality is important is in the food industry where the use of sensors made of glass, for example, is not always acceptable. The voltammetric electronic tongue has been studied in a number of different industrial applications. One example is in the pulp and paper industry where the increasing machine speed and system closure of the papermaking process have caused an increased need to control the wet-end chemistry of the paper machine. The main challenges have been to establish knowledge of its impact on product properties as well as the most important relations between wet-end chemistry and performance of stock trades such as paper chemicals and pulp in order to improve productivity and run ability. The voltammetric electronic tongue has been evaluated on pulp samples and the prediction ability of six reference parameters – pH, conductivity, chemical oxygen demand, cationic demand, zeta potential, and turbidity – was evaluated using PLS models. The results indicated that the electronic tongue studied had very promising features as a tool for wet-end control. Flexibility, fast response and wide sensitivity spectra make the electronic tongue suitable for a vast number of possible applications in the papermaking process [49]. Another example of an industrial application where the electronic tongue has been studied is as a sensor system in household appliances such as dishwashers and washing machines. The machines are today programmed to secure a good result, which often implies that the settings, such as temperature and washing time, are too high resulting in an unnecessarily large consumption of energy, water, and detergent. A sensor that can give information about the water quality, type of soil loaded, and when the rinse water is free from detergents would increase the efficiency of these machines. The voltammetric electronic tongue has, for example, been able to distinguish between different standardized soil types, even at high levels of detergents added to the solutions [50]. Much work remains to be done before the electronic tongue might be a conventional sensor technology in this type of machine, but these preliminary studies show its potential. The third example of industrial applications for electronic tongues is as a monitoring device in drinking-water production plants [46]. The quality of drinking water varies due to the origin and quality of the raw water (untreated surface or ground water), but also due to efficiency variations in the drinking-water production process. Problems 11.2 Electronic Tongues 281 can be related to occurrence of, e.g., algae, bacteria, pesticides, and herbicides, and industrial contamination, in the raw water. The character of the raw water, and the biological activity at the production plant as well as in the distribution net may all cause quality problems such as bad odor/taste and/or unhealthiness. A method for monitoring variations in the raw water quality as well as the efficiency of separate process steps would therefore be of considerable value. To evaluate the voltammetric electronic tongue for this purpose, water samples from each of eight parallel sand filters in a drinking-water production were collected and measured, as shown in Fig. 11.8. A PCA plot for the samples is shown in Fig. 11.9. The raw water samples are well separated from the treated water samples (slow and fast filter, and clean) in the plot. One interesting observation is that the water quality after flowing through some of the slow filters cluster close to that after the fast filter, which indicates that the chemical composition of the two are similar. This result suggests that these slow filters are not working properly. The water quality after flowing through three of the other slow filters cluster, however, much closer to the clean water (which has also been chemically treated), which in a similar way indicates that these filters are working properly. This implies of course that the quality of the clean water is acceptable. Figure 11.10 demonstrates a possible use of electronic tongues (and PCA) namely to check the performance of given filters of the production plant. The results for the drinking-water plant above suggest a possible use of the electronic tongue in continuous monitoring of the status of a given filter or other parts of the plant. After maintenance of a filter, for example, the initial position in a PCA plot of the

Fig. 11.8 Top: A water production plant. Bottom: Schematics of the production plant showing the inlet of raw water, a fast filter, eight parallel sand filters and the final pH adjustment and chlorination step. The sampling positions of the electronic tongue are indicated 282 11 Electronic Tongues and Combinations of Artificial Senses

Fig. 11.9 PCA plot of samples obtained from the water production plant

water coming out of the filter is determined. Through a continuous measurement on the water after flowing through the filter, the time evaluation of the position in the PCA plot is followed. As long as the points cluster together within an area (determined initially by experience) the filter is performing well enough. Deviations from the ’normal’ cluster indicate a malfunctioning filter (Fig. 11.10). To be able to associate a deviation from the ‘normal’ cluster to any specific parameter the reasons for malfunctioning filters must be studied. For this purpose traditional analytical chemical as well as biological methods must be used. The signals from the electronic tongue can then be correlated to these reference methods, and if there is a correlation, specific distur-

Fig. 11.10 Schematic illustration of time-dependent PCA analysis used to detect changes in performance of a part of a plant 11.2 Electronic Tongues 283 bances of the properties of a filter can be tracked to certain areas of the PCA plot. The possibility to detect a malfunctioning filter, regardless of the parameters causing it, is very valuable since it allows early measures to be taken against the problem. Other application areas that are under study are the use of the electronic tongue for detection of microbial activity [48, 51]. One important industrial area for such applications is in the food industry where the quality of food is very much determined by its microbial status. This can be unwanted microbial occurrence like pathogenic microorganisms as well as wanted microbial growth in, for example, fermented foodstuff. Studies have shown that it is possible to follow the growth of mould and bacteria, and also to separate between different strains of molds with the voltammetric electronic tongue [48, 51].

11.2.3 Piezoelectric Devices

Piezoelectric materials have an interesting property in that an electric field is generated by the application of pressure, and that it is distorted by the application of an electric field. The crystal will generate a stable oscillation of the electric voltage across it when an AC voltage is applied using an external oscillatory circuit. This resonance frequency is changed with the mass of the crystal according to the equation:

Df ¼ cf 2ðDM=AÞð8Þ where Df and DM are the changes in resonance frequency and mass, respectively, c is a positive constant, f the resonance frequency, and A the electrode area. Quartz crystals are widely used as sensors where the chemical sensitivity and selectivity is obtained from an adsorbent layer on the crystal. For a quartz crystal microbalance, analyte sorption on this layer will result in a frequency change [19]. Depend- ing on the affinity properties of the adsorbing layer, different chemical compounds can be measured. Using an array based on these kinds of devices coated with hydrophilic mono- and dicarbon acids, organic and inorganic acids, and amines in drinking water could be detected [52]. A quartz resonator coated with a lipid/polymer membrane has also been investigated. The oscillation frequency showed different responses depending on taste substances and the lipid in the membrane [53]. SAW devices have also been applied for sensors in the gas and aqueous phases. For use in liquids, shear-horizontal mode SAW (SH-SAW) must be used [20, 54]. Using a

368 rotated Y-cut X-propagating LiTaO3 device, a sensing system for the identification of fruit juices was developed. The device was divided in two parts: one metallized area as reference, the other area having a free surface that was electrically active. The sensor sensitivity was controlled by changing the excitation frequency. The phase difference and amplitude ratio between the reference and sensing signals were measured. A system was developed using three SH-SAW devices operated at the frequencies 30, 50 and 100 MHz, respectively, which was used to identify eleven different fruit juices 284 11 Electronic Tongues and Combinations of Artificial Senses

[54]. In another study, a similar system was used to classify thirteen different kinds of whisky samples [55]. The device has been further studied in an application for the discrimination of four commercial brands of natural spring water [56]. Transient frequency responses were studied, and using pattern recognition based on ANNs, all four samples could be easily discriminated. A review on recent efforts towards the development of both electronic tongues and electronic noses has been published [57], in which working principles, and the construction and performance of these systems mainly based on SAWs are discussed.

11.3 The Combination or Fusion of Artificial Senses

Appreciation of food is based on the combination or fusion of many senses, in fact for a total estimation all five human senses are involved: vision, tactile, auditory, taste, and olfaction. The first impression is given by the look of the food, thereafter information of weight and surface texture is gained by holding it in the hand. Thus, even before the food has come in contact with the mouth, a first conception is already made. In the mouth, additional information is given by the basic taste on the tongue, and the smell. Other quality parameters such as chewing resistance, melting properties, crisp sound, and temperature are added. This is often referred to as the mouth-feel, and is a very important property of the food. Individual properties correlated to special food products are especially important for their characterization, such as the crispness of crispbread or chips, the chilling properties of chocolate when melting on the tongue, or the softness of a banana. A challenging problem in the food processing industry is maintenance of the quality of food products, and, consequently much time and effort are spent on methods for this. Panels of trained experts evaluating quality parameters are often used, which, however, entails some drawbacks. Discrepancy might occur due to human fatigue or stress, sensory panels are time consuming, expensive, and cannot be used for on- line measurements. The development of replacement methods for panels for objective measurement of food products in a consistent and cost-effective manner is thus highly wanted by the food industry. In this respect, the combination of artificial senses has great potential to at least in part replace these panels, since the outcome of such a combination will resemble a human-based sensory experience. For these purposes, both simple and more complex combinations of artificial senses have been investigated. Depending on the art of the quality parameters to be investigated, different types of artificial senses are important. For estimation of the crispness of potato chips, the human sense analogs of olfaction, auditory, and tactile would be satisfactory, but for total quality estimations, all five human sense analogs should be represented. Applications of the combinations of artificial senses have so far only been developed for the food and beverage industry, dealing with classification and quality issues. In the future, however, it is expected that this approach also will find applications in other types of the process industry. 11.3 The Combination or Fusion of Artificial Senses 285

An important aspect is how to fuse the sense information. How a body of algorithms, methods, and procedures can be used to fuse together data of different origins and nature in order to optimize the information content has been discussed [58, 59]. The approach of abstraction level is introduced, namely the level at which the sensor data are fused together. A low level of abstraction means that the signals from the sensors are merely added together in a matrix. A high level of abstraction means that the data of each sensor system is analyzed as a stand-alone set, thus a selection of the most important features of each system can be selected and then merged together.

11.3.1 The Combination of an Electronic Nose and an Electronic Tongue

Various applications concerning the combination of an electronic nose and an electronic tongue have been reported. In a first study, different types of wine were classified using a taste-sensor array using lipid/polymer membranes and a smell-sensor array using conducting polymer electrodes [14]. A clear discrimination was found for the different samples. Also the effect of the ageing process was studied. Later investigations performed in more detail evaluate the different information obtained from the different sensor systems, thus, in one study of wines, an electronic nose based on eight QMB sensors using different metallo porphyrins as sensing layers, and an electronic tongue based on six porphyrin-based electrodes were used [59]. The data obtained were correlated with analysis of chemical parameters. PCA loading plots showed that the artificial sensory systems were orthogonal to each other, which implies the independence of the information obtained from them. The combination of an electronic tongue and an electronic nose for classification of different fruit juices has also been described [60]. The ‘electronic nose’ was based on an array of gas sensors consisting of 10 metal-oxide-semiconductor field effect transistors (MOSFETs) with gates of thin catalytically active metals such as Pt, Ir, and Pd and four semiconducting metal-oxide-type sensors. The electronic tongue was based on pulse voltammetry, and consisted of six working electrodes of different metals, an auxiliary electrode, and a reference electrode. Using PCA, it was shown that the electronic nose or the electronic tongue alone was able to discriminate fairly well between different samples of fruit juices (pineapple, orange, and apple). It was also shown that the classification properties were improved when information from both sources were combined, both in the unsupervised PCA and the supervised PLS. An original sensor fusion method based on human expert opinions about smell and taste and measurement data from artificial nose and taste sensors have been presented [12, 61]. This is achieved by a combination of ANNs and conventional signal handling that approximates a Bayesian decision strategy for classifying the sensor information. Further, a fusion algorithm based on the maximum-likelihood principle provides a combination of the smell and taste opinions, respectively, into an overall integrated opinion similar to human beings. 286 11 Electronic Tongues and Combinations of Artificial Senses

11.3.2 The Artificial Mouth and Sensor Head

Quality estimation of crispy products such as chips or crisp bread offers an intriguing problem. The human perception of crispy quality comes from the impressions collected when the product enters the mouth and is chewed. While chewing and crushing, impressions of chewing resistance, crushing vibrations, and crushing sound as well as the descriptive taste of the sample, will all contribute to give an overall quality impression. Methods developed so far only measure the crispness in terms of the hardness and brittleness of the sample. It appears that to give a better description of the crispness experienced, more subtle quality parameters referring to the ‘mouth feel’ should be accounted for. A special ‘artificial mouth’ or ‘crush chamber’ has been designed, in which information corresponding to three senses could be obtained: ‘auditory’ by a microphone, ‘tactile’ by a force sensor, and ‘olfaction’ by a gas sensor array, thus collecting information mimicking these three human senses [11, 12]. In this artificial mouth, crispy products could be crushed under controlled conditions. The schematic of the artificial mouth is shown in Fig. 11.11. A piston could be moved at a constant speed by the action of a stepping motor connected to a lever. The force applied to the piston was recorded by a force sensor, and a dynamic microphone was placed at the bottom of the chamber. The chamber was thermostated to 37 8C. The sensor array consisted of 10 MOSFET gas sensors, with gates of thin, catalytically active metals such as Pt, Ir, and Pd, and four semiconducting metal-oxide type sensors. Five types of crispbread have been investigated, one based on wheat flour, the other four based on rye flour. The information from the three information sources was first

Fig. 11.11 Schematics of the crush chamber or ‘electronic mouth’ 11.4 Conclusions 287 individually examined. Using information from the gas sensors, only the wheat flour based crispbread could be separated from the others. Using the sound information, a correlation to the hardness and brittleness of the samples could be obtained, and similar results were obtained from the force sensor. By combining all sense analogs, all five samples could be separated [11]. The quality of potato chips has also been investigated [12]. The aim of the study was to follow the ageing process during storage. For these studies, one set of experiments was performed on potato chips stored in an opened bag, the other set in a closed bag that was opened only for sample taking. PCA analysis of data obtained from the artificial mouth showed that the information from the single information sources was not sufficient to explain the ageing process, but with merged data, the ageing process could be followed. A closer examination of the loading plot revealed that much of the data were strongly correlated, and from this plot, a smaller subset of data could be collected. This was used for an ANN, in which the prediction of age was modeled, and it was found that predicted values of age correlated well with true values. To make a complete sensory evaluation, all five human senses are involved. A new approach for the assessment of human-based quality evaluation has been obtained by the design of an electronic sensor head [15]. The investigated sample enters an artificial mouth for detection of chewing resistance and recording of the chewing sound via a microphone. A video camera is used for the identification of color, shape, and similar properties of the sample. In parallel, aroma liberated during the crushing process is measured by a gas sensor array. Finally, the crushed sample is mixed with a saline solution, and an electrochemical multi-electrode arrangement analyzes the mixture. The artificial analogs to all the five human senses are therefore used for quality evaluation of the sample. All information obtained from the sensor system is then fused together to form a human-based decision. The arrangement was originally designed for quality studies of potato chips directly atline in the factory, hence it was also equipped with a robot arm, which could take out samples from the line. This sensor head has been used for quality estimation of crispy products, such as crispbread and chips. For the chips application, it was interesting to note that vision alone could predict the quality parameters of freshness, spots, and spiciness, the olfactory analogs the amount of spiciness, and the auditory and touch analogs the freshness. The freshness of the chips can thus be determined both by change in color and by change in texture. Also, the spiciness of a chip can be determined both by the smell and by the number and color of the spices as seen by the camera. If all senses are fused together, all quality parameters could of course be correctly predicted.

11.4 Conclusions

Biomimetic measurement methods, as illustrated by the electronic nose and the electronic tongue, are rapidly being introduced in different applications. It is an interesting development where new achievements in both hardware and software act together to 288 11 Electronic Tongues and Combinations of Artificial Senses

improve the performance of the sensor arrays. Some of the techniques used, such as the pulse voltammetric measurements on a number of different (metal) electrodes, produce an enormous amount of data, in most cases with a large redundancy. An efficient data evaluation method is therefore necessary in order to utilize the measurements in an optimal way. The further development of algorithms is therefore an important task especially for sensor arrays based on simple, but well investigated, individual sensors. The biomimetic concept should, however, not be exaggerated. The human senses are strongly connected in the brain and give rise to associations based on an integrated previous experience. With regard to taste, the human taste sensation can, in general, not be described by one of the five simple ‘basic’ tastes. In olfaction, the situation is similar. One should therefore be aware of the fact that the manmade sensor arrays give responses that are only related to the taste and smell, even when they correlate with the sensation obtained by humans. Sometimes the sensor arrays do not even respond to the same molecules which give rise to the human sensation. With this knowledge in mind, the sensor arrays are still extremely useful for quality control of products and processes as indicated in this contribution. In many applications there is no need to compare the sensor signals with sensory results, the signals themselves and their variations contain enough information. In many (industrial) applications the arrays will therefore not be calibrated against humans, but against traditional analytical techniques. Another interesting possibility is to follow the evaluation of the data in a ‘human dependent’ PCA plot. In this case, process or quality monitoring can be made using references in the PCA plot itself, as discussed in correlation with the clean water production plant. A combination of electronic noses and tongues with mechanical sensors and cameras of course increases the possibility to evaluate the properties of a given sample. The experiments made so far indicate that such a ‘biomimetic sensor head’ or robot has a large potential with regard to the evaluation of food, both of raw material and finished products. Such an approach will also have uses in process and product control in general.

References

1 This book. 5 A. Legin, A. Rudinitskaya, Y. Vlasov, 2 J. W. Gardner, P. N. Bartlett. ‘A brief history C. Di Natale, F. Davide, A. D’Amico. ‘Ta- of electronic noses’, Sensors and Actuators sting of beverages using an electronic tongue 1994 B18-19 211–220. based on potentiometric sensor array’, 3 F. Winquist, H. Sundgren, I. Lundstro¨m. Technical digest of Eurosensors X, Leuven, ‘Electronic Noses for Food Control’, Belgium 1996 427–430. in Biosensors for Food Analysis, 1998, 6 F. Winquist, P. Wide, I. Lundstro¨m. A.O. Scott, Ed., The Royal Society of ‘An electronic tongue based on voltam- Chemistry, Athenaeum Press Ltd, UK. metry’, Analytica Chimica Acta 1997 357 4 K. Toko. ‘Taste sensor with global 21–31. selectivity’, Materials Science and Engineering 7 K. Toko. ‘Taste sensor’, Sensors and Actuators 1996 C4 69–82. 2000 B64 205–215. 11.4 Conclusions 289

8 K. Toko. ‘A taste sensor’, Measurement 22 K. Toko, K. Hayashi, M. Yamanaka, Science and Technology 1998 9 1919–1936. K. Yamafuji. ‘Multichannel taste sensor 9 Taste Sensing System SA401, Anritsu Corp., with lipid membranes’ Tech. Digest 9th Sens. Japan. Symp., Tokyo, Japan 1990 193–196. 10 The Astree Liquid & Taste Analyzer, 23 K. Hayashi, M. Yamanaka, K. Toko, Alpha MOS, Toulouse, France. K. Yamafuji. ‘Multichannel taste sensor 11 F. Winquist, P. Wide T. Eklo¨v, C. Hjort, using lipid membranes’, Sensors and I. Lundstro¨m. ‘Crispbread quality evaluation Actuators 1990 B2 205–213. based on fusion of information from the 24 K. Toko. ‘Biomimetic Sensor technology’, sensor analogies to the human olfactory, Cambridge University Press 2000. auditory and tactile senses’, Journal of Food 25 K. Toko. ‘Electronic Tongue’, Biosensors and Process Engineering 1999 22 337–358. Bioelectronics 1998 13 701–709. 12 P. Wide, F. Winquist, A. Lauber. ‘The per- 26 T. Imamura, K. Toko, S. Yanagisawa, ceiving sensory estimated in an artificial T. Kume. ‘Monitoring of fermentation human estimation based sensor system’, process of miso (soybean paste) using Proc. IEEE Instrumentation and Measurement multichannel taste sensor’, Sensors and Technology Conference, Ottawa, Canada, May Actuators 1996 B37 179–185. 1997. 27 H. Yamada, Y. Mizota, K. Toko, T. Doi. 13 L. Rong, W. Ping, H. Wenlei. ‘A novel ‘Highly sensitive discrimination of taste method for wine analysis based on sensor of milk with homogenization treatment fusion technique’, Sensors and Actuators using a taste sensor’, Materials Science and 2000 B66 246–250. Engineering 1997 C5 41–45. 14 S. Baldacci, T. Matsuno, K. Toko, R. Stella, 28 T. Fukunaga, K. Toko, S. Mori, D. De Rossi. ‘Discrimination of wine using Y. Nakabayashi, M. Kanda. ‘Quantification taste and smell sensors’, Sensors and of taste of coffee using sensor with global Materials 1998 10(3) 185–200. selectivity’, Sensors and Materials 1996 8(1) 15 P. Wide, F. Winquist, I. Kalaykov. ‘The 47–56. artificial sensor head: A new approach 29 A. Taniguchi, Y. Naito, N. Maeda, Y. Sato, in assessment of human based quality’, H. Ikezaki. ‘Development of a monitoring Proceedings of the Second International Con- system for water quality using a taste ference on Information Fusion, FUSION ‘99. sensor’, Sensors and Materials 1999 11(7) Int. Soc. Inf. Fusion, Mountain View, CA, 437–446. USA 2 1999 1144–1149. 30 C. Di Natale, F. Davide, A. D’Amico, 16 A. J. Bard, L. R. Faulkner. ‘Electrochemical A. Legin, A. Rudinitskaya, B. L. Selezenev, Methods – Fundamentals and Applications’, Y. Vlasov. ‘Applications of an electronic John Wiley & Sons, Inc. 1980. tongue to the environmental control’, 17 J. Wang. ‘Analytical Electrochemistry’, Technical digest of Eurosensors X, Leuven, Wiley-VCH 1994. Belgium, 1996 1345–1348. 18 P. T. Kissinger, W. R. Heineman. ‘Labora- 31 C. Di Natale, A. Macagnano, F. Davide, tory Techniques in Electroanalytical A. D’Amico, A. Legin, Y. Vlasov, A. Rudi- Chemistry’, 2nd Edition, Marcel Dekker, nitskaya, B. L. Selezenev. ‘Multicomponent Inc. 1996. analysis on polluted water by means of an 19 R. Lucklum, P. Hauptmann. ‘The quartz electronic tongue’, Sensors and Actuators crystal microbalance. Mass sensitivity, 1997 B44 423–428. viscoelasticity and acoustic amplification’, 32 A. Legin, A. Rudinitskaya, Y. Vlasov, Sensors and Actuators 2000 B70 30–36. C. Di Natale, E. Mazzone and A. D’Amico. 20 T. Yamazaki, J. Kondoh, Y. Matsui, ‘Application of Electronic tongue for quan- S. Shiokawa. ‘Estimation of components titative analysis of mineral water and wine’, in mixture solutions of electrolytes using Electroanalysis 1999 11(10–11) 814–820. a liquid flow system with SH-SAW sensor’, 33 A. Legin, A. Smirnova, A. Rudinitskaya, Sensors and Actuators B 2000 83 34–39. L. Lvova, E. Suglobova, Y. Vlasov. ‘Chemical 21 P. Bergveld. ‘The ISFET’, IEEE Trans. sensor array for multicomponent analysis of Biomed. Eng. 1970 BME-19. biological liquids’, Analytica Chimica Acta 1999 385 131–135. 290 11 Electronic Tongues and Combinations of Artificial Senses

34 J. Mortensen, A. Legin, A. Ipatov, A. Rudi- 45 F. Winquist, S. Holmin, C. Krantz-Ru¨lcker, nitskaya, Y. Vlasov, K. Hjuler. ‘A flow I. Lundstro¨m. ‘Flow injection analysis injection system based on chalcogenide applied to a voltammetric electronic tongue’, glass sensors for the determination of heavy Int. J. Food Microbiology (at press). metals’, Analytica Chimica Acta 2000 403 46 C. Krantz-Ru¨lcker, M. Stenberg, F. Win- 273–277. quist, I. Lundstro¨m. ‘Electronic tongues for 35 Y. Kanai, M. Shimizu, H. Uchida, environmental monitoring based on sensor H. Nakahara, C. G. Zhou, H. Maekawa, arrays and pattern recognition: a review’, T. Katsube. ‘Integrated taste sensor using Analytica Chimica Acta 2001 426 217–226. surface photovoltage technique’, Sensors 47 T. Artursson. Licentiate Thesis no. 148: and Actuators 1994 B20 175–179. “Development of preprocessing methods for 36 Y. Sasaki, Y. Kanai, H. Uchida, T. Katsube. multivariate sensor data”. Linko¨ping Uni- ‘Highly sensitive taste sensor with a new versity 2000. differential LAPS method’, Sensors and 48 U. Koller. Licentiate Thesis no. 859, ‘The Actuators 1995 B24-25 819–822. electronic tongue in the dairy industry’, 37 M. George, W. Parak, H. Gaub. ‘Highly Linko¨ping University 2000. integrated surface potential sensors’, Sensors 49 A. Carlsson, C. Krantz-Ru¨lcker, F. Winquist. and Actuators 2000 B69 266–275. ‘An electronic tongue as a tool for wet-end 38 Y. Murakami, T. Kikuchi, A. Yamamura, control’, unpublished. T. Sakaguchi, K. Yokoyama, Y. Ito, 50 P. Ivarsson. Licentiate Thesis no.858, M. Takiue, H. Uchida, T. Katsube, ‘Artificial senses – New technology E. Tamiya. ‘An organic pollution sensor in household appliances’, Linko¨ping based on surface photovoltage’, Sensors University 2000. and Actuators 1998 B53 163–172. 51 C. So¨derstro¨m, H. Bore´n, F. Winquist, 39 S. Brown, R. Bear. ‘Chemometric techniques C. Krantz-Ru¨lcker. ‘Analysis of mould in electrochemistry: A critical review’, growth in liquid media with an electronic Critical Reviews in Analytical Chemistry 1993 tongue’, unpublished. 24(2) 99–131. 52 R. Borngra¨ber, J. Hartmann, R. Lucklum, 40 J. M. Diaz-Cruz, R. Tauler, B. Grabaric, S. Ro¨sler, P. Hauptmann. ‘Detection of ionic M. Esteban, E. Casassas. ‘Application of compounds in water with a new polycarbon multivariate curve resolution to voltam- acid coated quartz crystal resonator’, Sensors metric data. Part 1. Study of Zn(II) com- and Actuators 2000 B65 273–276. plexation with some polyelectrolytes’, 53 S. Ezaki, S. Iiyama.‘Detection of interactions Journal of Electroanalytical Chemistry 1995 between lipid/polymer membranes and 393 7–16. taste substances by quartz resonator’ Sensors 41 J. Menditeta, M. S. Diaz-Cruz, R. Tauler, and Materials 2001 13(2) 119–127. M. Esteban. ‘Application of multivariate 54 J. Kondoh, S. Shiokawa. ‘New application curve resolution to voltammetric data. Part 2. of shear horizontal surface acoustic wave Study of metal-binding properties of the sensors to identifying fruit juices’ Japan peptides’, Analytical Biochemistry 1996 240 Journal of Applied PhysicsK 1994, 33, part I, 134–141. 3095–3099. 42 J. Simons, M. Bos, W. E. van der Linden. 55 J. Kondoh, S. Shiokawa. ‘Liquid identifica- ‘Data processing for amperometric signals’, tion using SH-SAW sensors’, Technical digest Analyst 1995 120 1009–1012. of Transducers 95 – Eurosensors IX, Stockholm 43 F. Winquist, C. Krantz-Ru¨lcker, P. Wide, 1995 716–719. I. Lundstro¨m. ‘Monitoring of milk freshness 56 A. Campitelli, W. Wlodarski, M. Houm- by an electronic tongue based on volt- mady. ‘Identification of natural spring water ammetry’ Measurement Science and Technolgy using shear horizontal SAW based sensors’, 1998 9 1937–1946. Sensors and Actuators 1998 B49 195–201. 44 F. Winquist, S. Holmin, C. Krantz-Ru¨lcker, 57 V. Varadan, J. W. Gardner. ‘Smart tongue P. Wide, I. Lundstro¨m. ‘A hybrid electronic and nose’, Proc. SPIE International Soc. Eng. tongue’, Analytica Chimica Acta 2000 406 1999, 3673, 67–76. 147–157. 11.4 Conclusions 291

58 C. Di Natale, R. Paolesse, A. Macagnano, 60 F. Winquist, P. Wide, I. Lundstro¨m. ‘The A. Mantini, A. D’Amico, A. Legin, L. Lvova, combination of an electronic tongue and an A. Rudinitskaya, Y. Vlasov. ‘Electronic nose electronic nose’, Sensors and Actuators 2000 and electronic tongue integration for B69 243–347. improved classification of clinical and food 61 P. Wide, F. Winquist, P. Bergsten, E. Petru. samples’, Sensors and Actuators 2000 B64 ‘The human based multisensor fusion 15–21. method for artificial nose and tongue data’, 59 C. Di Natale, R. Paolesse, A. Macagnano, Proc. IEEE Instrumentation and Measurement A. Mantini, A. D’Amico, M. Ubigli, A. Legin, Technology Conference, St. Paul, Minnesota, L. Lvova, A. Rudinitskaya, Y. Vlasov. USA May 1998. ‘Application of a combined artificial olfaction and taste system to the quantification of relevant compounds in red wine’, Sensors and Actuators 2000 B69 243–347. Part C Advanced Signal Processing and Pattern Analysis

12 Dynamic Pattern Recognition Methods and System Identification

E. Llobet

Abstract The field of electronic noses has developed rapidly in the past few years. There are more than 25 research groups working in this area and many companies have developed commercial instruments. Most of the work found in the literature and commercial applications, however, relate to the use of traditional static pattern analysis methods, based on either statistical or neural approaches. In this chapter, the emerging field of the dynamic analysis of the gas/odor sensor response is reviewed. The different dynamic signal processing techniques used to date include well-established parametric and non-parametric methods borrowed from the field of system identification. These include linear filters, multiexponential models, functional expansions, time series neural networks and others. The way in which all these techniques may solve electronic nose problems such as lack of selectivity, interference effects, and drift, is analyzed and some examples are discussed. Finally, a few guidelines to select a suitable model for the dynamic modeling of application-specific electronic nose systems are suggested.

12.1 Introduction

It is only in the last few years that the use of dynamic signals from a multisensor system has received any significant attention. There are several reasons why dynamic signal processing techniques are of importance to the field of electronic noses. Recent reports suggest that the dynamic response of solid-state gas sensors contains useful information about the sensor kinetics and, these vary with both sensor and analyte. This additional information can be extracted from the transient response of a sensor to a controlled change in the analyte concentration (that is, concentration modulation) or to a change in the temperature of operation of the sensor (that is, temperature modulation). In some applications the use of these techniques has resulted in an enhancement of the sensor array selectivity [1–4].

Some sensors respond very slowly to weakly interacting odors. Non-steady state measurements are required when the environmental changes are on the same time-scale as the sensor response. This may help to broaden the field of application of intelligent sensor systems (for example, continuous pollution monitoring). The sample delivery system and the sensor array are both parts of a dynamic system. The time taken for the system to reach steady-state depends on parameters such as flow rate, volume of the test chamber, diffusion rate and reaction rate. When the sensors are modeled using steady-state values, the calibration time can be very long, especially when a multicomponent calibration is performed. The calibration time is the time needed to obtain the sensor response signal, which is a multi n-dimensional non-linear function of the analytical information of all detectable n components. From the calibration, important parameters such as partial sensitivities and selectivity can be deduced. Because the dynamic modeling allows for the estimation of the steady- state sensor responses [5, 6], it may significantly reduce the time of each calibration experiment. Even when sensors are exposed to identical gas mixtures, they do not give stable responses over a long period of time. In other words, sensor signals tend to show significant temporal variation, typically referred to as long-term drift. This variation may be due to unknown processes in the sensor system, like poisoning, aging or changes in the environment, (that is, temperature and humidity). Drift may seriously affect calibration. Therefore, when an intelligent sensor system is to be operated for a long period of time, long-term drift should be addressed by the pattern recognition algorithms [7, 8]. Finally, the baseline signal (in air) and response of a sensor can depend on its previous chemical history. These changes can be considered as a short-term drift. For example, a dynamic model that uses the knowledge of present and past inputs and outputs of the sensor would be able to predict its baseline behavior. In the next section, a review of the different dynamic methods usually applied in system identification is given. This is followed by a review of the techniques that are used to identify a model from measured data. Finally, the way, in which these techniques may solve electronic nose problems, ameliorate interference effects, the and drift experienced is shown. Some guidelines to select a suitable model for application- specific electronic nose systems are then suggested.

12.2 Dynamic Models and System Identification

The techniques that are typically used to model the dynamic sensor response are borrowed from the field of system identification. System identification is the process of developing a mathematical representation of a physical/chemical dynamic system using experimental input-output data. The majority of methods that have been developed to study engineering problems assume linearity and stationarity. In the context of sensors, linearity implies that their calibration curve for all detectable components is linear, while stationarity implies that their dynamic response is not affected by time- 12.2 Dynamic Models and System Identification 295 varying trends. However, almost all real chemical transducers are characterized by non-linear dynamics and response drift. This section reviews some models for the dynamic response of odor sensors.

12.2.1 Linear Models

Linear methods have been applied in diverse fields such as econometrics, biological systems, control systems, and many others. Their application to the identification of sensor array systems for gas analysis is recent [9]. The objective of the dynamical model is to forecast the output of the sensor from knowledge of the input signals in dynamic conditions (forward modeling). Only the inversion of the model would allow us to identify the input (gases/odors) given the output signals (inverse modeling). The most common models are ARMA (Auto-Regressive Moving Average), ARX (Auto-Re- gressive with eXtra Input, also Auto-Regressive eXogenous), ARMAX, and Box-Jen- kins. These models are of interest in digital signal processing because the time series can be considered to be the output of a linear filter with a rational transfer function. In the following, their mathematical expressions are given. x[n], y[n] and e[n] are input, output and residual term or noise signals respectively. The generic relationship between these variables is depicted in Fig. 12.1. Here x[n], y[n] and e[n] are discrete- time sequences, in which the time index n assumes integer values only. This is generally the case in the context of chemical sensors, where the output signal is a sampled version of the continuous-time sensor dynamic response.

Xq Xp

ARMA ðq; pÞ : y½n¼ aiy½n iþ bje½n jð12:1Þ i¼1 j¼0

The current value of the output is modeled using q past values of the output and the present and p past values of the noise. Two different sub-models of this one can be considered. The Auto-regressive (AR) and the Moving average (MA).

AR ðqÞ : yn¼ aiy½n iþe½nð12:2Þ i¼1 Xp

MA ðpÞ : y½n¼ bje½n jð12:3Þ j¼0

Moving average models are also known as all-zero models.

Xq Xr

ARX ðq; kÞ : y½n¼ aiy½n iþ ckx½n kþe½nð12:4Þ i¼1 k¼0

The present value of the output is modeled using a linear combination of the past q values of the output, and the present and past r values of the input. 296 12 Dynamic Pattern Recognition Methods and System Identification

Fig. 12.1 A generic black-box model describes the relationship between the output (y), the measured signal or input (x) and disturbance or noise (e)

Xq Xr Xp

ARMAX ðq; k; pÞ : y½n¼ aiy½n iþ ckx½n kþ bje½n jð12:5Þ i¼1 k¼0 j¼0

Similar to the previous one but including a moving average term:

Xr Xp

Box-Jenkins ðr; p : y½n¼ ckx½n kþ bje½n jð12:6Þ k¼0 j¼0

In this model, the prediction of the output is made without the use of past values of the output. It uses present and past values of the input in addition to filtered noise. For the previous models, the parameter vector h is defined as

h ¼ðal:::aq bl:::bp cl:::cr Þ. Identifying the model requires the identification of the parameters in h. The choice of which type of model to use is highly problem-dependent, however, and there are different means of choosing a model for a particular problem, which will be discussed later in this chapter.

State-Space Models In the state-space form, the relationship between the input, noise and output signals is

written as a system of first-order difference equations using an auxiliary state vector nn. This description of linear dynamical systems became increasingly important after Kal- man’s work on prediction and linear quadratic control [10]. Insights into the physical mechanisms of the system can usually be more easily incorporated into space-state models than into the models described previously. The state-space model can be expressed as:

nnþ1 ¼ AðhÞnn þ BðhÞx½nþep½nð12:7Þ

y½n¼CðhÞnn þ em½nð12:8Þ

where A, B and C are matrices of appropriate dimensions. h is a vector of parameters

that typically correspond to unknown values of physical coefficients, em is the measurement noise and ep is the process noise acting on the states. The disturbances em[n] and ep[n] are assumed to be sequences of independent random variables. 12.2 Dynamic Models and System Identification 297

12.2.2 Multi-exponential Models

The transient response of electrochemical and chemoresistive sensors when exposed to a volatile compound is of an exponential nature [11–13]. Therefore, it seems reasonable to model the response curves of these sensors by fitting a sum of exponential functions:

XN t=si xðtÞ¼ Gie ð12:9Þ i¼1 The task of modeling a curve with a set of exponential functions is not straightforward. Because exponential functions are not an orthogonal base of functions on the real axis, the determination of the set of coefficients (Gi, si, i ¼ 1, N) from finite-time and finite- precision samples of the response transient, will not have a unique solution. There- fore, an important issue is the determination of N, the number of exponential components that should be used to fit the response transient [14]. There are different de- convolution techniques that have been applied for data analysis. These include spectral methods, such as Gardner transform [15] or multiexponential transient spectroscopy (METS) [16] and non-spectral methods, such as non-linear least squares fitting [17], and Pade-Laplace [18] or Pade-Z transforms [19]. Spectral methods do not need previous knowledge of the exponential terms. The number of peaks in the spectrum gives directly the number of exponential terms used in the model. Furthermore, the shape of peaks can give information about the adequacy of the model. For example, wider peaks suggest that two or more similar time constants have not been resolved. On the other hand, the non-linear least squares fitting method approximates the response transient with a known number of exponential terms, and thus is not suitable for component detection. Unlike spectral methods, which return a distribution that needs further analysis, non-spectral methods such as the Pade-Laplace and Pade-Z transforms attempt to identify the finite set of coefficients (Gi, si, i ¼ 1, N). Pade-Laplace and Pade-Z transforms perform data compression and feature extraction simultaneously. The following briefly reviews some of the multiexponential modeling techniques. For further details, the reader is referred to the references given. Those readers who are not interested in the maths canskip this part andproceed to sub-section 12.2.3.Some results on the use of such multi-exponential models are revised in sub-section 12.4.1.

Gardner Transform This method, which is based on the Fourier transform, was introduced forty years ago by Gardner [15]. Later, the recovery of the spectrum was improved by applying a low- pass filter before the de-convolution step of the method [20]. Assuming an experimental response function x(t) such that:

XN ait xðtÞ¼ Gie ð12:10Þ i¼1 298 12 Dynamic Pattern Recognition Methods and System Identification

Eq. (10) can be rewritten using the spectrum gðaÞ:

ð 1ð ! 1 XN at at xðtÞ¼ gðaÞe da ¼ Gidða aiÞ e da ð12:11Þ 0 i¼1 0

Making the variable change p ¼ lnðtÞ, q ¼lnðaÞ, which changes the time axis from linear to logarithmic, Eq. (11) becomes:

1ð hi xðepÞ¼ gðeqÞeqexp eðpqÞ dq ð12:12Þ 1

Considering Eq. (12), the Fourier transform of epxðepÞ can be expressed as:

1ð 1ð 1 1 FðxÞ¼pffiffiffiffiffiffi epxðepÞejxpdp ¼ pffiffiffiffiffiffi 2p 2p 0 1 1 1 1ð hi @ gðeqÞepqexp eðpqÞ dqAejxpdq ð12:13Þ 1

Finally, by defining r ¼ p q, Eq. (13) can be rewritten:

ð 1ð 1 1 FðxÞ¼pffiffiffiffiffiffi gðeqÞejxqdq er exp½er ejxrdr ¼ GðxÞKðxÞð12:14Þ 2p 1 1

Therefore, the Fourier transform GðxÞ of the spectrum gðeqÞ can be found as the ratio of F(x) and K(x), the Fourier transforms of the functions epxðepÞ and exp½er, respectively. The spectrum g(a) is related to the inverse Fourier transform of G(x) by:

gðaÞ gðeqÞdq ¼ da ð12:15Þ a

The fact that g(a) and a are coupled in Eq. (15), biases the Gardner transform towards multiexponential curves for which the product of time constant and amplitude is similar for all the exponential components.

METS METS is based on a numerical multi-differentiation of the response transient [16]. The

first order signal METS1 is defined as follows: 2 3 1ð 1ð dxðtÞ d METS ðtÞ¼ ¼ 4 GðaÞeatda5 ¼ atGðaÞeatda ð12:16Þ 1 dlnðtÞ dlnðtÞ o 0 12.2 Dynamic Models and System Identification 299

Making the variable change s ¼ 1=a, p ¼ lnðtÞ and q ¼ lnðsÞ, which is equivalent to change the time axis from linear to logarithmic, Eq. (16) can be rewritten:

1ð

METS1ðpÞ¼ hðp qÞTGðqÞdq ¼hðpÞTGðqÞð12:17Þ 1 where hðpÞ¼exp½p ep and TGðqÞ¼eqGðeqÞ. The h(p) function has a bell shape with a peak located at y ¼ 0. Therefore METS1 will present peaks at every time constant. The relative amplitude of peaks is proportional to the amplitude of the exponential component. If the h(p) function were narrower, the method would give us the time constant distribution with improved resolution power. To reach this objective, we can p th substitute the h(p) function by hnðpÞ¼exp½np e in Eq. (17), obtaining the n order signal METSn:

METSnðpÞ¼hnðpÞTGðqÞð12:18Þ

The differentiation of Eq. (18) leads to a recurrent formula for the trivial computation of METS signals from experimental data:

dMETS ðpÞ n ¼ nMETS ðpÞMETS ðpÞð12:19Þ dp n nþ1

The fact that hnðpÞ presents a peak at p ¼ lnðnÞ implies a shift towards the right of the real axis. This results in a distortion of the spectrum. Because high order METS signals are obtained by successive differentiation, the method may become very sensitive to high-frequency noise.

Pade-Laplace This method is based on the theory of Pade approximants and the Laplace transform [18, 19]. The Laplace transform of the response function defined in Eq. (10) is:

1ð XN G XðsÞ¼ estxðtÞdt ¼ i ð12:20Þ =s i¼1 s 1 i 0

The Pade-Laplace method proceeds in three steps to estimate the Laplace transform of the response transient. First, the Laplace transform is approximated at an expansion point s0 by using a Taylor series:

XK k ^ 1 d k XðsÞ¼ XðsÞj ¼ ðs s Þ ð12:21Þ k! dsk s s0 0 k¼0 300 12 Dynamic Pattern Recognition Methods and System Identification

where

1ð dk k s0t XðsÞj ¼ ¼ ðtÞ xðtÞe dt ð12:22Þ dsk s s0 0

Second, a Pade approximant is computed for the expression (21). Pade approximants are rational expressions obtained by dividing two polynomials P(s) and Q(s). The power series expansion of a Pade approximant [M/N](s), agrees with the Taylor series up to the term sMþL.

PðsÞ p þ p s þ ::: þ p sM = 0 1 M : ½M NðsÞ¼ ¼ N ð12 23Þ QðsÞ q0 þ q1s þ ::: þ qNs

And third, the partial fraction expansion of the Pade approximant yields the time constants and amplitudes from the poles and residues of the expansion, respectively. When the order of the approximant exceeds the true number of exponentials, unstable (that is, artificial) poles will become noticeable. Therefore, the method requires the computation of the [i, i þ 1] approximants for i ¼ 0; :::; N.

Pade-Z The method is similar to the Pade-Laplace, but it employs the discrete Z-transform instead of the continuous Laplace transform. If x[k] is the sampled version of the response transient x(t):

XN kT=si x½k¼ Gie ð12:24Þ i¼1

then, the Z-transform is:

XN z X½k¼ Gi ð12:25Þ T=si i¼1 z e

Similarly to the Pade-Laplace method, the Z-transform is approximated by its Taylor

series expansion at a point z0 and the [i/i ](i ¼ 1; :::; N) Pade approximants are computed for the Taylor expansion.

12.2.3 Non-linear Models

Chemical sensors are non-linear for high concentrations. Most of them are inherently non-linear even at low concentrations. Transport, adsorption and reaction processes taking place at the sensor include intrinsic non-linear dynamics. Thus, an electronic nose instrument can be represented as a non-linear system. 12.2 Dynamic Models and System Identification 301

The analysis of non-linear systems poses many problems that do not appear in their linear counterparts. For instance, the law of superposition cannot been applied and the addition of two input signals may lead to unknown results. Traditionally, the methods used to identify non-linear systems are parametric methods that make assumptions about the structure of the system. If the structure is not accurate enough, the model will not work for all inputs. Recently, a few non-linear time series and other non-linear models have been proposed. Some of them will be reviewed briefly below.

Non-Linear Time Series Some of the non-linear models are introduced in this section. The reader is referred to the work of Tong [21] for a more comprehensive survey. One of the more important classes of non-linear models is the class of non-linear auto-regression. y[n] is said to follow a non-linear auto-regressive mode of order k if there exists a non-linear function f such that:

yn¼f ðy½n 1; y½n 2; :::; y½n k; e½nÞ ð12:26Þ where e[n] is noise. As a ‘dual’ to non-linear auto-regressive models, we may have non- linear moving average models (for example, of order q):

y½n¼gðe½n; e½n 1; :::; e½n q; qÞð12:27Þ q being a vector of parameters. Since the most important linear time series model is the ARMA model, it seems natural to develop a non-linear generalization of it. For suitable k and q:

T nn ¼ð1; e½n q þ 1; :::; e½n; y½n k þ 1; :::; y½nÞ ð12:28Þ nn is called a carrier vector. Choosing suitable matrices, F, G and H, we may achieve the non-linearisation of ARMA models by introducing:

nn ¼ Fðnn1Þnn1 þ Gðnn1nn; y½n¼Hnn1 ð12:29Þ

This is formally equivalent, under suitable choices of F, G and H to:

Xq Xp

y½n¼ aiðnn1Þy½n iþv½nn1þ bjðnn1Þe½n jð12:30Þ i¼1 j¼0

The carrier vector can be regarded as a state vector and the model above as a state- dependent model (SDM) [22].

Functional Expansions Functional expansions were studied by Volterra [23] and Wiener [24]. They are valid representations of non-linear systems under very weak assumptions (stationarity). The 302 12 Dynamic Pattern Recognition Methods and System Identification

concept of a functional was introduced to describe the input/output relationship of a system. Assuming that x(t) is the input and y(t) the output, then:

yðtÞ¼F½t; xðt 0Þ; t 0 tð12:31Þ

The task of modeling consists of obtaining a mathematical expression for the functional F. This is to identify the input/output map of the system, determining the effect of past values of the input on the output. In the case of a non-linear time invariant system, F can be expressed as a Volterra functional expansion of the form:

zfflfflfflfflffl}|fflfflfflfflffl{n 1ð ð X1 1 y½t¼ ... knðs1 ...snÞxðt s1Þ ...xðt snÞds1 ...dsn ð12:32Þ n¼1 0 0

th The kernels kn ðs1; :::; snÞ constitute the descriptors of the system dynamics. The n kernel attains the effect of the cross interaction of n past values of the input on the output. Wiener redefined the basis functionals so that they were orthogonal for white Gaussian inputs.

Block-Structured Network Models Block-structured network models consist of interconnections of two different classes of blocks, which implement either dynamic linear models or static non-linear models.

Fig. 12.2 Several block-structured models for bi-input systems. Ni

blocks are static non-linear models and Li blocks are dynamic linear models. (From S. Marco et al., Sensors and Actuators B, Vol. 34, pp. 213–223 ª1996 Elsevier Science, with permission) 12.2 Dynamic Models and System Identification 303

This modeling strategy is closely related to the functional expansion method, because a close examination of the relationship between the Wiener kernels is necessary to determine the topology of the network. This method is preferred by some authors to functional expansion because of the difficulty involved in interpretation of the kernels. Furthermore, block-structured models may be related to the inner structures of the system. Figure 12.2 shows some of the different topologies (for bi-input systems) typically used in the block-structured approach. The reader is referred to the work of Chen et al. [25, 26], where a systematic structural classification procedure employing Wiener kernels is reviewed.

Neural Networks In recent years, multi-layer perceptrons (series-parallel identification method) and time-delay or recurrent neural networks (parallel identification method) have been proposed for system identification and modeling purposes [27]. It has been proved that the output of an artificial neural network (ANN), whose inputs are delayed values of the input signals, can be expressed as an infinite Volterra series [27]. In this case, since the expansion is not limited to the first or second kernels, the network is able to model highly non-linear relations if there are enough hidden neurones. The output of the network is a non-linear function of q delayed outputs and p delayed inputs:

y½k þ 1¼f ðy½k; y½k 1; ...; y½k q; x½k; ...; x½k pÞ ð12:33Þ

From the point of view of system identification, a multilayer neural network can be assumed to be a non-linear map. The elements on the weight matrices are parameters,

Fig. 12.3 Architecture of a recurrent network, which could be used to identify a single-input system 304 12 Dynamic Pattern Recognition Methods and System Identification

Fig. 12.4 (a) In the series-parallel system identification method, the neural network is supplied with lagged inputs and outputs of the system to be identified. (b) The parallel system identification method uses a neural network with feedback

whose optimum values should be found by training the ANN over a training set. Fi- gure 12.3 shows the topology of a time-delay neural network and Fig. 12.4 shows the differences between the series-parallel and the parallel identification methods. The first method is generally applied for calibration. The stability of the second method, which uses a neural network with feedback, cannot be assured [28–30].

12.3 Identifying a Model

The techniques used to identify a model from measured data typically consist of parametric or non-parametric approaches. With non-parametric techniques, very few assumptions about the system to be modeled are required, and therefore apply more generally. However, parametric techniques can sometimes lead to better results, especially when the amount of data is limited (that is, short time series). This section reviews the different techniques available for model selection.

12.3.1 Non-Parametric Approach

A linear time-invariant system can be described by its transfer function or by the corresponding impulse response. A non-linear time-invariant system can be described using functional expansions (Wiener kernels). Transfer functions, impulse responses 12.3 Identifying a Model 305 and Wiener kernels may be determined by direct techniques. Such methods are often called non-parametric since they do not explicitly employ a parameter vector in the search for a best description.

12.3.1.1 Time-Domain Methods Time-domain methods include impulse-response analysis, step-response analysis and correlation analysis. Impulse response analysis is impractical because many processes do not allow impulse inputs of such amplitude that the error is insignificant compared to the impulse response coefficients. Step-response analysis can furnish some basic characteristics to a sufficient degree of accuracy (that is, delay time, static gain, dominating time constants). Using correlation analysis, an estimate of the impulse response gˆ(t) can be obtained, through the cross-correlation of input (white noise) ^ and output signals. If the input is white noise so that RxxðsÞ¼ads, then

R^ ðsÞ ^gðsÞ¼ yx ð12:34Þ a where

XN ^ 1 RyxðsÞ¼ yðtÞxðt sÞð12:35Þ N t¼s

If the input is not white noise, then an estimate of the auto-correlation of the input can be obtained as

XN ^ 1 RxxðsÞ¼ xðtÞxðt sÞð12:36Þ N t¼s and solve

XM ^ ^ RyxðsÞ¼ ^gðkÞRxxðk sÞð12:37Þ k¼1 to estimate gˆ(k). To identify ARMA models, the estimated auto-correlation and partial auto-correlation functions of the input signal provide valuable information. Auto-regressive processes of order 1,2,… are fitted successively and the residuals calculated. The partial auto-correlation is the correlation of these residuals and the input signal. If there is a sharp cut-off in the estimated auto-correlation function after lag k, the model can be identified as an MA(k). If the auto-correlation function tails-off but the partial auto- correlation function shows a sharp cut-off after lag q, the model can be identified as an AR(q). If both functions tail-off, an ARMA model is to be used. If the auto-correlation function does not tail-off nor cut-off, the process is non-stationary. If this occurs, the data can be successively differenced until the resulting time series appears to be stationary. Differencing provides a simple way of removing trends in the data. The first 306 12 Dynamic Pattern Recognition Methods and System Identification

Fig. 12.5 Flowchart illustrating the identification of ARMA/ARIMA processes

difference of a time series y[k], Dy[k] is defined by the transformation Dy½k¼y½ky½k 1. Higher order differences are defined by successive application of the transformation. In this case, an ARIMA (auto-regressive integrated moving average) model is identified. ARIMA is an extension to the ARMA class of processes as empirical descriptors of non-stationary time series. Differencing the input signal increases the noise level, therefore smoothing of the resulting signal may be necessary. Figure 12.5 illustrates the identification process. There are many different criteria that can be used to select the order of the model. In general, they do not provide the same model order for the analyzed series of data. The reader is referred to the works of Ljung [31] and Diggle [32] for a more detailed discussion. If the system shows a non-linear behavior, it is possible to use either a linear model (this can be a good check for the relative importance of the non-linear component in the system) or a non-linear model (Wiener kernels). The reader is referred to the work of Lee and Schetzen [33], where a non-parametric method based on correlation techniques is introduced for the estimation of Wiener kernels. This method uses Gaussian white noise as the input to the system. The idea of using white noise as a stimulus in order to identify a system is based on the fact that the system is tested on all the possible inputs regarding values and frequencies (depending on the length of the test). Another approach developed by Barker [34, 35] consists of using multi-level pseudo- 12.3 Identifying a Model 307 random sequences. In a Volterra series expansion, it becomes extremely difficult to identify kernels of order three or more. Therefore, these time-domain methods are aimed at identifying second-order kernels.

12.3.1.2 Frequency-Domain Methods The frequency response of a system H(jx) may be determined from an estimation of its transfer function H(s) by setting the complex Laplace s parameter to jx. More commonly it can be determined from the time-domain signals by taking a Fourier transform (continuous or discrete) of the input x(t) and output y(t) signals, namely

YðjxÞ HðjxÞ¼ ; ð12:38Þ Xðjx 1ð 1ð 1 1 where YðjxÞ¼pffiffiffiffiffiffi yðtÞejxtdt; XðjxÞ¼pffiffiffiffiffiffi xðtÞejxtdt 2p 2p 1 1

It should be noted that the Fourier transform is a linear integral transform and x(t) and y(t) must be non-trivial (that is, non-zero) to determine the frequency response using this method. When the input x(t) is a periodic signal, the estimate of the frequency response is only of significance at the frequencies present in the input. When the input is not periodic (that is, a realization of a stochastic process), the quality of the estimate falls at those previous frequencies but is a better estimate at the other frequencies. The estimates at different frequencies are asymptotically uncorrelated. This makes the estimate of the frequency response relatively crude in practical situations [31]. Spectral analysis for determining transfer functions of linear systems was developed from statistical methods for spectral estimation. The reader is referred to the work of Brillinger [36] for a detailed account of the method. The only way to improve the poor variance properties of the transfer function estimate is to assume that the values of the true transfer function at different frequencies are related. Since the transfer function estimates at neighboring frequencies are asymptotically uncorrelated, the variance can be reduced by averaging over these (for example, using a window such as Bartlett, Parzen or Hamming). While a broad window leads to biased estimates and low variance, a narrow window leads to unbiased estimates but high variance (appearance of spurious peaks). Another way of smoothing the transfer function estimate is to split the data set into different sub-sets. The estimates over different sub-sets will be uncorrelated and averages over these can be formed. In the frequency domain, the relationship between the input X(jx) and the output Y(jx) of a non-linear system is the Volterra functional series expansion of the form:

YðjxÞ¼HðjxÞXðjxÞþH2ðjx1; jx2ÞXðjx1ÞXðjx2Þþ...

þHnðjx1; jx2; K; jxnÞXðjx1ÞXðjx2Þ^XðjxnÞþ... ð12:39Þ where H(HðjxÞ) is the linear system frequency response. The identification of non- linear characteristics in the frequency domain is, in practice, restricted to the second- 308 12 Dynamic Pattern Recognition Methods and System Identification

order kernel transformation H2ðjx1; jx2Þ, because higher order Volterra kernel transformations are difficult to display and interpret [37]. Barker [38] described a method to estimate the kernel transformations, which uses signals obtained from multi-level maximum length pseudo-random sequences [37].

12.3.2 Parametric Approach

In this approach, a set of candidate models is selected and parameterized as a model structure, using a parameter vector h. The search for the best model within the set becomes a problem of determining or estimating h. To do so, two main strategies can be considered: minimizing prediction errors and correlating prediction errors with past data. The first approach employs well-known procedures such as the least-squares method and the maximum likelihood method, and is closely related to the Bayesian maximum a posteriori estimation. The second approach is based on the correlation between the prediction error and past data. Ideally, the prediction error of a good model should be independent of past data. A pragmatic way of checking this condition is that if the prediction error is correlated with the past data, then there was more information available in the past data about the actual output than was picked up by the model (predicted output). Therefore, the model was not ideal. See Ljung [31] for a detailed review of these methods. The non-parametric approach introduced by Lee and Schetzen [33] for the estimation of the kernels that characterize a non-linear system, requires long data sequences for optimum performance. Short data sequences lead to significant errors in the estimated kernels. Haber [39] introduced a parametric method to estimate the kernels which reduces their variance, leading to a better estimation when short data series are available. Billings [40] described a method to compute second-order kernel transfor-

mations, H2ðjx1; jx2Þ, which includes estimating a non-linear auto-regressive moving average (NARMA) model (see Eq. 33). The frequency responses can be computed from the postulated model [40, 41]. To estimate HðjxÞ, x[k] is set to ejxkD and the coefficients

jxkD jx1kD jx2kD e are equated. To estimate H2ðjx1; jx2Þ, x[k] is set to e þ e and the coeffi- ðx þx Þ D cients ej 1 2 k in the model are then equated. This method requires long data sequences to be accurate. When using block-structured models, accurate kernel estimation is crucial for the identification of the topology of interconnection. Since the estimation of high order kernels is impractical, especially with short data series, the topology of the system is usually selected from a set of universal representations [42]. This selection can be based on a previous knowledge (or postulation) of the inner characteristics of the system or by performing a structural testing procedure introduced by Chen [26]. If the system being studied does not satisfy the test criteria, the structure can be rejected and another selection can be made. On the other hand, if the system satisfies the test it cannot be concluded that it has this specific structure. Once the topology has been 12.4 Dynamic Models and Intelligent Sensor Systems 309 selected, the linear time-variant blocks can be identified using cross-correlation techniques and the static non-linear blocs are usually identified by fitting a polynomial [43]. A particular case of a parametric approach is the use of multi-exponential models. Multi-exponential models, such as Gardner transform, METS, Pade-Laplace and Pade- Z transforms are parametric because an exponential response transient is assumed. To implement the Gardner transform, the experimental transient must be sampled in the logarithmic scale [20]. As the transient is normally sampled at constant time intervals, an interpolation step must be performed, which can be difficult if the experimental curve is noisy. Furthermore, the de-convolution of the FFT of the spectrum favors high-frequency components (experimental noise). Therefore, the low-pass filtering of the FFT of the spectrum prior to the de-convolution process, leads to a better signal-to-noise ratio at the price of a lower spectral resolution. Similarly to Gardner transform, METS requires logarithmic sampling (or interpolation) of the experimental response transient. But the implementation of the method is easier compared to the Gardner transform [16]. Pade-Laplace and Pade-Z transform methods require the selection of an expansion point to approximate the Laplace (or Z) transform by a Taylor series. The selection of the expansion point is an important issue because both methods will not work properly for all the values of this point. If the expansion point is too small, the numerical integration in Eq. (22) will not converge in the time range provided by the samples of the experimental measurements. If the expansion point is too large, the numerical integration in Eq. (22) will truncate the data too early and the slowest poles will not be identified. There exist several heuristic search methods to find an optimal value for the expansion point [18, 44].

12.4 Dynamic Models and Intelligent Sensor Systems

In this section we briefly review the modeling techniques in the context of electronic nose systems. The models and techniques used so far aim to enhance the sensor array selectivity, to reduce the time necessary for calibration (for example, forecasting the steady-state response using the transient response) and to counteract drift. A summary of the main approaches is shown in Table 12.1. The main ones will be discussed in more detail later. Before applying any technique to dynamically model the sensor system, sensors that are not relevant for the specific application, or that do not work properly, should be eliminated. This requires careful ‘pre-analysis’ of the system. The use of classical techniques such as PCA may be very helpful in this preliminary stage. The reader is referred to Chapter 5 of this book for a detailed account of the different pre-processing techniques. 310 12 Dynamic Pattern Recognition Methods and System Identification

Tab. 12.1 Types of modeling approaches in intelligent sensor systems

Modeling Technique Identification Technology Application Ref.

Linear filters, and Parametric ARMA, Thick-film SnO2 Calibration time [5] state-space models sensor oriented model reduction Parametric ARX, 1 sensor 4 QMB Sensor response [45] sensor oriented Prediction Parametric Box-Jen- polymer coated 10 Identification of [8] kins, sensor oriented MOSFETs 2 gases. Drift rejection 2 thick-film SnO2 Parametric AR, 6 QMB Identification of [9, 46] sensor oriented 3 vapors Parametric state- polymer coated Quantitative analysis [47] space model, system of ternary mixtures oriented Parametric Box- 4 QMB polymer Quantitative analysis [48]

Jenkins FIR, sensor coated, 2 SnO2 of 2 vapors oriented 6 BAW polymer coated

Multiexponential Parametric, sensor Resistive Feature extraction [5, 13] models oriented (metal oxide and for odor recognition conducting polymer)

Functional expansions Non-parametric, 6 QMB Sensor response [48, 49] (non-linear) correlation polymer coated prediction [50] techniques. sensor oriented 4QMB Sensor response [45] Parametric, polymer coated prediction sensor oriented

Block-structured Parametric 6QMB Structure [45, 50] combining correlation polymer coated identification, and polynomial fitting response prediction

Neural networks

– SOM Non-parametric, Arrays of SnO2, Gas/aroma [45, 51, 52] – Time-delay system oriented, MOSFET and QMB identification, [6, 48, 53] – ART adaptive polymer coated drift rejection [54–60] – fuzzy ART

Other techniques Ad-hoc models Parametric, Metal oxides, Sensor selectivity [3, 61, 62] through odor or sensor oriented conducting enhancement, [2, 4, 63, 64] temperature models or FFT polymers, QMB gas/aroma [1, 65–70] modulation and techniques polymer coated identification and noise techniques quantification 12.4 Dynamic Models and Intelligent Sensor Systems 311

12.4.1 Dynamic Pattern Recognition for Selectivity Enhancement

To date most of the attempts to use transient information in the sensor signal are based on ad hoc models. These models allow for the estimation of parameters that characterize the transient response conferring some selectivity on the sensors. Gen- erally, an advantage of these models is that they account for physical and chemical properties of the sensing material (e.g. diffusion, reaction). Therefore, some insight into the sensors’ dynamic behavior can be realized [63, 64]. Their main weakness is

Fig. 12.6 Results of a PCA analysis of the response of a four-element tin-oxide electronic nose to three organic volatile compounds using static (a) and dynamic (b) signals. Results of a PCA of the response to binary mixtures using static (c) and dynamic (d) signals. (From E. Llobet et al., in Proceedings of IEEE Solid-state Sensors and Actuators Conference, Transducers, Vol. 2, pp. 971–974, ª1997 IEEE, with permission) 312 12 Dynamic Pattern Recognition Methods and System Identification

that transient signals are influenced by previous measurements (memory effect) and by drift (for example, aging of the sensor, variations in temperature or humidity). Since these aspects are not considered by the models, the pattern recognition ability of a sensor system which is initially learnt can deteriorate after a period of time. Fig- ure 12.6 shows the PCA results when an array of 4 thick-film tin oxide gas sensors were used to identify different volatile organic compounds and their binary mixtures [2]. The use of transient signals such as the rise time of the sensor conductance when the odor concentration varies stepwise, helps in the identification task. The identification of single components, using a feed-forward back-propagation trained neural network gave a 76 % success rate (using static signals only) and a 100 % success rate (using both static and dynamic signals). The success rate in the identification of binary mixtures increased from 75 % (using static signals) to 86 % (using static and dynamic signals). In a recent study [13], different techniques to identify multiexponential models were used to analyze the response transients of a 32-element sensor array. The sensors were based on conducting polymers and the modeling was carried out in the context of odor recognition. Figure 12.7 shows a typical response of the polymer sensors to fruit juice. Two spectral methods (Gardner transform and METS) and two non-spectral methods (Pade-Laplace and Pade-Z transform) were investigated. The results of applying these methods to the sensors’ responses are shown in Fig. 12.8. Both non-spectral methods outperformed the spectral ones. The slow sampling rate of the transients and the experimental noise required previous smoothing of the experimental signals. The Gard- ner transform was found to be very sensitive to the smoothing process. In METS, the differentiation of the transient and associated decrease of the signal-to-noise ratio, prevented higher-order signals to be of use. Therefore, both spectral methods were able to identify one exponential component. Non-spectral methods were found to be less sensitive to experimental noise and the response transients could be modeled with two exponential components. These methods led to very similar results. The dots

in Fig. 12.8 (bottom) are the exponential components (Gi, si) for each sensor. While the clusters with small time constants account for the initial transient of the signal, the scattered clusters with higher time constants represent the steady state. These scatter

Fig. 12.7 Typical response of a 32-element conducting polymer sensor array to fruit juice 12.4 Dynamic Models and Intelligent Sensor Systems 313

Fig. 12.8 Results of applying different multiexponential methods to model the tran-

sients shown in Fig. 12.7. Top left: Gardner transform. Top right: METS1. Time constants

(si) and amplitudes (Gi) derived from the Pade-Laplace (bottom left) and Pade-Z (bottom right) methods, respectively. After [13]

diagrams can be though of as odor signatures, which can be of use for odor recognition. However, since exponential functions are not an orthogonal basis of functions, further work is needed to check the repeatability of the extracted signatures. Other dynamic pattern recognition methods for selectivity enhancement consist of modulating the working temperature of the sensor or using an a.c. interrogation technique. The reader is referred to chapters 5 and 16 of this book for a detailed account of these methods. A variation of the a.c. interrogation technique is the pseudo-random binary sequence (PRBS) interrogation technique [69]. A PRBS voltage is applied to the gas sensor electrodes and the output signal is then taken from a resistive load connected in series to the resistive (conducting polymer) sensor. PRBS are easy to obtain and have a nearly uniform power spectral density (PSD) over a wide frequency band. Figure 12.9 shows a PBRS generator and the signal PSD. PRBS are interesting because they are deterministic, and thus measurements are repeatable. The output signal is processed using the FFT to convert it from the time domain to the frequency domain. The energy spectral density (ESD) of the output signal is a characteristic feature of the gas sensor in the presence of an odor. Figure 12.10 shows the ESD of a conducting polymer gas sensor in the presence of methanol and acetone [69]. The relative amplitude of peaks can be seen as a fingerprint for the tested odors. Another strategy consists of measuring the PSD of the random resistance fluctuations of a d.c. biased resistive sensor. It has been shown [70] that for a conductive polymer sensor, a significant variation in the PSD is obtained in the presence of odors. 314 12 Dynamic Pattern Recognition Methods and System Identification

Fig. 12.9 A pseudo-random binary sequence (PRBS) generator and power spectrum density (PDS) of the generated sequence. (Reprinted from M.E.H. Amrani et al., Sensors and Actuators B, Vol. 47, pp. 118–124 ª1998 Elsevier Science, with permission)

12.4.2 Calibration Time Reduction

Some applications of sensor response prediction aim to reduce the time necessary to calibrate the sensor array for the gases/odors of interest. Results with ARMA and multi-exponential models applied to the dynamic response of tin oxide sensor arrays have been reported [5]. The dynamic models were used to predict the static response of the sensors to small concentrations of nitrogen dioxide (0–9 ppm). Because the auto- correlation for the transient response of the sensors tailed-off and the partial autocor- relation cut-off after lag 1, an AR(1) model was identified (see Fig. 12.5). However, this AR model was found to underestimate the static response of the sensors. The computation of the first-order METS (see Eq. (17)) for the transients, which showed two peaks, suggested that two exponentials were suitable for the modeling of the sensor response. Table 12.2 shows the relative errors made by the dynamic multiexponential model, which performed better than the AR(1) model in the extrapolation of the gas concentration. In this application, the prediction of the static response from the initial part of the dynamic response permits a reduction of the calibration time by a factor of four. 12.4 Dynamic Models and Intelligent Sensor Systems 315

Fig. 12.10 (a) Energy spectral density of the gas sensor response to 500 ppm methanol vapor. (b) Energy spectral density of the gas sensor response to 500 ppm acetone vapor. (Reprinted from M.E.H. Amrani et al., Sensors and Actuators B, Vol. 47, pp. 118 –124 ª1998 Elsevier Science, with permission)

12.4.3 Building of Response Models

Dynamic measurements are interesting when the odors or the environmental conditions undergo changes with the same time-scale as the sensor response times. This situation is not uncommon because chemical sensors are often slow responding devices. In such cases, the inversion of the dynamical model allows for the concentrations input to the sensor or sensor array to be reconstructed. Another advantage of dynamical models compared with static models is the possibility of predicting future sensor responses from the knowledge of their past and present inputs and outputs. Methods of dealing with noise that allow for calculating the impulse response (of linear systems) or the Wiener kernels (of non-linear systems), using the correlation 316 12 Dynamic Pattern Recognition Methods and System Identification

Tab. 12.2 Relative errors made by a multi-exponential model in the

extrapolation of the concentration value of NO2 at different calibration times. (Reprinted from C. DiNatale et al., Sensors and Actuators B, Vol. 24–25, pp. 578–583, ª1995 Elsevier Science, with permission)

Time (s) Error at 1 ppm (%) Error at 6 ppm (%) Error at 9 ppm (%)

100 55.2 35.8 13.1 200 17.1 4.7 7.1 400 7.8 2.5 3.7 800 1.3 0.3 0.6

approach, appear to be useful for constructing models for the sensor response to different odors. Linear filters that use lagged values of the input and the output (i.e. previous values of these signals) to characterize the sensor (sensor oriented models) or the sensor array (system oriented models), identified using parametric approaches, such as the least-squares method, are also promising. In [8] Box-Jenkins linear filters were applied to model an array of metal oxide and MOSFET odor sensors in the presence of four alcohols and water vapor. Five models for each sensor were created (one for each alcohol and one for water vapor). The classification was done in prediction error space, and the alcohol whose model gave the lowest total squared prediction error for all sensors was identified as the unknown odor (Bayesian ap-

Fig. 12.11 Prediction errors for all the 5 models used for each sensor when the measured gas was 1-propanol. (Reprinted from M. Holmberg et al., Sensors and Actuators B, Vol. 35–36, pp. 528–535, ª1996 Elsevier Science, with permission) 12.4 Dynamic Models and Intelligent Sensor Systems 317

Fig. 12.12 Complete scheme of the estimated two-input Wiener model of a polymer-coated QMB. (Reprinted from F. Davide et al., Sensors and Actuators B, Vol. 24–25, pp. 830 –842, ª1995 Elsevier Science, with permission) proach). Figure 12.11 shows the total sum squared prediction error for all sensors and for every model when the measured gas was 1-propanol. The 1-propanol model gives the lowest prediction error in almost all cases, mostly leading to a correct classification. However, linear and non-linear models constructed using input-output data (black-box models) do not give any insight into the inner structure of the sensors. In other words, it is not possible to discuss the identified model in terms of physical or chemical properties of the system. On the other hand, block-structured models are more related to the intrinsic characteristics of the sensing mechanisms. Figure 12.12 shows the scheme of a two-input block-structured model of a polymer-coated quartz-microbalance sensor in the presence of n-octane and toluene [50]. The impulse response of the two linear blocks, which describe all the memory effects of the system, were estimated using the cross-correlation approach. The static input-output non-linearity was estimated by fitting a five-order polynomial. However, this method has not been widely applied because the identification of the model is complicated. In fact, the use of a non- parametric approach, such as the cross-correlation method, to estimate the impulse response with low errors, requires long data sequences. This can result in time consuming measurements to identify the sensor array or even worse, can be impractical in some applications.

12.4.4 Drift Counteraction

Because all of the approaches described above include memory effects, they are generally useful to address the problem of short-term drift (effects in the present response of the system due to measurements in its recent past). Another strategy consists of 318 12 Dynamic Pattern Recognition Methods and System Identification

Fig. 12.13 Comparison of the identification performances of non-adaptive and adaptive SOMs of a six-element tin-oxide gas sensor array in the presence of simulated drift. (From S. Marco et al., in Proceedings of IEEE Instrumentation and Mea- surement Technology Conference, pp. 904–907, ª1997 IEEE, with permission)

using neural networks with residual plasticity. This allows the networks to deal effectively with small variations in the sensor response [51, 55]. Long-term drift caused by sensor poisoning or aging implies that the system under identification is non-stationary. All the methods, except the neural network approach, assume that the sensor system is stationary and thus, are not suitable to analyze the effects of long-term drift. It has been shown that self-organizing maps (SOM) with residual plasticity can help to maintain the pattern recognition ability of a sensor system affected by drift [55]. The reader is referred to Chapter 13 for a detailed discussion on SOMs. Figure 12.13 shows the identification performances of an electronic nose based on six tin-oxide gas sensors and static and adaptive SOMs. The gases measured

were H2, CO, CO2,CH4, and binary mixtures of H2 with CO and CH4 with CO. It shows that if an adaptive SOM is used, the identification ability of the electronic nose remains almost unchanged when the drift in the sensor response is up to 20 %. However, SOMs with residual plasticity require the frequent measurement of all the patterns. If this requirement is not fulfilled, patterns that seldom occur will be forgotten. Recently, in some preliminary work, adaptive resonance theory (ART) neural networks have been proposed to deal with sensor drift [56]. The short-time memory of the network gives it some plasticity to adapt to sensor drift, while the long-time memory may give the necessary rigidity to avoid forgetting previously learnt patterns. ARTMAP (adaptive resonance theory supervised predictive mapping) and fuzzy ARTMAP are non-parametric, adaptive networks that are well suited to solve pattern classification problems [71, 72]. With other adaptive algorithms, the learning of new events tends to wash away the memory traces of previous, but still useful, knowledge. ARTMAP and fuzzy ARTMAP contain a self-stabilizing memory that permits accumulating knowledge to new events in a non-stationary environment [73]. Very recently, it has been 12.5 Outlook 319

Tab. 12.3 Incremental learning on the three data-sets with Fuzzy ARTMAP, [LVQ] and (MLP). For Fuzzy ARTMAP, the recode rate was fixed to b ¼ 0:1. Number of patterns correctly classified/Total number of patterns in the category. (Reprinted from E. Llobet et al., Meas. Sci. Technol., Vol. 10, pp. 538–548, ª1999 IOP Publishing, with permission.)

Category Performance

Learning/Test sets a b c d e f g (%)

1 / 1 – 21/21 – 7/8 20/20 – – 98.0 [20/21] [8/8] [19/20] [95.9] (20/21) (8/8) (18/20) (93.8) 2 / 1 and 2 10/10 28/29 7/8 16/17 19/20 – – 94.6 [9/10] [22/29] [8/8] [15/17] [18/20] [85.7] (9/10) (27/29) (8/8) (15/17) (0/20) (70.2) 3 / 1, 2 and 3 24/24 28/29 7/8 25/26 33/35 23/24 29/29 96.0 [21/24], [28/29] [8/8] [14/26] [23/35] [23/24] [23/29] [80.0] (23/24) (0/29) (0/8) (25/26) (15/35) (21/24) (27/29) (63.4)

shown that the incremental learning capability of fuzzy ARTMAP is very promising to address drift in electronic nose systems. In particular, the method has been successfully applied to the classification of alcohols and coffees [57], the non-destructive determination of fruit ripeness [58–60] and the classification of bacteria [74, 75]. The reader is referred to Chapter 13 for a detailed discussion on ART networks. During training the ARTa module was supplied with the response vectors of a four-element tin-oxide sensor arrays, while the ARTb module was supplied with the corresponding correct categories. Using fast node commitment and slow node re-code, this network performed incremental learning without forgetting previous knowledge. These results are shown in Table 12.3, where the performance of fuzzy ARTMAP is compared to other neural paradigms, such as multi-layer perceptron (MLP) and learning vector quantization (LVQ). The data were split in three data-sets to perform incremental learning.

12.5 Outlook

There is no universal sensor system that can solve all odor or gas mixture analysis problems. Instead there is a need to employ intelligent application-specific sensor systems that are appropriate to the application. This means building-in intelligence through the development of suitable sensor structures, sensor materials and pattern recognition methods [76]. New pattern recognition methods should make use of the transient information in the sensor signal to enhance the identification ability of the system. This requires the use of dynamic models for the sensor system that can account for the drift in sensor parameters and thus extend the calibration period. 320 12 Dynamic Pattern Recognition Methods and System Identification

The importance of many problems associated with current chemical sensor technology is application specific. If the system has to analyze low levels of low reactive species, sensors tend to perform well. If the system has to analyze high levels of reactive species, poisoning of the sensors is likely and drift effects become very significant. The baseline of sensing devices (for example, metal oxides, polymeric chemoresistors and polymer coated QMB) is sensitive to the operating temperature, the humidity and type of carrier gas [77]. Very often, the sensors require a long recovery time between measurements to reach their baseline. In continuous monitoring or repeated measurement applications, the response of the sensors is influenced by their previous history (short-term memory effect). Under these constraints, the choice of a suitable modeling strategy should be considered carefully:

* Non-adaptive models can be useful when the application implies the analysis of weakly reacting species with systems where temperature and humidity are strictly controlled by the sample delivery system. Drift is likely to be small in such a system. * Adaptive models are required when analysis of strongly reacting species is to be performed and the sensors are likely to drift due to poisoning. These models can also handle drift caused by slight variations in the temperature and humidity of the carrier gas. * Of the non-adaptive models, ad hoc parametric models are interesting because they may give some insight into sensor behavior. The measured parameters can be fed directly into well-established pattern recognition systems. Linear filters and non- linear models can be used to compensate for the short-term drift caused by the memory effect of the array when successive measurements are performed. * The development of non-linear, adaptive models in which competition between component gases occurs may best be solved using neural paradigms.

Fig. 12.14 Selection of a dynamical PARC method for linear or quasi-linear problems 12.5 Outlook 321

Fig. 12.15 Selection of a dynamical PARC method for non-linear problems

* SOMs with residual plasticity can be a good choice when frequent measurements of all the patterns are performed. When this condition is not fulfilled, the ART approach is a promising one. * The on-line incremental learning capability of fuzzy ART is a very promising feature for drift counter-action in electronic nose systems.

These basic ideas are contained in Figs. 12.14 and 12.15, where the suitability of a specific dynamic model to a particular type of problem is shown [78]. In Fig. 12.14 the sensor responses are considered to be linear or quasi-linear in concentration. This is generally the case when the species concentration is low, for example for conducting polymer resistive sensors, or when the concentration range is small and so is, step-wise, approximately linear. If the sensor response is non-linear in concentration in a well-defined manner, a pre-processing linearization algorithm can be used [79]. On the other hand, in Fig. 12.15, the selection assumes that the non-linear part of the sensor response is important and must be accounted for in the models. The first attempts to use the dynamic sensor signals in electronic noses have essentially consisted of the development of ad hoc sensor-oriented parametric models. To develop a new generation of electronic noses, there is a need to extend these models taking into account the effects of environmental variables such as temperature and humidity, and to implement improved adaptive models to counter-act sensor drift and poisoning. 322 12 Dynamic Pattern Recognition Methods and System Identification

References

1 R. E. Cavicchi, J. S. Suehle, K. G. Kreider, 19 S. A. Ivanov, V. N. Ivanova, V. B. Smirnov, M. Gaitan, P. Chaparala. Sensors and B. Z. Tabin. Opt. Spectros., 1992, 73, Actuators B, 1996, 33: p. 142–146. 150–53. 2 E. Llobet, J. Brezmes, X. Vilanova, 20 M. R. Smith, S. Cohn-Sfetcu. Techno- L. Fondevila, X. Correig. Quantitative vapor metrics, 1976, 18, 467–482. analysis using the transient response of non- 21 H. Tong, Non-linear Times Series. A Dyna- selective thick-film tin oxide gas sensors. mical System Approach. 1990, Oxford: In IEEE Transducers’97, 1997, Chicago, USA. Clarendon Press. Chapter 3, 96–120. Vol. 2, 971–974. 22 M. B. Prietsley, State dependent models: 3 F. J. Auerbach, Pattern recognition using a general approach to non-linear time series gas modulation.InTransducers’95 – Euro- analysis. J. Times Ser. Anal., 1980, 1, 57–71. sensors IX. 1995. Stockholm, Sweden. 23 V. Volterra, Theory of Functionals and 4 D. M. Wilson, S. P. Deweerth. Sensors and of Integro-differential equations. 1930, Actuators B, 1995, 28, 123–128. New York, Dover. 5 C. DiNatale, S. Marco, F. Davide, A D’amico. 24 N. Wiener, Non-linear Problems in Random Sensors and Actuators B, 1995, 24–25, Theory. 1958, New York: Wiley. 578–583. 25 H. W. Chen, N. Ishii, N. Suzumura. Int. 6 M. Schweizer, J. Goppert, A. Hierlemann, J. Systems Sci., 1986, 17, 371–377. J. Mitrovics, U. Weimar, W. Rosentiel, 26 H. W. Chen, L. D. Jacobson, J. P. Gaska. Biol. W. Go¨pel. Sensors and Actuators B, 1995, Cybernet., 1990, 63, 341–357. 26–27, 232–236. 27 S. A. Billings, H. B. Jamaluddin, S. Chen. 7 M. Roth, R. Hartlinger, R. Faul, Int. J. Control, 1992, 55, 193–224. H. E. Endres. Sensors and Actuators B, 1996, 28 J. G. Kuschewski, S. Hui, S. H. Zak. IEEE 35–36, 358–-62. Trans. on Control Systems Technology, 8 M. Holmberg, F. Winquist, I. Lundstro¨m, 1993, 1, 37–49. F. Davide, C. DiNatale, A D’amico. Sensors 29 C. L. Giles, G. M. Kuhn, R. J. Williams. and Actuators B, 1996, 35–36, 528–535. IEEE Trans. on Neural Networks, 1994,5, 9 M. Nakamura, I. Sugimoto, H. Kuwano, 153–55. R. Lemos. Sensors and Actuators B, 1994, 30 S. Z. Qin, H. T. Su, T. J. Mcavoy. IEEE Trans. 20, 231–237. on Neural Networks, 1992, 1, 123–130. 10 R.E. Kalman, On the general theory of control 31 L. Ljung, System Identification: Theory for the systems.inFirst IFAC Congress. 1960, User, 1987, Englewood Cliffs, New Jersey: Moscow, Butterworths, London. Prentice Hall. Chapters 6 and 8 to 11. 11 S. Vaihinger, W. Go¨pel, J. R. Stetter. 32 P. J. Diggle, Times Series: A Biostatistical Sensors, and Actuators B, 1991, 4, 337–343. Introduction, 1990, Oxford: Clarendon Press. 12 E. Llobet, X. Vilanova, X. Correig. Chapters 2 and 6. Proceedings SPIE, 1995, 2504, 559–566. 33 Y. W. Lee, M. Schetzen. Int. J. Control, 1965, 13 R. Gutierrez-Osuna, H. Troy Nagle, 2, 237–254. S. S. Schiffman. Sensors Actuators B, 34 H. A. Barker, R. W. Davy. Proceedings IEE, 1999, 61, 170–182. 1975, 122, 305–311. 14 C. Lanczos, Applied Analysis, Prentice-Hall, 35 H. A. Barker, R. W. Davy. Int. J. Control, 1956. 1978, 27, 277–291. 15 D. G. Gardner, J. C. Gardner, G. L. Laush, 36 D. R. Brillinger, Time series: Data Analysis W. W. Meinke, J. Chem Phys.. 1959, 31, and Theory, 1981, San Francisco, CA, 978–86. Holden Day. Chapter 6. 16 S. Marco, J. Samitier, J. R. Morante. Meas. 37 K. Godfrey, Perturbation signals for system Sci. Technol., 1995, 6, 135–142. identification, 1993, Prentice Hall, UK, 17 M. R. Osbourne, SIAM J. Numer. Anal., Chapter 2. 1975, 12, 571–92. 38 H. A. Barker, Nonlinear system identification 18 E. Yeremian, P. Claverie, Nature. 1987, 326, using pseudorandom signals with partially 169–174. orthogonal transforms, in 7th IFAC Symposium 12.5 Outlook 323

on Identification and System Parameter 55 S. Marco, A. Pardo, A. Ortega, J. Samitier. Estimation, 1985, York. Gas identification with tin oxide sensor array 39 R. Haber, IEE Proc. D, 1988, 135, 405–420. and self organizing maps: adaptive correction of 40 S. A. Billings, Springer Lecture Notes in sensor drifts,InIEEE Instrumentation and Control and Information Sciences, 1986, 79, Measurement Technology Conference, 1997, 263–294. Otawa, Canada. 41 S.A. Billings, K. M. Tsang. Mechanical 56 D. S. Vlachos, D. K. Fragoulis, J. N. Avarit- Systems and signal Processing, 1989,2, siotis. Sensors and Actuators B, 1997, 45, 319–339. 223–228. 42 M. J. Koremberg, Identifying noisy cascades 57 E. Llobet, E. L. Hines, J. W. Gardner, of linear and static non-linear systems, in 7th P. N. Bartlett, T. T. Mottram. Sensors and IFAC Symposium on Identification and System Actuators B, 1999, 61, 183–190. Parameter Estimation, 1985, York. 58 E. L. Hines, E. Llobet, J. W. Gardner. 43 D. T. Westwick, R. E. Kearney. Biol. Electronics Letters, 1999, 35, 821–823. Cybernet., 1983, 68, 75–85. 59 E. Llobet, E. L. Hines, J. W. Gardner, 44 N. J. Clayden, J. Chem. Soc. Faraday Trans., S. Franco. Meas. Sci. Technol., 1999, 10, 1992, 88, 2481–2486. 538–548. 45 S. Marco, A. Pardo, F. Davide, C. DiNatale, 60 J. Brezmes, E. Llobet, X. Vilanova, G. Saiz, A. D’amico, J. Hierlemann, M. Schweizer, X. Correig. Sensors and Actuators B, 2000, U. Weimar, W. Go¨pel. Sensors and Actua- in press. tors B, 1996, 34, 213–223. 61 M. Ippommatsu, H. Sasaki, J. Electrochem. 46 M. Nakamura, I. Sugimoto, H. Kuwano. Soc., 1989, 136, 2123–2128. Sensors and Actuators B, 1996, 33, 122–127. 62 H. E. Endres, W. Gottler, H. D. Jander, 47 T. Nakamoto, N. Okazaki, T. Morizumi. S. Drost, G. Sberveglieri, G. Faglia, Sensors and Actuators B, 1997, 41, 183–188. C. Perego. Sensors and Actuators B, 1995, 48 A. Pardo, S. Marco, J. Samitier. Dynamic 24–25, 785–789. measurements with chemical sensor arrays 63 X. Vilanova, E. Llobet, J. E. Sueiras, based on inverse modeling.InIEEE Instru- R. Alcubilla, X. Correig. Sensors and mentation and Measurement Technology Actuators B, 1996, 31, 175–180. Conference, 1996, Brussels, Belgium, 64 E. Llobet, X. Vilanova, J. Brezmes, 904–907. R. Alcubilla, X. Correig. J. Electrochem. Soc, 49 F. Davide, C. DiNatale, A. D’amico, 1998, 145, 1772–1779. A., A. Hierlemann, J. Mitrovics, 65 J. W. Gardner, P. N. Bartlett, E. L. Hines, M. Schweizer, U. Weimar, W. Go¨pel, F. Molinier, T. T. Mottram. IEE Proceedings: S. Marco, A. Pardo. Sensors and Actuators B, Sci., Meas. Technol., 1999, 146, 102–106. 1995, 26–27, 275–285. 66 H. Nanto, K. Kondo, M. Habara, 50 F. Davide, C. DiNatale, A. D’amico, Y. Douguchi, R. I. Waite, H. Nakazumi. A. Hierlemann, J. Mitrovics, M. Schweizer, Sensors and Actuators B, 1996, 35–36, U. Weimar, W. Go¨pel. Sensors and Actua- 183–186. tors B, 1995, 24–25, 830–842. 67 W. M. Sears, K. Colbow, R. Slamka, 51 F. Davide, C. DiNatale, A. D’amico. Sensors F. Consadori. Sensors and Actuators B, and Actuators B, 1994, 18–19, 244–258. 1990, 2, 283–289. 52 C. DiNatale, F. Davide, F., A. D’amico. 68 B. Yea, R. Konishi, A.K. Sugahar, T. Osaki. Sensors and Actuators B, 1995, 26–27, An advanced gas discrimination method utili- 237–241. zing the periodic operation of a semiconductor 53 C. DiNatale, A. Macagnano, A. D’amico, gas sensor. In IEEE Conference on Industrial F. Davide. Meas. Sci. Technol., 1997, Automation and Control, 1995. 1236–43. 69 M. E .H. Amrani, R. M. Dowdeswell, 54 B. W. Saunders, D. V. Thiel, A. Mackay-Sim. P. A. Payne, K. C. Persaud. Sensors and An artificial olfactory system using tiered Actuators B, 1998, 47, 118–124. artificial neural networks. In Australian and 70 P. Bruschi, A. Nannini, B. Neri. Sensors New Zealand Conferrence on Intelligent and Actuators B. 1995, 24–25, 429–432. Information Systems, 1994. 324 12 Dynamic Pattern Recognition Methods and System Identification

71 G. A. Carpenter, S. Grossberg, 76 W. Go¨pel, J. Hesse, J. N. Zemel Eds.. Sensors: J. H. Reynolds. Neural Networks, 1991,4, a comprehensive survey, Vol. 2, 1991, VCH, 565–588. Weinheim. Chapter 6. 72 G. A. Carpenter, S. Grossberg, 77 M. Sriyudthsak, L. Promsong, S. Panyakeow. J. H. Reynolds. IEEE Trans. Neural Sensors and Actuators B, 1993, 13–14, Networks, 1995, 6, 1330–1336. 139–142. 73 G. A. Carpenter, S. Grossberg. Fuzzy sets, 78 E.L. Hines, E. Llobet, J. W. Gardner. IEE neural networks, and soft computing, Proceedings, Circuits Devices and Systems, R.R. Yager, Zadeh, L.A., Editor. 1994, 1999, 146, 297–310. Van Nostrand Reinhold: New York, 79 S. W. Moore, J. W. Gardner, E. L. Hines, 126–165. W. Gopel, U. Weimar. Sensors and 74 H. W. Shin, E. Llobet, J. W. Gardner, Actuators B, 1993, 15–16, 344–348. E. L. Hines. C. S. Dow. IEE Proceedings, Sci. Meas. Technol., 2000, 147, 158–164. 75 P Boilot, E. L. Hines, S. John, J. Mitchell, F. Lopez, J. W. Gardner, E. Llobet, M. Hero, C. Fink, M. A. Gongora. Detection of bacteria causing eye infections using a neural network based electronic nose system, Proceedings of 7th ISOEN, 2000. 325

13 Drift Compensation, Standards, and Calibration Methods

M. Holmberg and T. Artursson

Abstract In Webster’s Seventh New Collegiate Dictionary, drift is defined as “a gradual change in any quantitative characteristic that is supposed to remain constant”. Thus, a drifting chemical sensor does not give exactly the same response even if it is exposed to exactly the same environment for a long time. Drift is a common problem for all chemical sensors, and thus needs to be considered as soon as measurements are made for a long period of time. First in this chapter, possible reasons for drift will be discussed. A distinction is made between drift in the sensors, and drift in the measurement system. After this, typical features of drift as seen in the measurements will be shown. These features include gradual increase or decrease, and jumps in the responses. At the end, many different methods for reducing the effects of drift will be described. These drift reduction methods try to compensate for the changes in sensor performance using mathematical models and thus maintaining the gas identification capability of the electronic nose. Many different methods have been applied for different situations. It is impossible to compare all the methods since each one of them makes some assumptions of how the measurements are made and/or how the drift is manifested. Not all examples discussed are for measurements with electronic noses, but the concepts may easily be transferred also to such applications. The purpose of describing all the methods is to show some possible ways of reasoning when dealing with a data-set from drifting sensors.

13.1 Physical Reasons for Drift and Sensor Poisoning

In this section, some of the common causes of drift in chemical sensors will be described. Other effects giving rise to similar phenomena will also be mentioned. The aim of this section is not to give detailed information of the chemical processes that occur, but only to give a brief introduction to these effects.

Ideally, a chemical sensor will always give the same response when exposed to an identical gas mixture. This will, however, not be true when the sensors are operated over a long time period. There will be changes in the size of the sensor response for a certain amount of a given gas; the selectivity of the sensor may change, i.e. the response changes differently for different gases; the speed of response may also change, see Figs. 13.1 and 13.2. These changes in the sensor behavior together give rise to drift in the sensor responses. Drift has plagued sensor researchers for a long time, but it is not until recent years that methods for reducing its effects have been developed [1, 2]. The response of chemical sensors depends upon chemical or physical interactions between molecules in the gas phase and the sensor surface and/or bulk material. A lot of effort has been made to find sensor materials which interact reversibly with the gas, such that the molecules that have reacted on the sensor will leave it as soon as the gas is

Fig. 13.1 Idealized sensor responses for a chemical sensor. The curve shows the three phases of a measurement: baseline measurement (usually made with pure air), test gas measurement, and recovery time (during which the sensor again is exposed to pure air, the recovery time is usually much longer, but the last part of the curve has been omitted here). Curve a) shows an example of a typical response curve for an arbitrary gas; curve b) shows how the sensor response for the same gas would be if drift has caused the speed of response to decrease; curve c) shows the sensor response to the same gas but where the sensor response has decreased compared to curve a). For a typical measurement, the x-axis shows time in seconds, while the y-axis is in arbitrary units, depending on the sensor type used 13.1 Physical Reasons for Drift and Sensor Poisoning 327

Fig. 13.2 Examples of the sensor response shown as bar charts for one sensor when exposed to ten different gases. a) shows the response as it appears without drift; b) shows the response to the same gases when the response of the sensor has decreased by the same amount for each gas. Note that the pattern is preserved, even though the absolute values change; c) shows the response to the same gases when the response has changed differently for the different gases. This is referred to as a change in the selectivity of the sensor 328 13 Drift Compensation, Standards, and Calibration Methods

no longer present at the sensor surface. In laboratory conditions with well-controlled atmospheres, this may be achieved. However, for “real” environments where a lot of different gases are present (several of them in very small amounts), little can be said beforehand regarding the chemical reactions and their reversibility. Therefore, some reactions will be irreversible, thereby blocking or creating reaction sites on the sensor surface and/or bulk of the sensing material. This will lead to a change in the sensitivity of the sensor towards other gases. Another effect that might occur is the re-organization of the sensing material, for example clustering of metal particles. This may hap- pen spontaneously with large time constants for all materials, but the effect may be speeded up by operating in reactive environments and/or at high temperatures. This aging of the material also changes the number of reaction sites and thus also the sensitivity of the sensor. A time-dependent change in the response to an identical chemical environment will therefore result, and this is how we see drift in our measurements. For different sensor types, different causes for drift will dominate [3–5]. Several papers have been published on work made to improve the long-term stability of gas sensors [6, 7]. However, some regeneration of the sensor may be performed by, for example, annealing of the sensor and thereby removing some of the irreversibly bound species. In the electronic nose concept it is also very important to consider drift in the measurement system. This may be due to temperature variations in the measured head space or on the sensors; reactions of gas species in the sampling system; variations in the gas flow; humidity variations in the sample; ambient pressure variations, or other physical/chemical processes. It is very difficult, or maybe even impossible, to distinguish between sensor drift and drift in the measurement system. It is, however, possible to optimize the system components for each application in order to remove as much of the system drift as possible. This may be done by careful control of the sample and sensor temperatures or by reducing the amount of tubing that the sample gas needs to flow through. In the remainder of this chapter, the effects of sensor drift and system drift will not be separated, but will always together be termed drift. There are also other phenomena that may give similar effects to drift. One that is worth mentioning is memory effects, i.e. that the response of the sensor depends on what it has recently been exposed to. The remnants of previous gases may be present either in the sampling system, or on the sensor surface itself. At the exposure of a test gas, these old remnants give an additional effect to the sensor response. This phenomenon is different from drift since it is a temporary effect that may last only for minutes or hours. For longer time constants, this effect will not be distinguishable from drift. The best way to deal with this phenomenon is not to use drift compensation algorithms, but to improve the measurement procedure, e.g. by limiting the size of the sampling system, or to introduce “cleaning cycles”, i.e. short pulses of clean air and/or high temperature annealing between the samples. Another effect that is often seen is that the sensors need some time before they give a stable response after start-up of a measurement series. This means that the response increases or decreases for the first minutes or hours of operation. This is sometimes called short-term drift, but the nature is different from ordinary drift and it will not be dealt with in this chapter. 13.2 Examples of Sensor Drift 329

13.2 Examples of Sensor Drift

In Fig. 13.3, the sensor responses as a function of time are shown for an experiment made using a gas mixing system and 39 sensors. The responses for the three sensors shown are all for one well-controlled gas mixture (“odor”), but other gas mixtures were also measured in between the measurements shown. In this experiment, all the sensors were freshly made. Since drift influences the sensors strongest in the beginning after their fabrication due to thermal relaxation of the device, the sensors show rather strong drift over the measurement period, which was about two months. There are some features in the graph that often can be seen in long-time measurement series:

* The most obvious feature is an exponential or linear decrease or increase in the sensor signal. This change comes either from changes in the sensitivity of the device, or from changes in the baseline. * There are some jumps in the data set, i.e. places where the sensor signal for no apparent reason suddenly changes value. In this case, the jumps are rather small (a few percent), but when the sensors are put in a more reactive atmosphere, the jumps may be much larger. * There is also some noise superimposed on all the sensor signals.

In real-life applications, the situation may be even more complex since also variations in the samples and/or the sampling system come into play. It is therefore important for all applications to carefully control the samples and the sampling system in order not to make the situation more complex than necessary.

Fig. 13.3 Sensor signals from three different sensors for one gas mixture as a function of time in an experiment using a gas mixture system 330 13 Drift Compensation, Standards, and Calibration Methods

Fig. 13.4 Sensor responses changing over time due to drift as seen in a PCA plot. The data set consists of data from 39 sensors, measuring on 9 different gas mixtures. The percentage shown after the PC number on the axes show how large part of the total variance that is explained by that PC. (Reprinted from J. Chemometrics, 14, T. Artursson, T. Eklo¨v, I. Lundstro¨m, P. Ma˚rtensson, M. Sjo¨stro¨m, and M. Holmberg, Drift corrections for gas sensors using multivariate methods, 711 –724, 2000, with permission from John Wiley & Sons Limited.)

The data as seen in Fig. 13.3 can be said to be univariate, which means that we study one variable (sensor) at the time. In a multi-sensor system it is, however, often convenient to study all sensors collectively using multivariate techniques such as Principal Component Analysis (PCA), see Chapter 6. A PCA gives you a mapping of the data, from the original multivariate space with the number of dimensions equal to the number of sensors, to a low-dimensional space which is much easier to visualize. Usually, the first few principal components are a good approximation of the data set for initial studies. Figure 13.4 shows the same experiment as in Fig. 13.3, but for nine different gas mixtures, and now using a PCA to visualize the data from all 39 sensors. As indicated by the arrows, the drift tends to move the sensor responses mainly in one direction, and the direction is similar for all gas mixtures. The reason why the drift tends to go in only one direction for each cluster is that the sensors used are exposed to the same (but not constant) environment all the time, so they tend to drift in a similar manner. This means that the drift may be described in only a few (in this case one) dimensions even though the process is rather complicated. The reason why all the clusters drift in a similar direction is that the gas mixtures are very similar, so when the sensor changes, this change affects the responses for all gas mixtures in a similar way. In a situation where many different sensor types are used, one cannot assume that the drift will occur in a few dimensions only, but rather that one dimension will be needed to describe the drift for each sensor type. It is also important to note that different gas mixtures might drift in different directions, so when choosing a reference gas for compensating drift, it has to be very similar to the test gases in the application. 13.3 Comparison of Drift and Noise 331

13.3 Comparison of Drift and Noise

In a real measurement series, it may sometimes be necessary to attribute a small change in the sensor response to either a change in the sample; noise in the measurements; or a drift induced change in the sensor response. If only one such change is occurring, this distinction is impossible to make. On the other hand, by analyzing a long time-series of data, much can be learnt about the intrinsic noise in the system. This information may then be used in statistical models to ascertain if the change is due to noise or to changes in the sample. However, sensor drift may change the statistical limits, thereby making the models useless. Some information regarding the frequency spectra of the noise and the drift may be obtained from such studies, but it may be difficult to use this information in practice, since the frequency spectrum of the sensors also may change over time due to drift. Very little can be found in the literature regarding the relationship between drift and noise. There has been one study [8] where a frequency analysis of a long time series was made. In the study, it was assumed that similar sensors drift in a similar manner. The signals from the sensors were passed through band-pass filters, and the correlation between the filtered signals was studied by seeing how well a model could predict

Fig. 13.5 (a) A model f1 is built on instrument 1 using samples representative of all possible measurements in the future. (b) In order to be able to use the information in instrument 1 without remaking all the

measurements used to build f1, some known samples are measured on both instrument 1 and i. Then, one searches for a transformation of

either the model f1 (initially equal to f1), Xi,orYi;1 that renders Yi;ref ¼ Y1;ref . (c) Depending on which transformation was chosen, the data evaluation for the new instrument n is made according to one of the three schemes shown. The first case with a transformation of f is not very simple, and therefore not so common. The second case with a transformation of X is referred to as direct transformation. The third case with a transformation of Y is called bias correction 332 13 Drift Compensation, Standards, and Calibration Methods

the output of a sensor using other sensor signals as inputs. A low prediction error meant that there was a high correlation and vice versa. For the frequency ranges where there was a correlation between the different sensors in the array, drift was said to dominate, while noise dominated in the frequency range where the variations were not correlated between the sensors. The sensor correlation was thus used as an instrument to distinguish the different frequency ranges of drift and noise.

13.4 Model Building Strategies

In general, some model, f, is used to map the measured sensor data, often termed X- data, to some output, Y, which gives us the information we desire, e.g. the class and/or quality of the sample, so Y ¼ f ðXÞ, see Fig. 13.5a. The model could be anything from a simple linear regression to more complex model types such as Artificial Neural Net- works, as described in previous chapters. When we study drift, we need to use further considerations in the model building and the model validation. The first thing to consider is the choice of sensors. Do some of the sensors vary more than the others, or maybe even stop to respond after some time? In that case, it might be wise not to include those sensors in the model building. Also, when a drift reduction method is tested, it is not a good idea to use parts of the data set from the whole time period in the data set for the model building. This could lead to the variations in the data being built into the model rather than actually being reduced by the drift reduction method. Instead, it is wise to build a model of the first measurements, and then apply the drift reduction method to subsequent measurements, thus validating both the model and the drift reduction method.

13.5 Calibration Transfer

The transfer of calibration methods from one instrument to another almost (but not exactly) identical instrument is a problem for many different instrument manufac-

turers. [9, 10] A model, f1, which maps X to Y, is built using measurements on instrument 1, see Fig. 13.5a. The measurements, X, used to build the model must represent all possible situations that may arise during later operation of the instrument in order to get a representative model. Many measurements are therefore necessary to build this model, the exact number depends on the instrument and model complexity but ranges from a few tens to several thousand. For all other presumably identical instruments manufactured, we want to avoid making all the measurements again since measurements in general are time-consuming and expensive. So, the aim is to trans-

fer the information contained in f1 (the model built using measurements made with instrument 1) to a model for instrument i (fi ) using as few measurements as possible. If the instruments were identical, there would be no need to make any new measure- 13.6 Drift Compensation 333 ments, since the models f1 and fi would be equal. If the instruments were completely different, there exists no common information for the two instruments, and thus all possible situations have to be measured also on instrument i. If we assume that the instruments are similar, but not equal, we can make a few new measurements on a third instrument n, and then assume that other measurements in similar environments have changed in a similar manner for all the instruments. We can then reduce the number of measurements necessary to build the new model, fi . If the sensors in an electronic nose have drifted slightly, this can be seen as having one instrument at time t1, and another slightly different instrument at a later time t. The concepts for calibration transfer and for drift reduction are therefore similar, even though the problem is different.

Mathematically speaking, the aim is to approximate a function, fi capable of mapping Xi to Yi, by using a low number of measurements, Xi, and another function, f1,as a first approximation:

fi ¼ Tðf1jYi ¼ fi ðXiÞÞ ð13:1Þ where T is a transformation operator, different for different calibration transfer methods, see Fig. 13.5b. For a complete description of possible transformations, see the references mentioned above. The model f is often changed by either pre-processing of the X-data, or post-processing of the Y-data, see Fig. 13.5c. In the second case in the figure, often called direct transformation, a relationship between the X-values for instrument 1 and i is calculated using some known samples. It is important that these samples are chosen so that they span as large part of the response space as possible in order to find a representative transformation for all possible X-values. The relationship is then used to transform the X-values obtained for instrument i to the same situation as for instrument 1. The originally built models on instrument 1 can then be used to predict Y also for instrument i. The third case in the figure, where the Y-values are corrected, is termed bias correction since it is assumed that the error between the measurements can be seen as a bias in the predicted Y-values. For the ith instrument, the original model for predicting Y from X is used, and a model to correct the predicted Y is built using some references. The Y-correction is then used for all subsequent measurements, making it possible to use the models for instrument 1 also for instrument i.

13.6 Drift Compensation

In order to get an estimation of the size of the drift, measurements are often made on one or several so-called reference gas (or gases). Measurements are made on the reference gas(es) in the beginning of the measurement series (time t1), and then with some intervals (usually a few times per week) as long as the sensors are used. The change in the sensor responses to the reference gas is taken as a measure of the re- 334 13 Drift Compensation, Standards, and Calibration Methods

Fig. 13.6 A diagram showing the difference between univariate, (a), and multivariate, (b), drift correction

sponse change for all other measurements, with different assumptions for different methods as described below. In order to get a good estimate of the drift for the real samples, the reference gas has to be well-chosen, meaning that the drift in the reference gas should reflect the drift for all other samples. Different researchers have chosen different approaches to find good reference gases; some use the same reference gas (often water) for all applications, while others choose the reference gas depending on the application (e.g. the head-space of a given concentration of ethanol in water for determination of the intoxication level with breath analysis). Usually, it is a good idea to choose a reference gas that is close to the real samples in sensor response space (as can be seen in a PCA score plot). A good reference gas also has to be stable over time (not degrade) and be easy to measure so that the variation in gas concentration over time becomes minimal. It can sometimes be helpful to categorize the different methods used for drift compensation. One such distinction is if the sensors are considered one at the time, or as a group. The first case means that the sensors are considered to operate independently of each other, which is called a univariate approach. In this case, one drift correction model is made for each of the sensors. In the second case, one drift correction model is made for a group (often all) of sensors. This is called a multivariate approach, see Fig. 13.6. Another way of distinguishing between different methods is to see where the adaptation due to drift is made. Basically, there are three strategies for compensating drift in the sensors: direct transformation (adaptation of X), bias correction (adaptation of Y), or the use of self-adapting models (adaptation of f). After the strategy has been chosen, one has to decide what model type to use for calculating the compensation, and what assumptions and information to use to build the compensation models. Many different solutions have been tested, and in the following sections we will 13.6 Drift Compensation 335 give an overview of some attempts that have been made. It is important to remember that a method that works well in one situation does not necessarily work in all other applications, so it is important to study the data to find what restrictions and possibilities that you have in your own data set before trying out a new method. As discussed previously, drift can be manifested in several ways. If drift causes the baseline of the sensor to change, the response will be increased or decreased by the same amount, so the drift will be additive. By measurement on a reference gas, the amount of change can be calculated and used for all measurements on the samples. If the drift causes the sensitivity of the sensor to change instead, the drift is termed as multiplicative, that is the response is increased or decreased by some factor. A reference gas can then be used, as for the additive case, to calculate the correction factor. These two corrections will be exact if the sensors are linear, but they will also work well as a first approximation for non-linear sensors. After applying a drift reduction method, it is a good idea to also check its effectiveness by comparing the prediction capabilities of the classification/quantification model with and without drift reduction. If the change in prediction error is not statistically significant, then the method should not be used since the increase in model complexity introduced by adding an extra algorithm might compromise the overall performance for future measurements. When the sensor array is used for quantification of one or several gases, the relative change in the RMSEP value can be used as a performance measure of the drift reduction method: sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi! sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi! with drift reduction without drift reduction Pm Pm 1 true pred 2 1 true pred 2 m ðyi yi Þ m ðyi yi Þ i¼1 i¼1 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi! without drift reduction Pm 1 true pred 2 m ðyi yi Þ i¼1 ð13:2Þ

true pred where m is the number of measurements, and yi and yi are the true and predicted quantification value, respectively. For the case where only classification is desired, the relative change of the Mahalanobis distance between the different clusters before and after drift reduction is a good measure of the effectiveness of the method [11]. It is also possible to obtain a measure of the performance of the drift reduction method simply by comparing the classification rate before and after drift reduction. The comparison can in this case be made with a simple k-nearest-neighbor classifier or other standard classification methods if desired.

13.6.1 Reference Gas Methods

It can sometimes be a problem to distinguish between sensor drift and changes in the sample. To separate drift from sample changes, a stable sample called a reference gas is often measured. Different approaches to correct gas sensor data using a reference 336 13 Drift Compensation, Standards, and Calibration Methods

gas as a reference value and then correcting all subsequent readings accordingly have been made. Five different examples will be given where a reference gas has been used to reduce the drift. All five examples make the assumption that there is a strong correlation between the drift in the response of the sensors to the reference gas and to the samples. After this, a sixth method is presented, where replicates of the samples are used as pseudo-reference gases. The first four methods and the sixth use direct transformation of the data, while the fifth method uses the bias correction procedure. Re- garding the data treatment, the first two methods and the last work in the univariate mode, while the third and fourth work in the multivariate mode. The fifth method only corrects one Y-variable, so here it is not meaningful to use these terms. The univariate methods assume that the changes in the relationship between the response to the reference gas and the response to the test gas can be compensated for one sensor at the time. The first method used by Fryder et al. [12] assumed that the drift was additive, i.e independent of the signal level. They reduced the drift in the measurements made with an electronic nose by subtracting the response to the reference gas from the sample responses, see Eq. (3).

0 xt;i ¼ xsample;t;i xreference;t;i ð13:3Þ

0 where xt;i, xsample;t;i, and xreference;t;i are the drift corrected sensor response, the uncorrected sensor response, and the response for the measured reference gas, respectively, all measured on sensor i at time t. The additive drift was removed and all the measurements were studied relative to the reference gas. The second example is closely related to the first one, but instead of reducing only the additive drift, the method corrects for multiplicative drift effects for measurements made within the same day see Haugen et al [13]. This method was successfully used in their experiments to reduce drift from fresh fish measurements with an electronic nose, measured over five days. The ratio between the responses at time t and at the initial time for the reference gas was calculated for each sensor (see Eq. (4)), and this ratio was used to compensate the responses for the samples. In this specific work, a linear trend line was also fitted to this ratio to find the correction factor for the sample measurements made between the reference gas measurements, see Eq. (5)

quote ¼ x =x ð13:4Þ reference;t1;i reference;t;i x 0 ¼ x f ¼ x ða t þ bÞ ð13:5Þ t;i sample;t;i t;i sample;t;i i

where ft;i is the trend line, a is the slope of the trend line, b is the intercept, t1 is the initial time, and i is the sensor number. An additive correction was used for correction between the different days. The third method is a drift reduction method based on PCA and PLS, called component correction, CC [14]. The method assumes that the drift has a preferred direction in the measurement space and removes this direction from the measurements. The direction of the drift, p, is calculated from measurements of a reference gas. If the sensor responses to the reference gas have significant drift, the first components, p,in 13.6 Drift Compensation 337

Fig. 13.7 A vector diagram showing the projection of the sample i down to the drift vector p1.

The projected value ti is the amount of drift for sample i

a PCA analysis of this gas will describe the direction of the drift. The vector p comprises the direction coefficients of the one dimensional principal component space, and can be used also to see which parameters contain the most drift. Projecting the sample gas measurements on this vector gives a score vector, t, which contains the amount of drift for each sample, see Fig. 13.7. The drift component, tpT, can then be removed from the sample gas data. The direction in the data set that is removed is a linear approximation of the drift direction. By removing this direction all the other directions are preserved and the important variances that separate different clusters and concentrations are maintained in the data set, unless the information is found in the same direction as the drift. This method was applied to data sets from measurements using a electronic nose and a gas-mixing system with mixtures of four different gases (hydrogen, ammonia, ethanol, and ethene) both for classification and quantification over a period of more than two months. The results for the data seen in Fig. 13.4 are shown in Fig. 13.8. A similar method, but based on canonical correlation analysis has been proposed by Gutierrez-Osuna, who used metal-oxide sensors for measurements on spices over a period of three months. [15] The fourth method uses several reference gases in a transformation model, linear or non-linear. A prediction model to predict Y from X is built at time t1, i.e. in the beginning of the measurement series. In later measurements a drift reduction model is built with the reference gas measurements at time t, t > t1, as inputs and the reference gas measurements at time 1 as outputs. This model should then be able to transform X- data at time t to the value they should have had at time t1, i.e. when no drift had afflicted the sensors. All the other data is then transformed using the drift reduction model. After this pre-processing the original identification model is then used to predict Y. 338 13 Drift Compensation, Standards, and Calibration Methods

Fig. 13.8 The data from Fig. 13.2 with drift reduced by the Component Correction (CC-) method. (Reprinted from J. Chemometrics, 14, T. Artursson, T. Eklo¨v, I. Lundstro¨m, P. Ma˚rtensson, M. Sjo¨stro¨m, and M. Holmberg, Drift corrections for gas sensors using multivariate methods, 711 –724, 2000, with permission from John Wiley & Sons Limited.)

Goodacre et al. [16] used artificial neural networks (ANN) both for the prediction model and the drift reduction model for data from pyrolysis mass spectrometry used to predict bacteria concentration with good results. The fifth method uses the bias correction procedure, i.e. the original model to predict Y is used for all data, but the predicted value is then corrected by a factor, which is calculated from measurements on reference gases, see Fig. 13.5c. In this case, the aim was to measure the alcohol content in breath samples from intoxicated persons using an electronic nose. [17] An ANN model was built to predict the alcohol content, using a gas chromatograph as a reference instrument. A reference sample with 109 mol-ppm EtOH in technical air was also included in the measurements. The measurements with the highest and lowest EtOH concentrations in the test set were also used as reference samples. A linear regression model between predicted and measured EtOH concentration in the reference samples were calculated. From this linear regression model correction factors such as slope and intercept were calculated and used for correction of the Y data in the test set. It may sometimes be difficult to find good reference gases for the measurements. If that is the case, it is possible to use replicates of the samples as pseudo-reference gases. This can be done as long as the samples are stable over time or reliable standardized samples are available. This has been done by Salit et al. [18], who used both an additive and a multiplicative drift correction algorithm with replicates of the samples as reference values for measurements made with inductively coupled plasma-optical emission spectroscopy. The signals, which suffered from additive and/or multiplicative drift, 13.6 Drift Compensation 339 were defined as the sum of the true value (xtruth), the drift influence (edriftðtÞ), and the noise (enoise), see Eqs. (6) and (7).

xmeasured ¼ xtruth þ edrift þ enoise ð13:6Þ

xmeasured ¼ xtruth ð1 þ edrift þ enoiseÞð13:7Þ

The aim was then to find the drift influence and remove it from the data. Instead of spending time frequently measuring standards they measured replicates of the samples, which is common when precise analytical results are wanted. For each sample the mean of the individual signals, xmean, is used as an estimate of xtruth. The drift and noise contribution was calculated as the difference between the measured sample and the estimate of the true value, xtruth. If the drift was additive a smooth curve was fitted to the deviation values. This curve was assumed to describe the drift, edriftðtÞ, and the residuals to the curve were defined as noise, enoise. For multiplicative drift, the deviations emeasured=xmean were fitted to a smooth curve. By these definitions it was then possible to reduce the drift. For additive drift, these corrections are directly predicted from the function edriftðtÞ, and for multiplicative drift the correction is xtruth ð1 edriftðtÞÞ. The use of replicates from all the samples, instead of replicates from one standard, reduced the uncertainty in the drift corrections.

13.6.2 Modeling of Sensor Behavior

The most exact drift counteraction model would probably be a physical one, where all physical changes of the sensor are modeled and accounted for. However, this type of model is very hard to make general for gas sensors. In a well-controlled system with very few gases it would be possible to know what reactions might occur, and thus to describe the drift with a physical model. The problem comes in a real application when the system is not so well controlled, and there are a lot of different gases and combinations of these. For the pH ISFET sensor measuring in liquid, a physical model for different pH can be made [19]. The origin of drift for these sensors is a chemical modification of the insulator surface, which is covered by a hydrated layer. The variation in thickness of this hydrated layer changes the capacitance of the insulator and thereby causes drift. By considering the correlation between the layer and its limiting factor for transport of water related species to the insulator, a model is built describing the drift. The model describes the drift behavior for Si3N4-gate and Al2O3-gate pH ISFETs measured in 0.1 M KCl solution. Another way of modeling the sensor behavior is not to consider the reactions that occur, but to study how the sensors behave in their operating conditions, and then assume that the sensors will always behave in the same manner when they are exposed to the same environment. This would be a mathematical model rather than a physical, but could still be useful in situations where the environment causes drift in the sensors, but does not change much over time. It can then be assumed that the 340 13 Drift Compensation, Standards, and Calibration Methods

sensors follow a certain mathematical curve over time. This requires well-controlled measurements as in the case for modeling of sensor behavior. Pearce et al. [20] used a linear fit to compensate for base-line drift in measurements with an electronic nose on beer over a period of 12 days. The base-line value was measured for each sensor, and a linear fit was used to determine the base-line drift and compensating for additive drift.

13.6.3 Pattern-Oriented Techniques for Classification

When a measurement is made, the responses of all sensors are measured. These responses can be said to form a pattern, imagine for example plotting the responses in a histogram that gives a pattern of bars, one for each sensor. We may then assume that each class in a classification problem has a typical pattern, preserved over time. The relative relationship between different sensors rather than their absolute outputs con- serves the pattern, see Fig. 13.9. If the relative relationships stay constant over time, a simple normalization (e.g. by setting one sensor to always have the value one and scaling the others accordingly) would do the trick. In reality things are not that simple. Noise and different amounts of drift for different sensors make it necessary to use other tricks to see if the pattern is conserved. By studying the steady state or the transient behavior of sensors in an electronic nose, and threshold the values into a binary output, Wilson et al. [21] managed to discriminate between different chemicals. The output voltage from an array of ten

Fig. 13.9 A diagram showing a constant relative relationship between different sensors, for time t ¼ 1andt ¼ 2 13.6 Drift Compensation 341

Fig. 13.10 Binary response pattern for (a) acetone, (b) ethanol, (c) hexane, (d) isopropyl alcohol, (e) methanol and (f) carbon monoxide. (Reprinted from Sens. Actuators B, 28, D. M. Wilson, S.P. De- Weerth, Odor discrimation using steady-state and transient characteristics of tin-oxide sensors, 123– 128, 1995, with permission from Elsevier Science.)

tin-oxide sensors were arranged in ascending order and the output from each sensor was set to zero if it was smaller than the median output, and set to one if it was larger than the median. The resulting output from the threshold function was a pattern of

Fig. 13.11 The VLSI circuitry for the winner-take-all signal processing. The output voltage is fed into the winner-take-all, WTA, and loser-take- all, LTA cells. Here, the WTA output is compared with its neighbor’s outputs giving slope left and slope right as outputs. (Reprinted from Sens. Actuators B, 26 –27, D. Bednarczyk, S.P. DeWeerth, Smart chemical sensing arrays using tin oxide sensors and analog winner-take-all signal processing, 271–274, 1995, with permission from Elsevier Science.) 342 13 Drift Compensation, Standards, and Calibration Methods

zeros and ones, see Fig. 13.10. This way of thresholding the signal removes much of the information, but the information which is left was sufficient to discriminate between acetone, ethanol, hexane, isopropanol, methanol and carbon monoxide. The resulting pattern is more robust than the use of absolute sensor values since it is relative, and it is not very sensitive to noise. The drawback with the method is that it also adjusts for changing concentration levels and is therefore useful only for classification purposes. Bednarczyk et al. [22] worked with a sensor array of ten tin oxide sensors. From this array they located the sensors with the highest (winner), and smallest (loser) outputs voltage for each sample, and also the slope between the winner/loser and its two nearest neighbors was calculated, see Fig. 13.11. This gives a total of six outputs from each sample giving a specific pattern that changed little over time. The winner and loser were used for a rough classification, and the slopes were used for finer classification. For example, ethanol is first classified as an alcohol from its winning and losing sensor, and after that as being ethanol from the values of the slopes. All calculation was processed in VLSI circuitry. The fact that the pattern of the responses rather than the absolute values was used allowed a robust chemical discrimination to be made. Another approach for using the pattern of responses rather than individual sensor responses was made by Holmberg et al. [23] In this case, four different alcohols and water were measured over a period of two months with large drift in the responses. It was assumed that the pattern was preserved over time for each class (i.e. the different alcohols or water), but only for sensors that were similar enough. For that reason, a small subset of three sensors was chosen for the model building. Then, for each class a model was built to predict the output of one sensor, using two other sensors as inputs to the model. These models were different for the different alcohols. When a new

Fig. 13.12 A block diagram describing the routines of training and prediction of sensor response, where time-invariant relationships between the sensor responses are used to reduce the influence of drift 13.6 Drift Compensation 343 measurement was made, the sensor responses were put into all the different models, and the new sample was identified as belonging to the class whose model gave the lowest prediction error. The approach was also improved by allowing updating of the models to adjust for possible changes in the relationships, see Fig. 13.12 [24]. How- ever, also in this case the models are insensitive to variations in concentration, and can thus only be used for classification.

13.6.4 Drift-Free Parameters

Another drift counteraction approach is to find parameters in the measurements that remain constant even though the responses changes. Roth and co-workers [25] used this approach to measure CO2 with gas sensors with organic coatings. They used an appropriate temperature profile in order to decrease the time the sensors were heated. This improved the overall lifetime of the sensor coatings. To further reduce the influence of drift they used the normalized response slope, instead of using the drift sensitive absolute values. The slope of the sensor signals was normalized with the overall amplitude of the signal, in this way the drift sensitivity of the parameters used was reduced, see Eq. (8).

slope slope slope ¼ ¼ ð19:8Þ norm amplitude max min where slopenorm is the normalized slope and max and min are the maximum and minimum response values. Effects like aging and poisoning, which alter the baseline, did not affect the calculated parameters.

13.6.5 Self-Adapting Models

Models that are adjusted on-line are usually called adaptive. This kind of model is useful if the process studied has large variations. Adaptive modeling can be used for both linear and non-linear models. Davide and co-workers [26] introduced an adaptive Self-Organizing Map (SOM) to reduce the influence of sensor drift. The basic idea was to follow the odor pattern that suffered from drift. For a SOM, different neurons are assigned to different classes in the model building process, see Fig. 13.13. When the SOM is used it learns in real-time by continuously moving the nearest neurons towards the input data, by adjustments of the weights. As the sensor responses change, so do the neurons and classification can thus still be made. In this way discrimination between different odors was possible. However, if one of the patterns is not measured for a long time, its neurons will be influenced by measurements of other classes and moved in an undesired way. One approach to avoid this was proposed by Distante et al. [27], who let each class be described by one SOM, thereby avoiding the confusion that might arise if the classes are encountered with different frequency. 344 13 Drift Compensation, Standards, and Calibration Methods

Fig. 13.13 Interpretation of the SOM, where each square symbolizes a neuron. Three distinct classes are visible: a, b and c

Another adaptive network, which can be used to reduce the effects of drift is Adap- tive Resonance Theory, (ART) see Chapter 16 for details. This kind of network has been used for classifications of odors subjected to drift in the chemical sensors [28]. ART networks have the ability to learn a new pattern in real-time and update the prototype vectors describing the different classes. The algorithm finds the prototype vector closest to the sample, and if the degree of match is higher than a threshold value the weights of the prototype vector are refined. If the degree of match is lower than the threshold value a new pattern will be created. A key to reliable results is to find a good threshold value, so the number of learned pattern becomes the right one. Vla- chos et al. [29] compared ART with a back-propagation neural network and showed that the probability to get a successful answer increased with the ART. Another way to use ANNs for drift reduction has been published by Smits et al. [30], where they have used signals (not necessarily from electronic noses) that change over time. They simulated drifting data and then compared the classification performance for an ANN with uncorrected data; with data corrected for additive drift; and with uncorrected data, but with an extra input to the ANN describing the amount of drift. Their results indicate that the last strategy gives the best result.

13.7 Conclusions

Drift is a common problem for electronic noses due to the varying and often reactive environment they are used in. The reasons for the drift vary, and stem from both the sensors and from the measurement system. Usually, the drift has a rather low frequency (the variations occur on the order of days), but it may be different when the sensors are fresh, or the environment contains aggressive gases. In order to reduce drift in the best possible way, measurements have to be made over a long time period so that the drift effects can be studied. It is important to establish that there is a drift, because if there is no drift, drift counteraction methods should of course not be used. If drift exists, the next step is to find trends and/or correlations in the data set that can be used as a drift reduction method. No drift reduction method has been found to be superior to the others in all different types of situations, so it may be necessary to use different algorithms for different applications. It is also important to remember that different methods put different requirements on the data set, such as a calibration gas is necessary; quantification of different gases is required; or the envir- 13.7 Conclusions 345 onment stays almost constant so the drift may be modeled. If the application allows the use of a reference gas it should be used, since it gives the user reliable information of the amount of drift in the sensor system. Furthermore, the use of a reference gas gives user the possibility to discriminate between sensor drift and changes in the sample over the time. The pattern-oriented techniques are attractive since they give stable results, but the drawback is that they are only useful for classification purposes. Both pattern-oriented techniques and the self-adapting approaches may fail when both the sample and the sensor system change over time, since they do not discriminate between drift and sample changes over time. In order to model the sensor behavior the systems need to be very well controlled, and these models are therefore hard to use in real applications. In any case, it is important to understand the method that is used. A good drift reduction method that is used in the wrong way may give confusing results, and therefore be more harmful than helpful.

Acknowledgements The authors would like to thank all colleagues that have contributed with valuable discussions and comments during this work, but we owe a special gratitude to Tomas Eklo¨v, David Lindgren, Fabrizio Davide, and Ingemar Lundstro¨m for their support and feedback.

References

1 W. Go¨pel, K.-D. Schierbaum. In Chemical 6 U. Schoneberg, H. G. Dura, B. J. Hosticka, and biochemical sensors, part I, Vol. 2 W. Mokwa. 1991 International Conference (Ed. W. Go¨pel, T.A. Jones, M. Kleitz, I. on Solid-State Sensors and Actuators, Lundstro¨m and T. Seiyama), VCH Verlags- San Francisco, USA, 1991. gesellschaft, Weinheim, Germany, 1992, 7 K. Dobos, R. Strotman, G. Zimmer. Sensors pp. 1–28. and Actuators, 1983, 4, 593–598. 2 J. W. Gardner, P. N. Bartlett. Electronic Noses 8 F. A. M. Davide, C. Di Natale, M. Holmberg, – Principles and Applications, Oxford Science F. Winquist. In Proceedings of 1st Italian Publications, 1999, 126–128 and 178–179. conference on sensors and microsystems 3 I. Lundstro¨m, A. van den Berg, B. H. van der (Ed. C. Di Natale and A. D’Amico), World Schoot, H. H. van den Vlekkert, Scientific, Singapore, 1996, pp. 150–154. M. Armgarth, C. I. Nylander. In Chemical 9 O. E. de Noord. Chemometrics and Intelligent and biochemical sensors, part I, Vol. 2 Laboratory Systems, 1994, 25, 85–97. (Ed. W. Go¨pel, T.A. Jones, M. Kleitz, I. 10 Y. Wang, D. J. Veltkamp, B. R. Kowalski. Lundstro¨m and T. Seiyama), VCH Verlags- Analytical Chemistry, 1991, 63, 2750–2756. gesellschaft, Weinheim, Germany, 1992, 11 P. Spangeus, D. Lindgren. Submitted to pp. 493–494 and 516–519 and references IEEE Sensors Journal. therein. 12 M. Fryder, M. Holmberg, F. Winquist I. 4 C. Caliendo, E. Verona, A D’Amico. In Lundstro¨m. In Proceedings of Transducers ’95 Gas Sensors (Ed. G. Sberveglieri), Kluwer and Eurosensors IX, Stockholm, Sweden, Academic Publishers, The Netherlands, 1995, 683–686. 1992, p. 281–306. 13 J.-E. Haugen, O. Tomic, K. Kvaal. Analytica 5 D. Kohl. In Gas Sensors (Ed. G. Sberveglieri), Chimica Acta, 2000, 407, 23–39. Kluwer Academic Publishers, The Nether- 14 T. Artursson, T. Eklo¨v, I. Lundstro¨m, lands, 1992,p.43–88 P. Martensson, M. Sjo¨stro¨m, M. Holmberg. Journal of Chemometrics, 2000, 14, 711–723. 346 13 Drift Compensation, Standards, and Calibration Methods

15 R. Gutierrez-Osuna. In ISOEN 2000 ab- 24 M. Holmberg, F. A. M. Davide, C. Di Natale, stracts (Ed. J. W. Gardner and K.C. Persaud), A. D’Amico, F. Winquist, I. Lundstro¨m. European Chemoreception Research Orga- Sensors and Actuators B, 1997, 42, 185–194. nisation, Brighton, UK, 2000, 137–138. 25 M. Roth, R. Hartinger, R. Faul, H.-E. Endres. 16 R. Goodacre, D. Kell. Analytical Chemistry, Sensors and Actuators B, 1996, 35–36, 1996, 68, 271–280. 358–362. 17 N. Paulsson, F. Winquist. Submitted to 26 F. A. M. Davide, C. Di Natale, A. D’Amico. Measurement Science and Technology. Sensors and Actuators B, 1994, 18–19, 18 M. L. Salit, G. C. Turk. Analytical Chemistry, 244–258. 1998, 70, 3184–3190. 27 C. Distante, T. Artursson, P. Siciliano, 19 S. Jamasb, S. Collins, R. L. Smith. Sensors M. Holmberg, I. Lundstro¨m. In Olfaction and Actuators B, 1998, 49, 146–155. and Electronic Noses 2, 2000 (Ed. J. W. 20 T. Pearce, J. W. Gardner. Analyst, 1998, 123, Gardner and K. C. Persaud), The Institute 2057–2066. of Physics, 2000. 21 D. M. Wilson, S. P. DeWeerth. Sensors 28 J. W. Gardner, E. L. Hines, C. Pang. and Actuators B, 1995, 28, 123–128. Measurement þ Control, 1996, 29, 172–178. 22 D. Bednarczyk, S. P. DeWeerth. Sensors 29 D. S. Vlachos, D. K. Fragoulis, and Actuators B, 1995, 26–27, 271–274. J. N. Avaritsiotis. Sensors and Actuators B, 23 M. Holmberg, F. Winquist, I. Lundstro¨m, 1997, 45, 223–228. F. A. M. Davide, C. Di Natale, A. D’Amico. 30 J. R. M. Smits, W. J. Melssen, Sensors and Actuators B, 1996, 35–36, M. W. J. Derksen, G. Kateman. Analytica 528–535. Chimica Acta, 1993, 284, 91–105. 347

14 Chemical Sensor Array Optimization: Geometric and Information Theoretic Approaches

Tim C. Pearce, Manuel A. Sa´nchez-Montan˜e´s

Electronic nose technology – which exploits arrays of broadly-tuned chemical sensors – has matured to the point where it is routinely applied to the quality control of a wide range of commercial products, such as foods, beverages, and cosmetics. Even though a large number of companies exist that design, implement, and sell this technology, the issue of how a practical system is configured and optimized to a particular application domain is, at best, carried out using heuristic methods, or more often, completely ignored. The key theme of this chapter is how the selection of different chemical sensors is crucial to the overall system performance of these analytical instruments. By taking a geometric approach combined with simple linear algebra analysis, we demonstrate how the ‘tunings’ of individual sensors affect the overall performance. New performance measures based on information theory are defined here that should be adopted for optimizing the performance of electronic nose systems.

14.1 The Need for Array Performance Definition and Optimization

Electronic nose instruments are used today for a very wide range of detection tasks from quality control of various food products to medical diagnosis. Clearly, each detection task requires sensitivity in the instrument to a number of different chemical compounds, which are likely to be very different from application to application. Over 10 000 odorous compounds are known to exist in nature, but only a handful of these are likely to be important in solving any discrimination task. The concept of a universal electronic nose instrument, able to solve all odor detection problems, is unlikely to become a commercial reality, particularly because creating sensor diversity within an instrument is expensive and most instruments are dedicated to a very restricted range of detection tasks. In practice, the entire instrument, from sample delivery to sensor array, signal processing and classifier stages, is usually optimized to a particular problem domain in order to provide suitable sensing performance. The optimization of signal processing, classifier, and sample preparation are dealt with elsewhere in this book.

In this chapter we consider exclusively the problem of tailoring a chemosensor array to a particular detection task. One approach might be to augment an existing array by adding sensors appropriate to the new task, but this is an expensive and wasteful solution. Most systems have a limited number of channels and, as we shall see, more sensors does not guarantee improved performance due to noise considerations – for example, when combined with a practical classifier adding a sensor with close to zero sensitivity to the compounds of interest but with significant noise will potentially degrade the performance of the array as a whole. In practical terms, optimization of chemosensors within an electronic nose instrument usually means selecting between a potentially large pool of different sensors (even comprising completely different sensing technologies). The optimization task is to select a combination of sensors best suited to the detection task, and ideally to be able to specify a detection limit for each compound of interest. Electronic nose instruments rely on a range of broadly tuned chemosensors in order to discriminate complex multicomponent odor stimuli. It is the pattern of response across the array that is used in discriminating between complex (multicomponent) odor stimuli. This sensing arrangement makes the question of detection performance definition and optimization non-trivial, because it is not usually possible to account for the sensitivity of the system to any one odor component in terms of any single chemosensor within an array. In the converse case, where a set of highly specific sensors each responding uniquely to a single component of the stimulus, optimization would be straightforward, because the signals from sensors responding weakly to the components of interest should be amplified, and those responding to interfering or unimportant components should be attenuated or ignored entirely. Furthermore, the detection performance would be simple to quantify, because the detection of the system for each compound would be uniquely defined by the signal-to-noise performance of the underlying sensor. The need for chemical sensor array optimization becomes obvious when we observe that one set of chemosensors used to solve a given problem may be poor at solving another, new detection problem. This is especially true for small array sizes where sensor diversity is limited, and sensor choice is more critical. But what properties of the array make the difference between it being suited to a particular detection problem or not? Clearly, before we can address the issue of performance optimization we must develop a rigorous framework for describing the criteria affecting the ability of an array to solve the problem. The inability of an array to solve a defined detection or discrimination task might result from one or more of four key factors:

1. There is insufficient sensitivity in any of the sensors within the array to the key compounds of interest at the concentration levels required to solve the new task. 2. Those sensors sensitive to the key compounds relevant to the new task are too noisy to yield sufficient information to solve the task. 3. The array response to a repeated and identical stimulus is not sufficiently reproducible to permit discrimination between similar stimuli. 4.2 Historical Perspective 349

4. There is insufficient sensor diversity within the array to discriminate between key compounds relevant to the new task. We will refer to this as the ‘tuning’ of the array.

Issues one and two are very closely related because ultimately sensitivity is limited by noise, therefore the real parameter of interest here is the signal-to-noise ratio. Issue three can be considered as a special case of issue two, because sensor-response reproducibility can be quantified probabilistically in a similar way to noise. So we see that, in general, the problem reduces to two basic issues, sensor noise (where we might choose to include sensor response reproducibility information) and sensor array tuning. Any comprehensive scheme for performance definition or optimization of chemosensor arrays needs to take both these aspects into account.

14.2 Historical Perspective

Zaromb and Stetter recognized very early the need to quantify sensor-array performance [1]. In 1984 they considered the case of using an array of non-specific chemical sensors for multicomponent gas analysis: a problem closely related to describing complex odors. By first assuming that the response of each sensor was binary to each stimulus (response vs. no response) they argued for a combinatorial measure of the number of sensors required to detect a given number of chemical species

XA m! 2n 1 ; ð14:1Þ ð Þ! ! i¼1 m i i where n is the number of sensors within the array, m is the number of different chemical compounds to be detected, and A is the maximum number of compounds (A m) appearing as a mixture at any one time. This inequality provided a lower bound on the number of sensors required to solve a particular sensing task. For example, according to Eq. (14.1) more than 18 sensors (n 18) would be required to detect a tertiary mixture (A ¼ 3) taken from 100 single chemical compounds (m ¼ 100). Because the derivation of Eq. (14.1) was made on the basis of each sensor responding in a binary fashion to the stimulus, this severely limits the information provided by each sensor and so the inequality produces a gross overestimate of the actual number of sensors required to solve a particular problem – in practice the bandwidth within the system is far higher than suggested here. This limitation can be partially overcome by considering each sensor to respond in an n-ary fashion by splitting the full-scale sensor range into p discrete domains and so the left hand side of Eq. (14.1) becomes (pn 1), yielding a more realistic estimate of the number of sensors required to solve a particular task. A more severe limitation, however, is the lack of any account of noise or sensor reproducibility in their analysis. This becomes obvious when considering that the bound given by Eq. (14.1) becomes meaningless in the extreme case where each sensor responds in some completely arbitrary (random) fashion to the stimulus be- 350 14 Chemical Sensor Array Optimization: Geometric and Information Theoretic Approaches

cause of extremely low signal-to-noise performance. Their analysis therefore applies to noiseless systems which cannot be obtained in practice. Not until Gardner and Barlett’s paper in 1996 on performance specification for chemosensor arrays was there any serious further treatment of this topic [2]. They were careful to consider noise to be central to the performance of these systems. By considering the chemical sensor array to perform a noisy (and therefore irreversible) mapping of a single point in the sample space to a spread of points in sensor space, they were able to quantity the effect of individual sensor noise on array perfor-

mance. They defined an error volume, Vn, as an ellipsoid within sensor space where the principal axes define the noise dispersion (or random error), r , of each sensor xi response, xi

n=2 n 2p Pi¼1rx V ¼ i ; ð14:2Þ n nCðn=2Þ

where CðÞ is the standard Gamma function. This equation provides a useful measure of the error introduced by noisy sensors. They then went on to define an important

quantity of the total number of array response vectors that may be discriminated, Nn, in view of this noise as

n Pi¼1FSDðxiÞ Nn ð14:3Þ Vn

this being the total volume within sensor space divided by the error volume of a single

hyperellipsoid feature, where FSDðxiÞ gives the full-scale deflection of sensor xi. While this is a useful measure of the theoretical limit to the number of distinct features identifiable by an array in principle, in practice it is unlikely to be attainable because not all of the sensor space may be accessible by the system, depending on the range of the stimulus and the tuning of the array. As an extreme case, consider an array of sensors each with identical sensitivities (tunings) to the stimulus. As we will see from the geometrical arguments below, the response of such an array would be confined to a 1 D sub-space (line) oriented within sensor space and would be unable to discriminate between any two compounds. As the dimensionality (n) of the array increases, this effect becomes more severe and usually electronic nose systems use an extremely small portion of the available sensor space as a result of the non-orthogonal sensor tunings and dynamic range of the stimulus. Consequently, the effects of array tuning and range of the input are as fundamental as noise in defining the system performance. Also note that Eq. (14.3) is an approximate bound because it assumes optimal packing of error hyperellipsoids in sensor space. Although the factors defining the performance of chemical sensor arrays for odor analysis have been given some consideration during the development of electronic nose technology over the past twenty years, there still exists no comprehensive theory of performance that can be widely applied to these systems. Without such a theory it is not possible for a manufacturer, user, or researcher to specify the likely performance of a given sensor array for a particular problem domain and, even more importantly, 14.3 Geometric Interpretation 351 optimize its performance for a given task. The lack of a clearly defined performance specification is a real barrier to the uptake of electronic nose systems, because the manufacturers of competing chemical sensing technologies such as gas chromatographic or mass spectrometric-based instrument manufacturers are able to rigorously specify detection limits for particular analytes, either individually or in combination. Current methods of specification for electronic nose systems are largely empirical, requiring vast numbers of measurements to be made to a wide range of single analytes. Since these individual measurements cannot predict the overall system performance to complex mixtures of analytes that are routinely encountered in the real world, this makes a complete empirical specification impossible for all but the most constrained and artificial of cases. Furthermore, system performance cannot be quantified in any meaningful manner. Empirically based optimization strategies, which rely on databases of measurements to different stimuli, may be used, but usually the number of parameters to be optimized is prohibitive. The lack of a performance theory also means that any attempts at array and system optimization must be carried out using empirically-based heuristic methods. There are no guarantees of optimizing the performance for chemical sensor arrays designed using these methods, and the user cannot be sure that they have the best array for their task. In this chapter we discuss the recent work on this topic by the authors, which relates both the array tuning and noise aspects to sensor-array performance. We believe this represents a unified framework within which to rigorously define system performance that provides the means to specify, and the foundation to optimize electronic nose systems. Optimization measures are developed to characterize different aspects of sensor array performance including system detection limits to specific odor stimuli, a theoretical maximum of the number of odor features that may be detected by a chemical sensor array (to a given confidence interval), and the resolution of an array to neighboring odor stimuli (closely related to the signal-to-noise ratio). These measures may be widely applied independently of sensor technology, sensor preprocessing methods, pattern recognition techniques, or the odor delivery system. Finally, we consider how these measures may be used within an optimization scheme to select the best chemical sensor array for a particular problem domain.

14.3 Geometric Interpretation

In order to demonstrate the effects of noise and tuning on array performance, we need to show the mapping between odor space and sensor space as carried out by a sensor array. We firstly assume this to be linear although we will later drop this restriction. 352 14 Chemical Sensor Array Optimization: Geometric and Information Theoretic Approaches

14.3.1 Linear Transformations

We begin by considering a linear stationary chemical sensor model

x ¼ a0 þ a1c1 þ a2c2 þ :::ajcj þ ::: þ amcm; ð14:4Þ

where x is the sensor response (note here there is no noise and so x is a deterministic function of the stimulus – later we will consider x to be a random variable which

fluctuates around some mean response value), cj gives the concentration of analyte j, and aj defines the sensitivity of the sensor to the same analyte. The term a0 gives the sensor response when no stimulus is present, often referred to as the ‘baseline’ response for the sensor. Although this linear model only applies to a subset of chemical sensor technologies (e.g. electrochemical cells and fluorescent indicators), and only then up to an operating limit, more general models of sensor response will be considered in Section 14.5 after results have been developed for the linear case. An electronic nose may be modeled as comprising n sensors within an array, each

with potentially different sensitivity terms, aij. This linear model is convenient since we may apply linear algebra to represent the array as 0 1 0 10 1 0 1 x a a ... a c a B 1 C B 11 12 1m CB 1 C B 10 C B x2 C B a21 a22 ... a2m CB c2 C B a20 C B . C ¼ B . . . . CB . C þ B . C ð14:5Þ @ . A @ . . .. . A@ . A @ . A

xn an1 an2 ... anm cm an0

or simply

x ¼ Ac þ a0; ð14:6Þ

where A is termed the sensitivity matrix and a0 the residual baseline vector for the array. Using this simplified view we may consider the array of sensors to be carrying out a linear (affine) geometric transformation between odor space, c, and sensor space, x. We may choose any basis for representing c and x, but the simplest for the purposes of visualization is over Rm and Rn respectively. Within this representation we can uniquely define any combination of odor stimuli and with it a specific sensor array response. From Eq. (14.6) it is clear that the nature of the transformation between odor and

sensor space is uniquely defined by A and a0, which are properties of the array. In terms of the capability of the array to detect changes in the stimulus, the residual baseline vector is of no interest, because it has no effect on the response of the array to different odor compounds – it acts only as an offset term. Consequently, we will not

consider a0 any further in our analysis. On the other hand, the sensitivity matrix is fundamental to the system performance as it determines the array response to the stimulus in the linear case, and so this will be the main focus of our discussion. It is instructive to visualize the action of the sensor array directly, by considering the trivialized example of a 2-odor to 2-sensor transformation for a variety of sensitivity 14.3 Geometric Interpretation 353

Fig. 14.1 Visualization of a 2-odor to 2-sensor transformation for different examples of linear sensitivity matrices, A, a) orthogonal sensors through to d) identical sensors. V0: Hypervolume of accessible odor space, Vs: Hypervolume of accessible sensor space. (Reprinted with permission from Pearce [3])

matrices, as shown in Fig. 14.1. It is clear that the sensitivity matrix has a profound effect on the nature of the transformation between the odor space (domain) and the sensor space (range). In particular, for perfectly orthogonal sensors (with unit gain) as shown in Figu. 14.1a, where the sensitivity matrix is simply the identity matrix, I,no transformation occurs from the domain onto the range and so it preserves the area of the original odor space, in other words the transformation is isometric. However, as the orthogonality of the individual sensor sensitivities decreases, as shown in Fig. 14.1b, c, there is a noticeable collapsing of the domain onto the range so as to restrict the possible array response. In the other extreme, where the sensors are identical, as shown in Fig. 14.1d, all points within the domain are mapped onto a single line in the range. Clearly, such an array would be unable to distinguish between the two odor compounds, but would only be able to provide an estimate of the combined analyte concentrations. From these observations we can define an important performance parameter for an array, the hypervolume of accessible sensor space, Vs, which in each example is equal to the area spanned by the transformation of the domain onto range.

It is noticeable in Fig. 14.1 that the total transformed area, Vs, is related to the orthogonality of the two sensors. A well-known result from linear algebra states that, given an affine transformation defined by a square matrix D, then a region of unit volume within the domain is transformed into a region within the range, the volume of which is equal to the absolute value of the determinant of the transformation matrix, that is, jDj [4]. Consequently, if the possible linear combinations of odor stimuli covers a 354 14 Chemical Sensor Array Optimization: Geometric and Information Theoretic Approaches

2 :52 Fig. 14.2 Visualization of a 3-odor to 2-sensor transformation, A ¼ 122

defined volume in the domain, which we term the hypervolume of accessible odor space,

V0, then in the m-odor to n-sensor (m ¼ n) case we have

Vs ¼ V0absðjAjÞ; ð14:7Þ

0 0 where V0 ¼ Pici gives the volume in odor space (c is the maximum concentration considered for a specific odor component). The absolute value must be taken because the determinant gives the ‘oriented volume’ which may be negative. The form of Eq. (14.7) is very similar to the array optimization measure proposed by Zaromb and Stetter as long ago as 1984 [1]. However, they never discussed how this measure applies generally to chemical sensor arrays, because the determinant is only defined for a square matrix, and so can only be used when the number of odors is equal to the number of sensors. In general, electronic noses map many more odor components onto fewer sensors in order to discriminate between complex odors using as simple an array as possible. Consequently, we need to generalize the measure defined by Eq. (14.7) for a transformation of arbitrary dimensionality that may be carried out by a chemical sensor array. To do this we need to consider an example transformation for which the sensitivity matrix is not square, as shown in Fig. 14.2. This visualization shows how the cube of unit side within odor space is mapped onto the plane in sensor space. Clearly, in this

example the area defining Vs cannot be found by a single determinant. If we consider the three 2 2 minors (of order 2) of A, then each of these represents how a single face of the cube is transformed into the range. That is, each face of the cube is transformed into a region in sensor space defined by its corresponding minor of A. So, for example, the face of unit area {(0,0,0),(0,0,1),(0,1,1),(0,1,0)} in the domain has a transformed area equal to the absolute value of the determinant of the 2nd-order minor, :52 abs det ¼ 3, in the range. 22 14.4 Noise Considerations 355

Furthermore, it is evident from Fig. 14.2 that the total region Vs comprises of the three transformed perpendicular faces of the cube, suggesting the general result

Xm Xm Xm 0 0 0 Vs ¼ ... cpcq ...cr absðjMpq...r jÞ for m n; ð14:8Þ p¼1 q6¼p;q¼1 r6¼p;r6¼q;r¼1 where Mpq...r is the minor of order n which is obtained by taking the columns (p, q,…,r) of A. Again the absolute value is taken because the areas defined by the minors must be additive. This result can be shown to apply generally m n to any affine transformation between m-dimensional odor space and n-dimensional sensor space [5], and so may be used to calculate the allowable space that may be accessed by a given array for a stimulus volume. For an array of linear chemosensors Eq. (14.8) completely specifies the role of the array tuning in terms of defining the total volume of accessible sensor space, which may be considered as the range of the system as a whole. For instance, applying Eq. (14.8) to the example shown in Fig. 14.2 gives a value for Vs ¼ 8:5, which can be easily verified using elementary geometry.

14.4 Noise Considerations

Although the performance of perfectly specific chemical sensor array (such as one where the off-diagonal terms of A are zero) is simple to characterize – by simply measuring the detection limit of the sensors individually – the case for cross-sensitive sensors is less straightforward. In the latter case, the overall sensitivity of the array to an individual compound arises from the combined sensitivity of a number of devices. Consequently, it is necessary to understand how these individual sensitivities contribute to the array performance.

14.4.1 Number of Discriminable Features

So far we have considered the transformation carried out by a sensor array to be noiseless, that is, there is a perfect correspondence between points within odor space and points within sensor space. In the noiseless case, the magnitude of Vs is unimportant since it is always possible to perfectly resolve neighboring points in odor space, no matter how close in proximity. In practice, of course, all measurements are limited by noise and so chemical sensors generate a non-reproducible response to the same stimulus. Instead of there being perfect correspondence between odor and sensor space, we must now view the noise process as mapping single points in the stimulus space onto a region (usually small) in sensor space where the likelihood of obtaining a particular measurement is determined by some probability density function. When the noise process is introduced into the transformation, the magnitude of Vs becomes of great importance be- 356 14 Chemical Sensor Array Optimization: Geometric and Information Theoretic Approaches

Fig. 14.3 Sensor space representation where sensor noise has been

represented as ellipses superimposed on the region Vs. The addition of the noise components for each sensor dx is equal to r . For illustration i xi purposes we assume here that the noise is independent of the stimulus or sensor response. (Reprinted with permission from Pearce [3])

cause it restricts the total number of discriminatable features for a given significance level. In the simplest case, where the noise in each sensor is considered to be independent of both the stimulus and the response magnitude, we may define a confidence interval in sensor space as an m-dimensional hyperellipsoid, where the cross-section along the principal axes is given by dx ¼r , the standard deviation of the noise (or random i xi error) for sensor i. A representation of the noise process combined with the sensor

array transformation is shown in Fig. 14.3 for the 2-sensor case where the region Vs is packed by the error ellipses. After Gardner and Bartlett [2], each ellipse corresponds to a single stimulus point in the domain, the number of ellipses that may be packed into

the region Vs gives the number of discriminable odor features, Nn, a bound for which was given in Eq. (14.3). By also taking into account the accessibility of the sensor space for a defined region of the sensor space, as discussed, we can estimate the number of features which can be discriminated by the array on the average

Vs Nn : ð14:9Þ Vn

Most importantly Eq. (14.9) provides an estimate of the number of discriminable features that can be coded by a chemosensor array, taking into account both noise and 14.4 Noise Considerations 357 array tuning. The value of Vs limits the access to the sensor space depending on the dynamic range of the stimulus and the array tuning, through Eq. (14.8). By using the formulæ given for Vs in the linear case, Eq. (14.8), and non-linear case, Eq. (14.23) (as will be discussed in Section 14.5), it is possible to produce an estimate for Nn for any chemical sensor array.

14.4.2 Measurement Accuracy

Of particular interest is how the noise generated in sensor space determines the measurement accuracy of the array to individual components of the odor stimulus. This may be achieved by considering the inverse mapping of noise in sensor space onto odor space. We first define a noise matrix, gx, which comprises each of the sensor errors as a diagonal matrix of the form 0 1 r 0 ... 0 B x1 C B . C B 0 r . C g ¼ B x2 C: ð14:10Þ x B . . C @ . .. 0 A 0 ... 0 r xn

1 We can now quantify the inverse transformation of the noise matrix, gx, via A so as to generate the noise components in terms of the odor space, to give a corresponding detection limit for each individual odor component, Dc. This corresponds to solving the system of equations gx ¼ A Dc for Dc. Depending on the form of A there are three possible cases to consider, as shown in Table 14.1. The most straightforward case is where there are the same number of odor components as there are sensors, which produces a square matrix A, and is of full rank (all sensors are linearly independent but not necessarily orthogonal, if this is not the case then we consider the system to be underdetermined). The overdetermined case occurs when there are more sensors than individual chemical compounds, given that the rank of A is m. Because of the typically high dimensionality of the stimulus in the case of olfaction, the overdetermined case would not be usual. However, it is of direct interest to researchers who use arrays of broadly tuned chemical sensor arrays for single gas analysis or sensing mixtures of gases using such systems. This case is dealt with in Appendix 14.A. More usual in electronic nose systems is the underdetermined case where there are more odor compounds than independent sensors within the array. This case is studied in Appendix 14.B, assuming that the distribution of the stimuli is Gaussian. In the case where n ¼ m and A is of full rank, there is no loss of dimensionality during the forward transformation, i.e. Vs > 0. A unique two-sided inverse exists, A–1, and each point within the domain has a one-to-one mapping with points in the range (subject to noise constraints). Now

1 DC ¼ A gx; ð14:11Þ 358 14 Chemical Sensor Array Optimization: Geometric and Information Theoretic Approaches

Table 14.1 Possible cases of transformations between odor and sensor space, showing examples for each case

Odor space Sensor space

Uniquely determined

ðn ¼ mÞ\ðVs > 0Þ

Overdetermined ðn > mÞ

Underdetermined

ðm > nÞ[ðn ¼ m \ Vs ¼ 0Þ 14.4 Noise Considerations 359 and so the detection limit for the array is simply the noise matrix scaled by the elements of the two-sided inverse of the sensitivity matrix and is therefore simple to calculate. The solution is then of the form 0 1 dc dc ... dc B 11 12 1n C B dc21 dc22 ... dc2n C DC ¼ B . . . . C; ð14:12Þ @ . . .. . A

dcm1 dcm2 ... dcmn

T where each column ðdc1i; dc2i; ...; dcmiÞ gives the noise vector for sensor i projected onto odor space, and each row (dcj1; dcj2; ...; dcjn) gives the noise components for each sensor projected onto the odor component j. These noise components may act in the same or opposite directions and so the total squared error for the array is

Xn Xm 2 2 e ¼ dcji; ð14:13Þ i¼1 j¼1 whereas the overall contribution of sensor i to error in odor space is

Xm e2 ¼ dc2; ð14:14Þ xi ji j¼1 and finally the total error produced by all the sensors for odor component j is

Xn e2 ¼ dc2: ð14:15Þ cj ji i¼1 The latter expression is particularly important because it provides measure of the detection limit of the noisy chemosensor array to each compound j owing to the array tuning and noise properties. Finally, it is also useful to define a signal-to-noise ratio for neighboring points in odor space. This tells us how easy it will be to discriminate between these points given the array tuning and noise performance. For two given stimuli separated by Dc we see that this corresponds to a sensor response of magnitude

Dx ¼ ADc; ð14:16Þ which leads to the local signal-to-noise ratio for stimulus difference Dc

kDxk2 SNR 0 ¼ ð14:17Þ Dc T trðgxgx Þ where kk is the Euclidean vector norm and tr() is the matrix trace operation. This measure is extremely useful because it allows us to predict the likelihood of discrimination between two neighboring points in odor space using a particular chemosensor array. 360 14 Chemical Sensor Array Optimization: Geometric and Information Theoretic Approaches

To apply this theory, the experimentalist or practitioner needs to be able to provide suitable values for the parameters within Eq. (14.11) and Eq. (14.17). In particular,

measuring values for A and gx is the key requirement. The sensitivity matrix, A, can be measured directly by varying individual stimulus components and calculating the regression parameters of a linear fit between concentration and sensor response (least squares). Because at this point the model assumes a linear behavior, it is straightforward (although time consuming) to estimate all of the values for A, because the sensitivity of each sensor to a particular compound can be measured independently and then assumed to sum linearly in our model. Therefore, over some linear operating region (often assumed to be for low concentrations), regression parameters for the concentration dependence of each sensor to each compound can be estimated directly from the sensor response data. Of course, the number of individual compounds may be too high to be able to realistically estimate an individual sensitivity between each sensor and each compound. However, note that many compounds may be grouped together to act as a single component (dimension in our model) as long as the sensor responds linearly to the mixture over the operating region.

Estimating values for gx provides more of a challenge because it requires estimation of noise properties in each of the sensors. The model assumes the noise for each sensor is constant over the stimulus range and is independent of noise sources in other sensors (later we will show that we can also deal with stimulus-dependent noise properties). This assumption makes estimation of the standard deviation of the noise straightforward. There may be two forms of noise that the practitioner might wish to take account of when using the model. First, intermediate to high-frequency noise in the sensor response (arising from instantaneous noise sources in the sensor or interface electronics), which may be quantified from the fluctuations in the time series of data to no stimulus or constant stimulus. The second form of noise is the reproducibility of the response of the sensor to repeated stimulus. This would require repeating identical stimuli many times and quantifying the dispersion of responses in each of the sensors. Because the noise is assumed here to be independent, then the noise can be characterized independently in each sensor. If the noise varies over the stimulus range then a mean value can be assumed for the purposes of the linear model. If any of these restrictions do not seem reasonable given the data available for the sensors being optimized, then a more complex, non-linear model such as those discussed below, will need to be considered.

14.4.3 2-Sensor 2-Odor Example

Some of the concepts become clearer through a trivialized example of a 2-linear sensor array responding to 2-odor compounds. To simplify the calculations, we assume that both sensors have the same noise r and that this is independent of the stimulus or sensor response. The sensitivity matrix is then simply 14.4 Noise Considerations 361

Fig. 14.4 The effect on the optimal squared estimation error, e2, from variations in the tuning of one sensor within an odor sensing array of two sensors, after fixing the sensitivities for the other sensor. The array is composed of 2-linear sensors with Gaussian

noise, where the tunings of one of the sensors is fixed, a11 ¼ 1; a12 ¼ 0:5. (Reproduced with permission from Sa´nchez-Montan˜e´s and Pearce [6])

a a 11 21 ; ð14:18Þ a12 a22 and gx is r 0 ; ð14:19Þ 0 r and so applying Eq. (14.11) we obtain the solution r a a DC ¼ 11 21 ; ð14:20Þ a11a22 a12a21 a12 a22 giving the formula for the total squared error for the array as

2 2 2 2 2 2 a11 þ a12 þ a21 þ a22 e ¼ r 2 : ð14:21Þ ða11a22 a12a21Þ

As an example of performance optimization we might wish to choose a11, a12, a21, and a22 in order to minimize this error. Clearly, a unique solution is not possible, but by fixing the sensitivities of one of the sensors, say a1j, we can visualize the effect on the error as we vary the tunings for the other sensor, say a2j (Fig. 14.4). The results are intuitive by considering the situation when one sensor possesses sensitivity terms that are multiples of the other (i.e. the sensors are identical after normalization). In this 362 14 Chemical Sensor Array Optimization: Geometric and Information Theoretic Approaches

Table 14.2 Models of concentration dependence for a variety of chemical sensors and their behaviors. All models assume that no competition for sites within the sensor takes place, and that chemicals act independently on the sensor. (Reprinted with permission from Pearce [3])

Device Model Behavior P m Electrochemical fuel cell, Linear x ¼ j¼1 aj cj þ a0 fluorescent indicators

P m i Metal oxide semiconductor Power x ¼ j¼1 aj cj þ a0

P Conducting polymer Langmuir x ¼ m ½bj aj cj þ a j¼1 1þaj cj 0

case, the array is unable to distinguish between the individual stimuli so the reconstruction error tends asymptotically towards infinity, reflecting the impossibility of discrimination between the separate stimuli in this case. This is represented by the ridge along the center of Fig. 14.4, (left), where the ratio between the sensitivity

terms a21 : a22 is 2:1. If we constrain each of the sensitivity terms to the range [0, 1] (i.e. the sensor response can only increase from its baseline value and its sensitivity is

constrained), then the best performance is obtained when a21 ¼ 0 and a11 ¼ 1, that is, when the second sensor is as different as possible from the first sensor within the specified constraints. 14.5 Non-linear Transformations 363

14.5 Non-linear Transformations

Because only a subset of chemical sensors is considered to behave linearly up to an operating limit, it is necessary to extend the methods developed in Sections 14.3 and 14.4 so that they may be applied more generally. The concentration dependence of the most popular chemical sensor types to be used within electronic nose systems are shown in Table 14.2. Of these, metal-oxide semiconductor sensors are arguably the most widely used in existing systems. These have been modeled by a power law, where ri typically lies between 0.6 and 0.8 but may also

Fig. 14.5 (a) Visualization of 2-odor to 2-sensor transformation using the non-linear power law model for metal oxide semiconductor devices: r1 r1 r2 r2 x1 ¼ a11c1 þ a12c2 , x2 ¼ a21x1 þ a22c2 , where r1 ¼ r2 ¼ 0:8 and a11 ¼ 0:8, a12 ¼ 0:25, a21 ¼ 0:6, and a22 ¼ 0:25. (b) Plot of the determinant of the Jacobian for the same 2-sensor metal oxide device array showing how the localized feature volume varies with the stimulus. (Reprinted with permission from Pearce [3]) 364 14 Chemical Sensor Array Optimization: Geometric and Information Theoretic Approaches

depend on the stimulus. Conducting polymer devices are also very popular for use within chemical sensor arrays and these have been described as behaving according to a Langmuir isotherm model. For these and other non-linear sensors, a sensitivity matrix may be formed in the non-linear case from the Jacobian matrix, A

Fig. 14.6 (a) Visualization of 2-odor to 2-sensor transformation using the non-linear Langmuir isotherm model for conducting polymer devices: x ¼ a11 c1 þ a12 c2 , x ¼ a21 c1 þ a22 c2 , (b) plot of the deter- 1 1þa11c1 1þa12 c2 2 1þa21 c1 1þa22 c2 minant of the Jacobian for the same 2-sensor conducting polymer device array showing how the localized feature volume varies with the stimulus. (Reprinted with permission from Pearce [3]) 14.5 Non-linear Transformations 365 0 1 @ @ @ x1 x1 ... x1 B @c1 @c2 @cm C B @x2 @x2 @x2 C B @ @ ... @ C B c1 c2 cm C A ¼ B . . . . C ð14:22Þ @ . . . . A . . . . @xn @xn @xn @c @c ... @c 1 2 m c1;c2;...;cm for some operating point (c1, c2,…,cm) in odor space. This linearized sensitivity matrix may then be used in place of A as defined by Eq. (14.6) so that the analysis developed in Sections 14.3 and 14.4 may then be applied in the general non-linear case. The determinant of the Jacobian, | ¼jAj, may then be used to approximate the localized hypervolume for the transformation for a particular operating point, which we call the localized feature volume. Furthermore, the Jacobian may also be applied to calculate the hypervolume of accessible sensor space in the non-linear case, because

ð 0 ð 0 ð 0 cm c2 c1 Vs ¼ ... | dc1dc2 ...dcm: ð14:23Þ 0 0 0

The fitting of experimental data for the practitioner using these non-linear models is straightforward. Rather than finding the regression parameters that fit the concentration dependence of sensor reponse in the linear case, we should now estimate the regression parameters of the model in the general non-linear case. Such non-linear regression can be achieved by most statistical software packages. Because of the nature of the models described in Table 14.2, the action of each of the compounds still sums linearly (even though their dependence on individual compounds may be non-linear) and so the sensor response to each compound may be analyzed independently. More complex models of analyte competition for sites in each sensor could be developed and may still be applied using Eqs. (14.22) and (14.23). As with the linear models, the noise is considered to be independent of the stimulus. The Fisher information approach, to be described below, should be used in the case of stimulus-dependent noise. Examples of calculations for the noiseless non-linear case are shown for metal-oxide semiconductor sensors in Fig. 14.5 and for conducting polymer sensors in Fig. 14.6, using the sensor models summarized in Table 14.2. For both examples, the non-linear mapping of 2-odor space is shown, showing how the non-linearity in each sensor contributes to the transformation as a whole. The nature of the models implies that the metal oxide semiconductor devices are far more linear in their behavior, which is verified by contrasting the mappings onto sensor space for both sensor varieties. In particular, the localized volume of the transformation in the conducting polymer case is shown to tend towards zero with increasing stimulus concentration. This is also shown by Figure 14.6 which shows the linearized Jacobian at different points in the stimulus space. As c1; c2 ! 1 then | ! 0, verifying the observation. In contrast, the determinant of the Jacobian for the metal-oxide array never reaches zero, because the behavior of these sensors is more linear. Note that the same analysis on an array of 366 14 Chemical Sensor Array Optimization: Geometric and Information Theoretic Approaches

Fig. 14.7 A hypothetical statistical estimator takes the response vector x, from a sensor array and uses this in order to estimate (reconstruct) the stimulus. The tuning parameters for each of the sensors are represented as parameters to the sensor array. (Reproduced with permission from Sa´nchez-Montan˜e´s and Pearce [6])

linear sensors would produce a perfectly flat feature localized volume. Hence the localized feature volume map provides an intuitive visualization of the performance of the sensor array in detecting the stimulus.

14.6 Array Performance as a Statistical Estimation Problem

We can also consider the definition of chemosensor array performance in a different context, one which we will show provides certain advantages in the calculation of the array error. Here we consider the data produced by a chemosensor array as being part of a statistical estimation problem as outlined in Fig. 14.7. Each sensor within the array produces a response dependent on its tuning to the stimulus plus some noise. A hypothetical statistical estimator (one produced using, for example, maximum likelihood or Bayesian estimation methods) uses the noisy response from the array to attempt to reconstruct the stimulus. Because of this noise, if we present the same stimulus c to the system several times, the estimator response cˆ will not be the same on each occasion but will fluctuate around a certain mean value. An estimator should be right on the average, that is, if we present the same stimulus c many times, the mean of the different estimations cˆ should be equal to c. If the estimation satisfies the property we call it unbiased. Moreover, the variance of the response of the estimator when the stimulus is fixed should be as small as possible. If the estimator is unbiased, its squared error in the estimation coincides with its variance. Depending on the tuning parameters of the individual sensor elements and their noise properties, the accuracy of the overall sensor system in estimating the stimulus varies in addition to the range of stimuli that may be tested. A typical goal in choosing 14.7 Fisher Information Matrix and the Best Unbiased Estimator 367 which sensors to incorporate into an artificial olfactory system is to maximize the accuracy with which the sensory system can estimate/predict the stimulus or optimally discriminate between similar stimuli. By considering a hypothetical unbiased statistical estimator that uses the sensor array response in order to estimate the individual stimuli within a complex odor mixture, we can define and test how different tuning parameters of the sensor array effect the accuracy of stimulus reconstruction.

This arrangement is shown in Fig. 14.7 where each sensor, i, generates a response, xi, to the multicomponent stimulus c. Conveniently, our problem when placed in this context is well known to the field of statistical estimation, and classical results exist that we can call upon here. For example, the variance of any unbiased estimator that might be constructed for this purpose has a well defined limit through the “Crame´r-Rao bound”, which we will make use of [7]. Furthermore a direct relationship between the Crame´r-Rao bound and Fisher information exists that allows us to calculate this bound, and therefore quantify the performance of the array in reconstructing the stimulus.

14.7 Fisher Information Matrix and the Best Unbiased Estimator

When a multicomponent odor stimulus c is exposed to the sensor array, the array of sensors gives a response x, of which component i denotes the response of sensor i. Because of the noise and nonreproducibity of the sensor, the array response is not deterministic so it follows some probability density function pðxijcÞ conditioned on the stimulus. The elements of the Fisher information matrix (FIM), Jjj 0 ðcÞ, are defined as [7] ð ! ! @ @ Jjj 0 ðcÞ¼ dx pðxjcÞ ln pðxjcÞ ln pðxjcÞ ; ð14:24Þ @cj @cj 0 where j and j 0 are both individual stimulus components. Some understanding of what is being measured by the Fisher information can be gained by considering the simplified case of a single sensor responding to a single odor component. For ease of the analysis, let us assume that the sensor responds to stimulus concentration with a Gaussian tuning curve (note this is not physically reasonable for a chemosensor but is for illustrative purposes only). In this case we have the situation shown in Fi- gure 14.8a. Now Eq. (14.24) reduces to 1 df ðcÞ 2 JðcÞ¼ ; ð14:25Þ r2 dc where f(c) is the mean sensor output to the stimulus c, hxjci, in this example following the Gaussian curve (see Fig. 14.8a) and r is the standard deviation of the noise shown as error bars in the same figure. From this simplified example we see that the Fisher information scales inversely with the noise variance but is linearly dependent on the 368 14 Chemical Sensor Array Optimization: Geometric and Information Theoretic Approaches

square of the slope of the tuning curve (here concentration dependence). We see that the slope of the tuning curve is greatest at the inflexion points of the Gaussian, which from Eq. (14.25) is also where the Fisher information is maximum, Fig. 14.8b. At the peak of the Gaussian where the slope is zero, the Fisher information is also zero, Figure 14.8c. This result is intuitive because if we wish to measure a small change in the stimulus it is far better to be operating on the slopes of the tuning curve, where we obtain a relatively large change in sensor output for a given stimulus change, compared to at the peak, where the change in sensor response will be close to zero. So we see that the Fisher information concisely describes the combined role of sensor tuning and noise in defining estimation performance. Although the Fisher information may not be straightforward to interpret it directly, we can relate it to the reconstruction error of the stimulus through the Crame´r-Rao bound. This states that for every unbiased estimator that uses the data x for estimating the stimulus c,ascˆ, the squared error for stimulus component j satisfies DE 1 2 varð^cjjcÞðJ ðcÞÞjj ec ; ð14:26Þ j opt

where var means variance, hi is the expected value or mean, and ^cj is the estimation of the component j of c, j ¼ 1; ...m [7]. And so this also provides a valuable link to the geometric theory of array error considered in Appendix 14.c. This result allows us to directly calculate the minimum expected reconstruction error for a given stimulus component j from the jth diagonal element of the inverse of the FIM. Furthermore, the total expected squared reconstruction error across the entire array is equal to the summation of the errors in each of the components. That is,

Xm Xm ^ ^ 1 2 varðcjcÞ¼ varðcjjcÞ ðJ ðcÞÞjj e opt; ð14:27Þ j¼1 j¼1

Fig. 14.8 In this example the sensor is characterized by (a) a bell- shaped tuning curve with overlapping Gaussian noise. The bars show the standard deviation of the noise. (b) The points that maximize Fisher information are those where the slope of the receptive field is higher. (c) Points where the slope is zero make the Fisher information minimum 14.8 Fisher Information Matrix Calculations for Chemosensors 369 and so the overall performance of the array in detecting all of the stimuli is defined by the elements of the FIM, Jjj 0 . In Appendix 14.C we show that the Fisher information and geometric approaches to sensor array optimization are equivalent when the noise in the sensors is independent of the stimulus. In the case of stimulus-dependent noise, the Fisher information approach should be used. There is another notion related to Fisher information called ‘discriminability’. This measures the ability of the system to distinguish between two similar stimuli c1 and c2. If we call Dc ¼ c2 c1, the ability of the system to discriminate between these is given by

d 0 DcT FDc: ð14:28Þ

The maximization of this quantity can be shown to be equivalent to the maximization of the local signal-to-noise ratio defined in Eq. (14.17). We now need to be able to calculate the FIM for different sensor array configurations in order to proceed.

14.8 FIM Calculations for Chemosensors

First, the FIM for an individual sensor i is given by the elements of the matrix ð ! ! i @ @ Jjj 0 ðcÞ¼ dxi pðxijcÞ ln pðxijcÞ ln pðxijcÞ : ð14:29Þ @cj @cj 0

It can easily be shown that when the array of sensors has uncorrelated noise, the FIM of the entire array,PJ, is equal to the summation of the individual FIM matrices for each i sensor i, that is i J . This is valid in a general sense – in other words the noise and concentration dependence of the sensors can be different across the array and can comprise different sensor technologies, noise properties and tunings. We now calculate the FIM elements for two example cases of chemical sensor by substituting the appropriate probability density function into Eq. (14.29) and rearran- ging. Case 1: Analog chemical sensor with Gaussian noise:

@f ðcÞ @f ðcÞ @r ðcÞ @r ðcÞ i 1 xi xi 1 xi xi 0 : Jjj ðcÞ¼ 2 þ 2 2 ð14 30Þ r ðcÞ @c @c 0 r ðcÞ @c @c 0 xi j j xi j j where f is the mean response for sensor x , i.e. f ðcÞ¼ x jc , which would be ex- xi i xi i pected to follow some model of concentration dependence, e.g. a simple linear model such as given by Eq. (14.1). However, the sensor model for the concentration dependence can be a far more complex, non-linear one. Note also that, in principle, the noise dispersion can depend on the stimulus. The Gaussian noise case is most appropriate for describing metal-oxide semiconductor and conducting polymer chemosensors 370 14 Chemical Sensor Array Optimization: Geometric and Information Theoretic Approaches

used within electronic nose systems, where the partial derivatives can be calculated for the sensor models given in Table 14.2. Case 1: Analog chemical sensor with Laplacian noise:

@f ðcÞ @f ðcÞ @a ðcÞ @a ðcÞ i 1 xi xi 1 xi xi 0 : Jjj ðcÞ¼ 2 þ 2 ð14 31Þ a ðcÞ @c @c 0 a ðcÞ @c @c 0 xi j j xi j j

where a (c) is the dispersion parameter of the Laplacian noise for that sensor. The xi Laplacian case is most appropriate for describing fluorescence-based optical chemosensors used within artificial olfactory systems, where the concentration dependence is approximately linear up to saturated vapor pressures of analyte [8].

14.8.1 2-Sensor 2-Odor Example

To illustrate these concepts we again consider two linear sensors to generate an analog response that is corrupted by Gaussian noise, identical to the example given in Sec- tion 14.4.1. This is a linear model and so the sensitivity of sensor i to stimulus com- @f ðcÞ ponent j is a constant a xi . Using Eq. (14.30) we can calculate the FIMs for each ij @cj sensor 2 2 1 1 a11 a11a12 2 1 a21 a21a22 J ¼ 2 2 J ¼ 2 2 r a11a12 a12 r a21a22 a22

Adding these to form the FIM of the array, then substituting into Eq. (14.27) and rear- ranging we obtain exactly the same form as Eq. (14.21), demonstrating equivalence between the geometric and information theoretic approaches in this case. In Appen- dix 14.C we show that this equivalence holds for any input dimension.

14.9 Performance Optimization

An outline of the optimization problem we will consider is shown in Fig. 14.9. A pool of k different sensor types is available, each with a unique profile of response to the m distinct molecular species relevant to the problem. Our instrument provides n channels, each of which we can assume may house any of the available sensors. Further- more, we will not consider duplication of sensor types in the array since this yields no additional information about the stimulus, but acts to reduce the noise in the system if averaging is employed (this case can be dealt with for independent noise by replacing the l identical sensors in the calculations with a single sensor of the same type but with r2 noise variance, r 0 2 ¼ xi Þ. xi l The optimization problem is then to select the single configuration that provides the k k! best sensing performance to the compounds of interest out of n ¼ n!ðknÞ! possible 14.9 Performance Optimization 371

Fig. 14.9 A cartoon of the optimization problem for chemical sensor arrays configurations. What is best here depends on the detection task to be solved. We envisage three possible criteria to be optimized in a practical system

1. Maximize the total number of Nn separate features that can be detected by an array. This is optimizing the range of the system and can be directly quantified from the geometric approach (Eq. 14.9). Shannon information theoretic approaches are more suited to calculating this value than the Fisher information [7]. 2. Maximize the signal-to-noise ratio obtained from the array for some vector or set of vectors in stimulus space. This is optimizing the resolving power or discrimination ability of the array and may be quantified using either the geometric (Eq. 14.17) or Fisher information (Eq. 14.28) approaches. 3. Estimate the concentrations of some of the compounds or some function of these, e.g. interfering compounds (distractors) could be present. This is optimizing the detection threshold of sensitivity for the system to specific components, which can be quantified using either the geometric or Fisher information approaches.

The case where we are interested in reconstructing the concentration of all the stimulus compounds has been extensively described in this chapter.

14.9.1 Optimization Example

We will illustrate the Fisher information maximization principle with a simple example. Consider a set of linear chemosensors each responding to combinations of three single odor components (m ¼ 3). The noise in all the sensors available to us is assumed to be Gaussian, independent of each other and the stimulus, and with equal 372 14 Chemical Sensor Array Optimization: Geometric and Information Theoretic Approaches

Table 14.3 Best three groups of solutions in the optimization example, and one singular solution (dependent sensors). The errors are given in units of the noise variance. The best error is achieved by just one solution (shown in the table), while the second best error and third best error are each achieved by six solutions (corresponding to replace a 1 sensitivity with 0.75 and 0.25 respectively). The table shows in- stantiations of these sub-optimal solutions

Best 2nd Best 3rd Best Singular

Sensitivities a1 a2 a3 a1 a2 a3 a1 a2 a3 a1 a2 a3 Sensor 1 0 1 1 0 0.75 1 0 1 1 0.5 0.25 1 Sensor 2 1 0 1 1 0 1 1 0.25 1 0.25 0 0.25 Sensor 3 1 1 0 1 1 0 1 1 0 0.25 0.25 0.75 e2 2.25 2.51 2.65 infinity

variance. For this example we assume each sensor is available with five graded levels of sensitivity to each of the three components, that is, 0, 0.25, 0.50, 0.75 or 1.0. Therefore, there are k ¼ 53 ¼ 125 possible sensor types. We would like to select a sensor array consisting of any three of these available sensors (that is, n ¼ 3). Therefore, we should select those three sensors from the 125 available that optimize the performance of the system in terms of the overall reconstruction error (criteria 3 above). For each of the 125! ð1253Þ!3! ¼ 317,750 possible array configurations we calculate the system Fisher information as previously described, in order to evaluate their performances. In Table 14.3 we show the three best groups of solutions. Note that the optimal configuration is formed by sensors with non-zero sensitivities as well as zero sensitivities i.e. they are mixed. The non-zero sensitivities are maximum in each case, showing that intermediate sensitivities are disregarded. This is intuitive, because providing as much gain as possible to each of the analytes will maximize the performance under all three optimization criteria discussed above – increased gain is always advantageous as long as it is not commensurate with equal amounts of noise. Importantly, the specific case (in which each sensor responds to a different component with maximum gain while its sensitivity to the others is zero) is not the best in our example (the expected squared error of this configuration is exactly 3r, in units of the standard deviation of the noice). This demonstrates that even if it were possible to develop perfectly specific sensors for given compounds, this would not yield the best possible performance for electronic nose systems, because some amount of overlap in sensor response is shown to be advantageous. Interestingly, the sensors forming the optimal configuration tend to have the same number of zero and non-zero sensitivities as the input dimension increases (data not shown). The number of zero sensitivities in each sensor of the optimal configuration tends to be the same as the number of zero sensitivities as the input dimension increases (data not shown). Arrays formed by non-independent sensors (some linear relationship between the sensitivities exists within the array) have infinite expected error because they are not able to discriminate between three-dimensional stimuli. For example, the singular configuration shown in Table 14.3 has only 2 independent sensors. Therefore, it can only discriminate between two-dimensional stimuli. 14.10 Conclusions 373

Fig. 14.10 Expected squared error for all possible sensor array configurations. The configurations are ranked according to their squared error. The grey zone indicates solutions where an unbiased exti- naty is comfortable its construct (dependent sensors). The error is normalized by the best expected squared error. Dotted line indicates the configuration in which the squared error starts to be greater than 100. The percentage of configurations whose error is greater than 100 is 22.46 %

The errors of all the 317 750 possible arrays are sorted and shown in Fig. 14.10. Critically, the error of any given configuration can be orders of magnitude greater than the error of the optimal configuration. Therefore, if the sensory array is designed randomly choosing three of the available sensors, we are likely to select a far-from- optimal configuration. We stress this point to indicate the importance of optimization in chemical-sensor-array design. For example, the probability of having an expected squared error more than 100 times the optimal one is 22.46 % (see Fig. 14.10). The technique illustrated in this example can be analogously used in general conditions: non-linear sensor noise that depends on the stimulus, other types of noise (non-Gaussian), bipolar sensor sensitivities, and arrays of sensors with different kinds of responses and noises [6]. Many more complex examples can be easily constructed.

14.10 Conclusions

In this chapter we have described two unified theories of chemical-sensor-array performance, using both geometric-based linear algebra and Fisher information approaches. The theories may be applied in a variety of conditions such as different sensor noise properties and different concentration-dependence models. More generally, any variety of different sensor types may be optimized within the same array. The geometric theory is particularly suited to visualization of the sensor array performance and the Fisher information copes with more complex scenarios, where for example the sensor noise is dependent on the stimulus. The utility of the approaches for array optimization is demonstrated using a number of simple examples that serve as the basis for more realistic applications of the theory. Manufacturers of electronic nose instruments may easily apply this theory in order to optimize the sensing performance of the systems they sell. Furthermore, we can envisage a catalog of parameters for each sensor used within practical systems today, which would make the optimization of sensor arrays to particular detection tasks a simple and routine operation. 374 14 Chemical Sensor Array Optimization: Geometric and Information Theoretic Approaches

Acknowledgments T.C.P. was supported by grant IST-2001-33066 from the European Commission and GR/R37968/01 from the United Kingdom Engineering and Physical Sciences Re- search Council. M.A.S-M. was supported by grant BFI2000-0157 from MCyT.

References

1 S. Zaromb, J. R. Stetter. Theoretical basis for 5 J. R. Wicks. personal communication. identification and measurement of air con- 6 M. A. Sa´nchez-Montane´s, T. C. Pearce. taminants using an array of sensors having Fisher information and optimal odor partly overlapping selectivities. Sensors & sensors, Neurocomputing, 38–40 (2001) Actuators, 6 (1984) 225–243. 335– 341. 2 J. W. Gardner, P. N. Bartlett. Performance 7 T. M. Cover, J. A. Thomas. Elements of definition and standardization of electronic Information Theory, John Wiley, 1991. noses. Sensors & Actuators B, 33 (1996) 8 T. C. Pearce, P. F. M. J. Verschure, J. White, 60–67. J. S. Kauer. Robust stimulus encoding 3 T. C. Pearce. Odor to sensor space trans- in olfactory processing: hyperacuity and formations in biological and artificial noses, efficient signal transmission, in Neural Neurocomputing, 32–33 (2000), 941–952. computation architectures based on neuro- 4 J. R. Wicks. Linear algebra an interactive science, (eds. Wermter S., Austin, J., and approach with Mathematica, Addison- Willshaw D.), Spinger-Verlag 2001. Wesley, 1996. 14.10 Conclusions 375

Appendices

14.A Overdetermined Case

The Fisher information approach described in the main text operates correctly in the overdetermined case. However, for the geometric approach described in the main text, we must find the least squares solution which leads to

T 2 1 T DC ¼ðA gx AÞ A gx ð14:32Þ and Xm Xn 2 2 e ¼ dcji: ð14:33Þ j¼1 i¼1

It can be easily verified that when A is square and non-singular these equations are the same as Eqs. (14.11) and (14.13) respectively.

14.B General Case with Gaussian Input Statistics

Here we consider the global optimal estimator (biased or unbiased) which minimizes the global expected error. When the sensors are linear and the noise is Gaussian this T 2 1 1 minimum error can be shown to be trðA gx A þ V Þ Þ, where V is the covariance matrix of the input stimuli, which are assumed to be Gaussian distributed. This equation is valid for all the cases (square A, underdetermined and overdetermined cases) as well as when the input statistics are not homogeneous, and so is the most general result.

14.C Equivalence Between the Geometric Approach and the Fisher Information Maximization P 2 T Because i;j xij ¼ trðxx Þ, using Eq. (14.32) we can rewrite Eq. (14.33) as

2 T 2 1 T T T 2 1 T 2 1 e ¼ trððA gx AÞ A gxgx AððA gx AÞ TÞ¼trððA gx AÞ Þð14:34Þ

On the other hand, if the sensors are linear and their noise does not depend on the stimulus, Eqs. (14.30) and (14.31) can both be expressed as

i 2 Jjj 0 ðcÞ¼ðgxÞii aijaij 0 ð14:35Þ

T 2 Then the total Fisher information matrix is J ¼ A gx A so by Eq. (14.27) the optimal T 2 1 error is simply trððA gx AÞ Þ, which coincides with that derived for the geometrical approach. A similar proof can be shown for when the sensors are non-linear. 377

15 Correlating Electronic Nose and Sensory Panel Data

Robert W. Sneath, Krishna C. Persaud

Analytical methods such as gas chromatography-mass spectrometry (GC-MS), or near infrared spectroscopy provide the mainstay for measurement of volatile components in food, agricultural, chemical, or environmental industries. Although data obtained give very precise measurements of individual components in a mixture, they give very poor indication of the sensory quality perceived by the human nose or tongue. The control of odor quality within these industries is associated with problems that are unique, because they also rely on human perception and preference for particular types of odors or tastes. It is difficult to relate the output of traditional analytical instruments to human perception, because the chemosensory systems of smell and taste use information gathered from the interaction of complex chemical mixtures with the biological sensors without separation of individual components. Many such industries therefore rely on human sensory panels that are trained to discriminate subtle nuances of smell and taste in a given product or raw material, or to quantify the odor level in a sample. This in itself presents problems because such panels can only cope with relatively few sample assessments per day, and are very costly to run. They may be used for optimization of a new product, periodic sampling of problematic systems, and random quality control. This highlights the need for automated chemical sensing systems that produce data that are easily correlated to human odor perception. The human nose contains a large array of chemical sensors, and patterns of information are processed in the olfactory brain of an animal in order to achieve quantification and discrimination of odors based on previous learning experiences. With instrumental means of odor measurement, the human user interface needs to be considered very carefully, as the results need to be presented in a form that can be easily interpreted by the user. If an electronic nose is applied, the signals produced by an array of sensors consist of measurements of responses to odors producing different patterns that are projected into multidimensional space. In many instances we are dealing with complex mixtures of compounds in which only relatively few components (which may be at very low concentrations relative to other components) are important in the determination of odor quality by a human sensory panel [1, 2]. Ol- factory data depend strongly on individual physiological differences, on measurement methods, and on psychological factors. Classifications of odors are necessary to put

some order in odor descriptions that are used in structure-odor relationships. Pub- lished classifications have been based on empirical, semi-empirical, or statistical approaches. In the last category, data may be obtained using semantic descriptions or profiles, or similarity estimations. The intensity data are perceived as the strength of a stimulus. They also present a huge variability, which makes it difficult to relate them to physicochemical properties.

15.2 Sensory Panel Methods

Correlation of human sensory panel responses with data from electronic nose instrumentation demands that both sets of information have good repeatability and accuracy, which are usually accounted for by frequent calibration against known standards. Novel methods sometimes need to be developed to calibrate a sensor array. Compli- ance of sensory panel data to accuracy and repeatability standards is often neglected. Unless this feature of data collection is attended to, correlations are likely to be poor. There are few sensory panel standards but one that is relevant for correlations with electronic noses’ is the European standard EN13725 [3]. However, although it only sets criteria for detection threshold measurements, it has many features that can be taken on board when measurements of the other dimensions of odor require standardization.

15.2.1 Odor Perception

Sensory perception of odors has four major dimensions: detectability, intensity, quality, and hedonic tone, and problems arise when we want to assign values to perception.

1) Detectability. There is no conscious subjectivity to this dimension: either the smell is detected or it is not, but every person will have their own detection threshold, which will vary in people depending on their own situation at the time. 2) Intensity, which refers to the perceived strength of the odor sensation, and the odor has to be at a supra-threshold level. 3) Quality, i.e. what the substance smells like; assessors usually work from an agreed list of descriptors. 4) Hedonic tone. This is a category judgement of the relative pleasantness or unpleasantness of the odor, which is a very personal description and can only have any objectivity assigned to it if a comparison is made with other odors.

Perceptions are qualitative, and will of course vary from person to person so we have to devise ways of standardizing the descriptions of the odor and/or standardizing the people who make the assessments. In all odor or taste-related industries this is common practice, selected and trained staff are used when blending teas, coffees, and 15.2 Sensory Panel Methods 379 perfumes, and they use a set of agreed descriptors between themselves in an attempt to make the descriptions objective.

15.2.2 Measurement of Detectability

Detectability is the only one of those dimensions that can be reduced to an objective perception. The only answers to the question “Can you detect the odor?” are “Yes” or “No”, although the value of the response depends on the assertiveness and honesty of the subject. The threshold of detection is different for each individual and can be affected by factors such as where the person is, by background odors, or by familiarity with that odor. Therefore, threshold values are not fixed physiological facts or physical constants, but represent the best statistically estimated value from a group of individual responses. Odor thresholds are estimated in one of two ways, by getting a yes/no response, as above, or by a ‘forced choice’ response where the subject is forced to choose which air stream, from two or more, smells. In the former classical evaluation, yes/no answers are, amongst other factors, dependent on the subjects’ honesty and motivation. If odors at a range of concentrations, alternating with blanks, are presented a sufficiently large number of times, yes/no answers may be evaluated with the aid of signal detection theory, to eliminate the effects of context. The forced-choice procedure is an attempt to measure a subject’s sensitivity, which is not influenced by fluctuations in criteria. Two or more choices are presented to the subject at a range of odorant concentrations, and it is the subject’s task to choose the one that is odorous from the other that is not. The assumption is made that the ob- server chooses the one that gives the largest sensory excitation, provided that there is no response bias towards one or more of the options. If the comparison stimuli (blanks) have been carefully defined and controlled, the proportion of correct responses can be used as a measure of sensitivity, because it will always be measured in comparison to blanks.

15.2.3 Transforming the Measurement of the Subject to the Subject’s Measurement of an Odor

The detection threshold value is a measure of the sensitivity of the assessor, but what we need to do is to measure, in a reliable way, the odor we are interested in. In all measurements, two criteria must be satisfied: accuracy and repeatability. This usually means manufacturing a sensor that produces the correct answer and will produce the same answer repeatedly. In olfactometry our sensor is the human nose. These sensors have been produced in a manufacturing process that has no quality control: therefore, from the production run, we must choose sensors that fit our criteria for accuracy and repeatability. The machine that presents the odor sample to the sensors must also be constructed and operated to achieve the criteria of accuracy and repeatability. 380 15 Correlating Electronic Nose and Sensory Panel Data

15.2.4 Assessor Selection

The key part of accurate odor measurement is the selection of the odor assessors. In order to select odor assessors, n-butanol has been specified in EN13725 as the reference material. Although it is recognized that a single component reference gas is not the ideal, no representative odorant mixture has yet been formulated. Only people with a mean personal threshold for n-butanol in neutral gas of between 20 ppb and 80 ppb and a log standard deviation of less than 2.3, calculated from the last 10 to 20 individual threshold estimates (ITEs), are acceptable. These assessors are continually checked for their detection threshold (at a minimum after every 12 odor measurements) and have to remain within these limits to be a panel member. This selection criteria used at the Silsoe Research Institute (UK) laboratory leads to the rejection of about 43 % of those tested because they are not sensitive enough and 12 % because they are too sensitive to n-butanol. The complete distribution of sensitivities of all 164 people tested in the Silsoe Research Institute laboratory, to date, is illustrated in Fig. 15.1. The butanol thresholds are grouped into 0.3 log intervals, i.e. less than 1.0, 1.0 to 1.3, 1.3 to 1.6, etc. plotted as a linear scale on the y-axis of Fig. 15.1. Of those who have a qualifying sensitivity, about two thirds have a threshold above the accepted reference value of 40 ppb (log 1.6). Selection of the panel members using the above method will lead to acceptable accuracy and precision and enable a laboratory to comply with the criteria set in EN13725 (Section 15.2.1).

15.2.5 Types of Dynamic Dilution Olfactometry

15.2.5.1 Choice Modes Two different choice modes can be used to obtain an individual threshold estimate. These choice modes and their requirements are described here. They all produce the

Fig. 15.1 Distribution of n-butanol olfactory thresholds for 164 subjects. The histogram highlights subjects who would qualify for sensory panel measurements 15.2 Sensory Panel Methods 381 common result of an ITE. The use of the ITE derived from either of these methods in the calculation of an odor concentration is then identical throughout this standard.

15.2.5.2 Yes/No Mode In the yes/no olfactometer; (Fig. 15.2) either neutral gas or diluted odor passes from the single port. The panel member is asked to evaluate gas presented from the single port and to indicate if an odor is perceived (yes/no). The panel members are aware that in some cases blanks (only neutral gas) will be presented. (A second port always presenting neutral gas may be made available to the assessor to provide a reference.) The samples may be presented to the assessors either randomly or in order of increasing concentration. When using the yes/no mode, 20 % of the presentations in a set of dilution series must be blanks to satisfy the operator that the panel members are giving the correct response when there is no odor present. For each panel member the measurement must include a dilution step at which they respond ‘No’ to a diluted odor and for two adjacent dilutions they must respond, ‘Yes’.

15.2.5.3 The Forced Choice Mode A forced choice olfactometer (Fig. 15.3) has two or three outlet ports, from one of which the diluted odor flows, while clean odor-free air flows from the other(s). In this method, panel members assess the ports of the olfactometer, from one of which the diluted odor flows, neutral gas flows from the other port(s). The measurement starts with a dilution of the sample large enough to make the odor concentration beyond the panel members’ thresholds. The concentration is increased by an equal factor in each successive presentation: this factor may be between 1.4 and 2.4. The port carrying the odorous flow is chosen randomly by the control sequence on each presentation. The assessors indicate from which of the ports the diluted odor sample is flowing, using a personal keyboard. They also indicate whether their choice

Fig. 15.2 Schematic diagram of a ‘Yes/No’ olfactometer. When the presentations are sorted in order of ascending concentration, the geometric mean of the dilution factors of the last FALSE and the first of at least two TRUE presentations determines the individual threshold estimate (ITE) for a panel member. The odor concentration for a sample is calculated from the geometric mean of at least two ITEs for each panel member 382 15 Correlating Electronic Nose and Sensory Panel Data

Fig. 15.3 Schematic diagram of a forced choice olfactometer. Panel members assess the ports of the olfactometer, from one of which the diluted odor flows, neutral gas flows from the other port(s). The port carrying the odorous flow is chosen randomly by the control sequence on each presentation

was a guess, whether they had an inkling, or whether they were certain they chose the correct port. Only when the correct port is chosen and the panel member is certain that their choice was correct is it taken as a TRUE response. At least two consecutive TRUE responses must be obtained for each panel member. The geometric mean of the dilution factors of the last FALSE and the first of at least two TRUE presentations determines the ITE for a panel member. The odor concentration for a sample is calculated from the geometric mean of at least two ITEs for each panel member. 3 The odor concentration has units of ouE m (European odor units per cubic meter). For measurements on reference odorants, this value can be converted to a detection threshold, expressed as a mass concentration using the known concentration of the reference gas divided by the ITE.

15.2.5.4 Laboratory Conditions For laboratories to conform to the required standard, they must be guaranteed to be free from odor. They are usually air-conditioned with activated charcoal filtration. They must also have a source of odor-free air, i.e. neutral gas, with which to dilute the odor sample. The olfactometer, which is a dilution device, is made entirely from approved materials, glass, tetrafluoroethylene hexafluoropropylene copolymer, or stainless steel. Samples are processed within 30 hours of collection.

15.2.5.5 Laboratory Performance Quality Criteria The EN13725 is based on the following accepted reference value, which shall be used when assessing trueness and precision: 15.2 Sensory Panel Methods 383

1ouE 1 EROM (European reference odor mass) ¼ 123 lg n-butanol When 123 lg n-butanol is evaporated in 1 m3 of neutral gas at standard conditions (20 8C) for olfactometry the concentration is 0.040 lmol mol1 (40 ppb). Two quality criteria, as below, are specified to measure the performance of the laboratory in terms of the standard accuracy and precision, respectively. Accuracy reflects the trueness or closeness to the correct value, in this case the true value for the reference material is 40 ppb and the precision is the random error. The standard specifies how these two quality criteria are calculated [3]. The criterion for accuracy Aod (accuracy of the odor measurement) i.e. closeness to the accepted reference value is:

Aod 0:217

In addition to the overall accuracy criterion, the precision, expressed as repeatability, r, should comply with

r 0:477

This criterion for repeatability can also be expressed as:

10r 3:0

This repeatability requirement implies that the factor that expresses the difference between two consecutive single measurements, performed on the same testing material in one laboratory will not be larger than a factor of 3 in 95 % of cases.

15.2.5.6 Compliance with the Quality Criteria The performance of an olfactmetry laboratory is monitored continuously by checking the accuracy and repeatability of the results of measurements of n-butanol. Fig- ures 15.1, 15.4 and 15.5 illustrate this over the first five months of the year 2000 at the Silsoe Research Institute laboratory. Each point on the graphs is the result of the previous 20 panel threshold n-butanol measurements. The panel thresholds are shown in Fig. 15.4. This shows the accuracy to be slightly biased to the high side of the accepted reference value of 1.6. This is explained by reference to Fig. 15.1, the distribution of threshold values. To date, panel members are selected randomly from our list of qualified assessors, thus the panel is biased towards the higher n-butanol threshold. Closer agreement with the accepted reference value can be achieved by selecting panel members more rigorously. In Fig. 15.5 the record of accuracy and repeatability criteria over the same period shows that the laboratory exceeded the quality criteria of the standard (accuracy criterion shown as &, and repeatability criterion shown as ~). 384 15 Correlating Electronic Nose and Sensory Panel Data

Fig. 15.4 Five-month history of average panel threshold at the Silsoe Research Institute Laboratory

Fig. 15.5 The accuracy and repeatability of the daily measurements of n-butanol with the chosen sensory panel

15.2.6 Assessment of Odor Intensity

The second dimension of the sensory perception of odors, intensity, refers to the perceived strengths of the odor sensation. Intensity increases as a function of concentration. The dependence may be described as a theoretically derived logarithmic function according to Fechner [4]:

S ¼ k log I= ; ð15:1Þ W I0

where: S ¼ perceived intensity of sensation (theoretically determined) I ¼ physical intensity (odor concentration) 15.2 Sensory Panel Methods 385

I0 ¼ threshold concentration kW ¼ Weber-Fechner coefficient. Stevens [5] suggests a power relationship should be applied:

S ¼ k In; ð1Þ where: S ¼ perceived intensity of sensation (empirically determined) I ¼ physical intensity (odor concentration) n ¼ Stevens’ exponent k ¼ a constant.

Which one of these two descriptions applies depends on the method used. To date, no theory has been able to derive the psychophysical relationship from knowledge about the absolute odor threshold of various substances [6]. Odor intensity is measured using this category estimation technique. After determining the odor concentration of the samples, a range of suprathreshold dilutions is presented in random order to panel members. They are required to indicate their perception of intensity at each dilution according to the scale shown in Table 15.1. Intensity scores are obtained from each panel member at each of 12 presentations of suprathreshold dilutions and the average score for each presentation plotted against log10concentration. A linear regression is performed on intensity vs. log10concentration and the line of best fit plotted on the graph. Examples of two such measurements are shown in Figs. 15.6 and 15.7. The fresh 3 landfill material has an intensity of 2.5 (faint to distinct odor) at 0.5 log10ouE ·m , 3 (3.2 ouE ·m ), whereas at the same odor concentration the stale landfill gas has an intensity of only 1.5 (very faint to faint odor). This means that at the same odor concentration the odor from fresh landfill material will be perceived to be the stronger odor. If these data had been obtained from an odor source for which an odor-abatement plant needs to be designed, then it could be that the intensity of a ‘faint odor’, at a complainant’s premises, was considered as the unacceptable limit. In that case the outlet concentration from the abatement equipment would have to be designed so 3 as to deliver an odor with a concentration of less than 2 ouE ·m (fresh landfill 3 material) or 6 ouE ·m (stale landfill gas), respectively, to the nearest complainant.

Table 15.1 Scaling of odor intensity by a human sensory panel

0 No odor; 1 Very faint odor; 2 Faint odor; 3 Distinct odor; 4 Strong odor; 5 Very strong; 6 Extremely strong odor 386 15 Correlating Electronic Nose and Sensory Panel Data

Fig. 15.6 Plot of odor intensity versus odor concentration for volatiles from fresh landfill material

15.2.7 Assessment of Odor Quality

Some useful information about the characteristics of an odor can be obtained if quality assessments are made at a range of dilution ratios close to the panel detection threshold, although these are not included in the standard.

Fig. 15.7 Plot of odor intensity versus odor concentration for volatiles from stale landfill gas 15.3 Applications of Electronic Noses for Correlating Sensory Data 387

One assessment we often carry out is a description of the odor. Our odor panel members are asked to smell the odor at a dilution ratio of between 12 and 100 and indicate, from a choice of descriptors, which comes closest to their perception of the odor. Typically the panel is asked if the odor sample smells like: sewage, fish, rotten cabbage, rotten eggs, bleach, earthy, compost, tarry, smoky, or other. This method is useful for diagnosing if a piece of abatement equipment is changing the odor as well as reducing the concentration. For a food or beverage application such as wines, the requirements for the descriptive terms have to be specific and analytical and not be hedonic or the result of an integrated or judgmental response. Floral is a general but analytical descriptive term, whereas fragrant, elegant or harmonious are either imprecise and vague (fragrant) or hedonic, and judgmental [7, 8], and often an ‘odor wheel’ containing a series of descriptive terms is used to guide the human panel. Each application presents its own specific problems, and appropriate descriptors need to be devised and standardized. For Scotch whisky production for example, the key characteristics arising during production are: estery (the fruity, fragrant, pear-drops aromas that characterize certain malts particularly), phenolic (from woodsmoke to tar, iodine to sea-weed – typified by some malts), aldehydic (leafy, grassy scents, sometimes like Parma violets, often found in various types of malts ) and feinty. The aromas associated with feints are not pleasant – they are notes of sweat, vomit, and rotten fruit – but they give Scotch whisky its character and are essential to the overall flavor. They are present to a greater or lesser extent in all malts [9]. Similar odor descriptor wheels are available for beers, coffee, tea, and many other commodities. The data from such an assessment is usually presented as a histogram of the panels’ response.

15.2.8 Judgment of Hedonic Tone

Hedonic tone is a judgement of the un/pleasantness of the odor. In a similar way to the assessment of the intensity, the panel members are asked to score their perception of the odor on a scale from 1 to 5 at a range of odor concentrations above the odor threshold. A graph similar to the intensity graph can be plotted.

15.3 Applications of Electronic Noses for Correlating Sensory Data

Using an array of sensors together with appropriate data processing may allow mapping of sensory panel attributes to electronic nose data. Multivariate analysis was applied to electronic nose data to correlate sensory panel data for marjoram assessment [10]. Discriminant analysis and neuro-fuzzy treatment of electronic nose and/or color measurement data of marjoram were applied. The aim was to investigate if the judgements of a sensory panel regarding taste, smell, and appearance or genetically determined differences of marjoram samples can be predicted. Frank et al. [11], in co-op- 388 15 Correlating Electronic Nose and Sensory Panel Data

eration with packaging material suppliers and a food manufacturer investigated the quality of different kinds of wrapping foils for chocolate bars using a hybrid modular sensor system (MOSES II). A GC-MS unit connected to a headspace-sampler was used as an analytical reference. A human sensory panel using a sniff-test also qualified all analyzed samples. The different packaging material species could be distinguished in a principal component analysis (PCA). With the aid of a principal component regression (PCR) a correlation between human and technical odor perception was carried out, to determine the spoilage of fish [12], storage of chicken [13], evaluation of tomato quality [14], dairy products [15], and correlation of malodors from sewage [16]. Other sensory attributes may be equally important. For example Benedito et al. [17] investigated methods of improving Mahon cheese texture assessment, where the relationship between instrumental and sensory measurements was sought. For that purpose, 30 pieces of Mahon cheese from different batches and 2 different manufacturers were examined. Textural characteristics at different curing times were evaluated by uniaxial compression, puncture, and sensory analysis. Significant linear correlations were found between instrumental and sensory measurements. A logarithmic model (We- ber-Fechner) fitted data better than a linear one. Pearce and Garner [18, 19] describe a novel method for predicting the organoleptic scores of complex odors using an array of non-specific chemosensors. The application of this method to characterizing beer flavor was demonstrated, which predicted a single organoleptic score as defined under the joint European Brewing Companies/American Association of Brewing Chemists/ Master Brewing Association of the Americas (EBC/ASBC/MBAA) international flavor wheel for beer.

15.4 Algorithms for Correlating Sensor Array Data with Sensory Panels

One problem that needs to be solved is to map responses from a sensor array to analytes of various concentrations (or mixtures) to psychophysical measurements from a human sensory panel, so as to correlate parameters such as quality (in terms of a descriptor) or intensity (in terms of a nonlinear scale). In the ideal situation, we have knowledge of the physical processes underlying the relationship between sensor responses and human panel responses. A theoretical formula can then be used to calculate some meaningful number from the input variables. Usually we do not have this sort of information to hand, however, we can see that there is a relationship there. This is where calibration becomes important. Instead of trying to calculate the theoretical relationship between input and output variables, we make simple assumptions as to the underlying relationship. Using some given examples of input and output variables, we then try to estimate the parameters of this relationship. We can take as X variables the quantities we wish to measure using an electronic nose. Typically, these are more convenient to measure than the values we wish to model. The Y variables are the quantities we wish to predict. These will be the values estimated from the X values using the model. Together with measurement there is some variance, which is a measure of the spread of a variable about its average value. In multi- 15.4 Algorithms for Correlating Sensor Array Data with Sensory Panels 389 dimensional data from sensor arrays the covariance is also important. This is a measure of the similarity of two variables. Variables having high covariance are strongly related to each other. To know the strength of this relationship, we also need to know the variance of the individual variables. For multidimensional data, matrices become important representations of data.

15.4.1 Multidimensional Scaling

Multidimensional scaling (MDS) encompasses a collection of methods that allow us to gain insight in the underlying structure of relations between entities by providing a geometrical representation of these relations [20]. MDS has its roots in two important traditions within psychology. The first is in psychophysics and the other in psycho- metrics. These methods belong to the more general category of methods for multivariate data analysis. MDS can be characterized by the generality of the type of observed relations, which can be submitted to the data analysis on the one hand, and by the specificity of the type of geometrical representation of these relations on the other hand. Whatever kind of relation between a pair of entities that can be translated into a proximity measure, or conversely into a dissimilarity measure, can be considered as possible input for MDS. The choice of a particular type of spatial representation can be considered to be the most important part of the modeling which goes together with the application of a specific MDS-algorithm on the set of proximities. Young and Householder [21] wanted to extend the methodology of unidimensional scaling of perceptual characteristics of stimuli to the simultaneous scaling of several characteristics. Guttman [22] was interested in a less restrictive model than the factor analytic model to represent the relations between several assessment variables. This would allow for a much more systematic way to formulate hypotheses on the underlying structure for assessment variables. The psychophysical approach led to algorithmic developments, which soon came to be known as MDS, while the psychometric approach preferred to label its own production of algorithms under the heading of ‘smallest space analysis’.

We can use the symbol pij to refer to the proximity measure between entities i and j. If a subject has to indicate the perceived dissimilarity between two odors on a rating scale (0 for ‘no difference’ and 10 for ‘maximal difference’), then this rating can be considered to be a reversed measure of the proximity between the two odor stimuli. Or a correlation coefficient between variables i and j can be considered to be a proximity measure for these two variables. The proximities are then represented in a geometrical space, e.g. in a Euclidean space. The distance d between two vectors u and x in a j- dimensional space is given by the formula: ! 1= X l l d ¼ juj xjj ð3Þ j where l ¼ 2 for the Euclidean distance measure commonly used. 390 15 Correlating Electronic Nose and Sensory Panel Data

Three methods of analysis are closely related to MDS: PCA, correspondence analysis and cluster analyis (CA). These are described in detail in Chapter 6.

15.4.2 Regression Methods

Univariate linear regression may be used for establishing a correlation. In its simplest form, this will be familiar as finding the line of best fit through a cloud of points. We assume that the relationship between a single X variable and one Y variable is linear, i.e.

Y ¼ bX þ a ð2Þ

where b is the slope of the line, and a is the intercept at the Y axis. Univariate linear regression estimates the values of b and a by minimizing the sum of squared vertical distances from points to the line. We choose a candidate slope, b and intercept, a. For each recorded (X, Y) pair, we square Y – bX – a and add it to the total. The line having the smallest total is the best-fit line. In practice, calculus gives us a formula for estimating b directly, and thence a, ^b ¼ CovðX; YÞ=VarðXÞ. The indicates that the value is an estimate of b. We can ignore a if we center all our variables before using them. To center each variable, we calculate its average value, and then subtract this value from all sample values. a can be calculated after modeling using the estimated value of b and the subtracted averages. When working with centered data, we can express the linear regression equation for b in matrix form as ðX T XÞ1X T Y. Note that if the variance of X is zero, then we cannot estimate b. This occurs when the X variable has the same value for all values of Y. The matrix form of the linear regression also works for multiple X values, and so in using a multisensor array, the resulting estimate of b is a vector containing the weights applied to the X variables, and this is termed multiple linear regression (MLR). There are many situations when ðX T XÞ1 cannot be calculated, and so some care has to be taken when using MLR. Note that if the number of recorded samples is less than the number of X variables, then collinearity (correlated X variables) is guaranteed to occur. In this situation, the usual solution is to discard variables. The process of selecting variables for MLR is known as stepwise MLR. Because of difficulties in carrying out MLR without prior inspection of the data, methods of visualizing structure in multidimensional data have to be used. PCA provides a method for finding structure in such data sets (See Chapter 6). This method rotates the data into a new set of axes, such that the first few axes reflect most of the variations within the data. By plotting the data on these axes, we can spot major underlying structures automatically. The value of each point, when rotated to a given axis, is called the principal component value. Correspondence analysis is classically used with the aim to visualize the relations (i.e. deviations from statistical independence) between the row and column categories. The unfolding models do the same: subjects (row categories) and objects (column categories) are visualized in a way that the order of the distances between a sub- 15.4 Algorithms for Correlating Sensor Array Data with Sensory Panels 391 ject-point and the object-point reflects the preference ranking of the subject. The measure of ‘proximity’ used in correspondence analysis is the chi-square distance between the profiles. Cluster analysis models are equally applicable to proximity data. The main difference with the MDS models is that most models for cluster analysis lead to a hierarchical structure. Path distances under a number of restrictions approach the dissimila- rities. The path distances are looked for in a way that minimizes the sum of squared errors.

15.4.3 Principal Components Regression

PCA selects a new set of axes for the data. These are selected in decreasing order of variance within the data. They are also perpendicular to each other so that the principal components are uncorrelated. Some components may be constant, but these will be among the last selected. The problem with MLR is that correlated variables cause in- stability. So the strategy adopted is to calculate principal components, throwing away the ones that only appear to contribute noise (or constants), and using MLR on these: this process is known as PCR. Rather than forming a single model, as we did with MLR, we can now form models using more than one component, and decide how many are optimal. If the original variables contained collinearity, then some of our components will contribute only noise. So long as we drop these, we can guarantee that our models will be stable. This method is commonly used to correlate instrumental analyses with human sensory panel data.

15.4.4 Partial Least Squares Regression

The intention in using PCR was to extract the underlying effects in the X data, and to use these to predict the Y values. In this way, we could guarantee that only independent effects were used, and that low-variance noise effects were excluded. This improved the quality of the model significantly. However, PCR still has a problem. If the relevant underlying effects are small in comparison with some irrelevant ones, then they may not appear among the first few principal components. So, we are still left with a component selection problem – we cannot just include the first n principal components, as these may serve to degrade the performance of the model. Instead, we have to extract all components, and determine whether adding each one of these improves the model. This is a complex problem that may be solved using partial least squares regression (PLSR). The algorithm used examines both X and Y data, and extracts components (now called factors) that are directly relevant to both sets of variables. These are extracted in decreasing order of relevance. So, to form a model now, all we have to do is extract the correct number of factors to model relevant underlying effects. A combination of MLR, PLS, factor analysis, and PCR are often used [10, 11, 14, 23–27]. 392 15 Correlating Electronic Nose and Sensory Panel Data

In data from sensor arrays there are often underlying effects. In multivariate calibration, these are called latent variables. A latent variable is one that we do not observe directly, but we can infer its existence by the properties of our observed variables. We can view latent variables in several ways: Assuming that all relationships between latent and observed variables are linear, we can use PCA (if we assume that only the X variables are affected by the latent variables), or PLSR (assuming that both X and Y are affected). If the relationships are thought to be nonlinear, then PCA and PLSR are not ideal, since these assume linearity. If we have an idea of the mathematical form of the nonlinearity, we can try transforming the X and Y variables to linearize them. Failing that, we can use artificial neural networks (ANNs), which use a latent variable model that does not assume linearity.

15.4.5 Neural Networks A parametric regression model usually refers to the regression model where the form of the functional relationship is known (e.g. the linear regression or a specified polynomial regression). Nonparametric regression does not need to specify the form of the unknown functional relationship. The function is modeled using an equation containing unknown parameters but in a way that allows the class of functions that the model can represent to be very broad. Typically the equation, in some functional form, has many unknown parameters, and none of the parameters have any physical meaning in relation to the problem to be solved. Neural networks, including multilayer perceptrons and radial basis function (RBF) networks are nonparametric regression models and these have been described in Chapter 6. Various ANN algorithms can be used to discriminate gases and odors, but the multilayer perceptron network has been adapted for various industrial applications from among many models of ANNs. The development of a learning algorithm, called back-propagation, by Rumelhart et al. [28] revolutionized pattern recognition methodology. An example of the use of neural networks for classification is given by Stetter [29] who used a sixteen-element electrochemical sensor array to identify different grades of wheat, and reported an excellent identification accuracy using the multilayer neural network. For mapping sensor array responses to human sensory panel responses the generic interpolation problem must be solved. The RBF method solves the interpolation problem by constructing a set of linear equations of basis functions [31]. The RBF network makes a linear function space that depends on the positions of pattern vectors according to an arbitrary distance measure. RBF networks can be combined with fuzzy algorithms for enhanced effectiveness in array sensing applications [30].

15.4.6 Fuzzy-Based Data Analysis Fuzzy set theory was introduced in Chapter 6. There are many areas of uncertainty in sensor systems, and fuzzy set theory offers opportunities in many aspects of signal 15.5 Correlations of Electronic Nose Data with Sensory Panel Data 393 processing. These include the evaluation of noisy signals, automatic fault diagnosis, the use of indirect measured values to measure process variables, the automation of measurement and evaluation procedures based on expert knowledge, and the fusion of sensor information in a multisensor environment. The latter application is of importance in the mapping of multisensor data to a human sensory scale. The data provided by the sensors may contain information from several variables, or the information from several sensors is used to provide measurement of a single variable. One way of approaching this is to fuzzify sensor data from each sensor in the array i.e. the numerical value is transformed into a linguistic variable. The results of this step are analyzed by a fuzzy rule base that describes the various relationships between the possible sensor array outputs. The possible outcomes of the fuzzy analysis are then combined and defuzzified to produce the crisp measurement values. This method applied to an odor measurement scenario allows both the ‘quality’ and the ‘intensity’ of the odor to be mapped to sensor responses.

15.5 Correlations of Electronic Nose Data with Sensory Panel Data

At Silsoe Research Institute we use our odor panel selected and monitored as required by the EN13725 when we need to correlate electronic nose responses with human sensory perception. The results of one example of this technique is discussed below.

Fig. 15.8 Sensory panel evaluation of grain from 1998 harvest. Classes are divided into good and bad, a grade mark of 2.5 being the threshold 394 15 Correlating Electronic Nose and Sensory Panel Data

Fig. 15.9 Classification of grain by an RBF network. Of the 50 samples analyzed, the system correctly classified 38, the remainder were not graded by the system as bad or good. In the majority of cases this corresponded with an intermediate rating of the grain i.e. somewhere between good and bad. Each bar G1/8–G5/8 represents good grain samples and B1/8 represent bad samples. The error bars are the standard deviations of repeated presentations to the panellists.

15.5.1 Data from Mouldy Grain

An odor panel selected according to EN13725 was trained at Silsoe Research Institute to evaluate commercial grain samples from 1998. The samples were presented three times in random order and graded into very good, good, bad, and very bad classes, and the panel were asked to mark the samples 1–4 respectively. Grain samples of one variety from the 1998 harvest were used as the training set for the electronic nose this data and the odor panel assessments were the input data for the RBF network described in [31]. Selections of 13 varieties, of wheat from the 1999 harvest were presented as the unknowns. The odor panel classification of the 1998 grain is shown in Fig. 15.8 with classes divided as good and bad, a grade mark of 2.5 being the threshold. The neural network, trained with these grain samples and classes was then used on-line with the electronic nose to classify the 1999 harvest grain. In some instances, discrimination between good and bad grain types has merely been as a result of different moisture content of the grain samples. In our work we could show this is not the case, as illustrated by Fig. 15.9. Of the 50 samples analyzed, the system correctly classified 38, the remainder were not graded by the system 15.5 Correlations of Electronic Nose Data with Sensory Panel Data 395 as bad or good. In the majority of cases this corresponded with an intermediate rating of the grain i.e. somewhere between good and bad. A PCA plot of the data, Fig. 15.10, shows that the good and bad grain fall into distinct groups although the data could be considered as part of one elliptical cluster with subcategories within the cluster with opposite ends representing the best and worst grain samples. The training data for the RBF network was collected with the grain analysis sensor prototype (GASP) three weeks prior to classifying the unknown samples, indicating that sensor drift had minimal affect on the result. The PCA plot is merely a representation of the data for visualization purposes, to give an indication of the ‘sense’ of the data. In this instance the simplistic view of grain as being either good or bad somewhat limits the data and forces the decision making into arbitrary choices. The benefit of the RBF network is that it produces an intermediate result. Further work with an enlarged data set should enable the grain to be re-evaluated against a more robust classification system such as good, intermediate good, intermediate bad and bad, because the grain (as can be seen from the PCA plot) does not instantaneously transform from good to bad but follows a gradual transition from an ‘optimal’ good state through an intermediate stage and on to bad. However, the initial panel-evaluated data set was not large enough to give a reliable enough training set to produce sub-classified data against which real grain samples could be evaluated. Further development will investigate the robustness of the system over long periods of use and across a range of different grain samples. Grain-quality classification into more groups would be a welcome improvement. The principal drawbacks of enlarging the number of grain classifications are the enlarged training set required, obtaining reliably classified training examples, and the time involved in acquiring the data before real samples can be run.

Fig. 15.10 A PCA plot of the training and test data for the neural network. It shows that the good and bad grain fall into distinct groups although the data could be considered as part of one elliptical cluster with subcategories within the cluster with opposite ends representing the best and worst grain samples. The training data for the RBF network was collected with the GASP three weeks before classifying the unknown samples, which indicated that sensor drift had minimal affect on the result 396 15 Correlating Electronic Nose and Sensory Panel Data

Developments of the system to enable characterization of other grain contaminants of interest such as invertebrates are planned. A dedicated system capable of determining the quality of wheat at points of transfer has been developed. The dedicated nature of the system has enabled a more robust and user-friendly system to be developed. The design of the instrument has ensured that classification of large samples can be repeatable. Important factors in the design are temperature and humidity control, consistent presentation of the sample, sensor cleaning and a neural network that is robust and quickly trained.

15.6 Conclusions

In many sensory panel measurements arbitrary scales are used. This makes it difficult to correlate instrumental data, unless some standard can be utilized in both sensory panel measurements and the instrument. Odor measurements no longer need be the arbitrary assessment they have often been perceived to be. Olfactometry to the CEN draft standard, EN13725, ensures a measurable accuracy criterion for the laboratory, and ensures reproducibility of results between laboratories. Once an odor concentration measurement has been made on a sample, then the other three dimensions of odor can be investigated systematically. Measurements of odor intensity can give useful indications of the amount of abatement required, especially when combined with an assessment of hedonic tone. A variety of multivariate analysis methods are applicable to electronic nose data, provided that the questions are clearly defined. The use of neural networks provides a powerful tool when the parameters defining complex relationships between sensor responses and human responses are not well understood. As can be seen from the results, the GASP system is capable of classifying grain at a level equivalent to a trained odor panel, with the implementation of a RBF network. The classifications are independent of grain moisture content. The RBF network is interpolative and allows both qualitative as well as quantitative mapping of sensor array outputs to odor descriptors, intensity, or other sensory parameters. The combination of sensor arrays with multivariate algorithms for mapping complex relationships opens a new route for measuring a percept rather than individual components in a mixture. Incorporation of array sensing technology, signal processing, and computation to produce integrated, low-cost measurement devices is on the horizon, and this will make them increasingly useful in quality control applications in a large number of industries. Thus, industries that rely on human perception and preference for particular types of odors or tastes will now have access to instrumental measurement and control of odor. 15.6 Conclusions 397

References

1 H. Guth, W. Grosch. Journal of Agriculture 15 F. R. Visser, M. Taylor. Journal of Sensory and Food Chemistry 1994, 42, 2852–2866. Studies 1998, 13(1), 95–120. 2 P. Semmelroch, W. Grosch. Journal of 16 R. M. Stuetz, G. Engin, R. A. Fenner. Water Agriculture and Food Chemistry 1996, 44, Science and Technology 1998, 38(3), 331–335. 537–543. 17 J. Benedito, R. Gonzalez, C. Rossello, 3 ‘Air quality – Determination of odor A. Mulet. Journal of Food Science 2000, 65(7), concentration measurement by dynamic 1170–1174. olfactometry’; Draft prEN 13725; European 18 T. C. Pearce, J. W. Gardner. Analyst 1998, Committee for Standardization, editor, 123(10), 2047–2055. CEN: Brussels, 1999. 19 T. C. Pearce, J. W. Gardner. Analyst 1998, 4 G. T. Fechner. Elemente der Psychophysik 123(10), 2057–2066. Breitkopf and Hartel: Leipsig, 1860. 20 J. B. Kruskal, M. Wish. Multidimensional 5 S. S. Stevens. Psychological Review 1957, 64, Scaling. Beverly Hills, California: Sage, 1978. 153–181. 21 G. Young, A. S. Householder. Psychometrika 6 R. L. Doty. Perceptual and Motor Skills 1997, 1938, 3, 19–22. 85(3), 1439–1449. 22 L. Guttman. Psychometrika 1968, 33, 469– 7 A. C. Noble. Abstracts of Papers of the Ame- 506. rican Chemical Society 1998, 216, 130-AGFD. 23 A. HenryBressolette, B. Launay, M. Danzart. 8 C. D. Owens, P. Schlich, K. Wada, A. C. Sciences des Aliments 1996, 16(1), 3–22. Noble. Olfaction and Taste Xii 1998, 855 24 P. J. Hobbs, T. H. Misselbrook, T. R. Cumby. 854–859. Journal of Agricultural Engineering Research 9 M. MacLean. Pocket Whisky Book; Reed 1999, 72(3), 291–298. InternationalBooks Ltd.: 1995. 25 J. E. Parker, G. M. E. Hassell, D. S. Mottram, 10 M. Hirschfelder, A. Forster, S. Kuhne, R. C. E. Guy. Journal of Agricultural and Food J. Langbehn, W. Junghanns, F. Pank, Chemistry 2000, 48(8), 3497–3506. D. Hanrieder. Sensors and Actuators 26 A. K. Thybo, M. Martens. Journal of Texture B-Chemical 2000, 69(3), 404–409. Studies 1998, 29(4), 453–468. 11 M. Frank, H. Ulmer, J. Ruiz, P. Visani, 27 M. C. Zamora, A. M. Calvino. Journal of U. Weimar. Analytica Chimica Acta 2001, Sensory Studies 1996, 11(3), 211–226. 431(1), 11–29. 28 D. Rumelhart, G. E. Hinton, R. J. Williams. 12 G. Olafsdottir, E. Martinsdottir, Nature 1986, 323 533–536. E. H. Jonsson. Journal of Agricultural and 29 J. Stetter. Chemical sensor array: practical Food Chemistry 1997, 45(7), 2654–2659. insights and examples, in Sensors and sensory 13 B. Siegmund, W. Pfannhauser.Zeitschrift systems for an electronic nose, Gardner, J.; fu¨r Lebensmittel-Untersuchung Und-Forschung Bartlett, P., editors; Springer-Verlag: Berlin, A-Food Research and Technology 1999, 1992. 208(5–6), 336–341. 30 W. Ping, X. Jun. Sensors and Actuators 14 F. Sinesio, C. Di Natale, G. B. Quaglia, B-Chemical 1996, 37(3), 169–174. F. M. Bucarelli, E. Moneta, A. Macagnano, 31 P. Evans, K. C. Persaud, A. S. McNeish, R. Paolesse, A. D’Amico. Journal of the R. W. Sneath, N. Hobson, N. Magan. Sensors Science of Food and Agriculture 2000, 80(1), and Actuators B 2000, 69(3), 348–358. 63–71. 399

16 Machine Olfaction for Mobile Robots

Hiroshi Ishida and Toyosaka Moriizumi

Abstract Olfaction often plays an important role in orienting behaviors of animals. Famous examples are ants following pheromone trails marked on the ground and moths tracking aerial pheromone plumes. Inspired by these olfactory-guided behaviors, robotic systems that perform chemical trail following and plume tracking have been developed. In this chapter, the achievements so far are reviewed to demonstrate the current status of this new application of chemical sensor technologies.

16.1 Introduction

The development of electronic noses has seen a successful transfer of knowledge from biological studies to engineering products. The fundamental mechanism of animals’ olfaction, i.e., an array of sensors combined with a pattern recognition algorithm, has become a key element in artificial odor sensing systems. There are, however, other interesting features of olfaction that can be used as models to build engineering systems. One of those features is the close interaction between olfaction and behavior. Olfac- tion often plays an important role in orienting behavior, and many species of animals rely for their survival on this ability. Swimming up or down the gradient of chemical concentration is one of the oldest types of behavior and can be even seen in microorganisms [1]. For some animal species, olfactory cues are far more effective than visual or auditory cues in search for objects such as foods and nests [1]. Olfaction is also used for various types of pheromonal communications [2]. Inspired by these olfactory-guided behaviors, research has been initiated on the use of chemical sensor technologies for navigation of robots. There have been two types of robotic systems developed so far. One is to follow odor trails marked on the ground; its biological model is ants following pheromone trails. The other type of robots track aerial or underwater plumes of chemicals to find their sources. A wide range of animals from bacteria to insects and mammals show this type of search behavior.

Complete understanding of the two types of behavior has not yet been attained. However, biological studies have been gradually revealing the underlying mechanisms, and some of them have been successfully transferred to engineering systems. In this chapter, achievements made to realize mobile chemical sensing systems are reviewed after a brief overview of animal behaviors.

16.2 Olfactory-Guided Behavior of Animals

Among a variety of olfactory-guided behaviors, here we focus on two fundamental search behaviors, plume tracking and trail following. A brief overview of the behaviors is shown below to give useful insights for designing mobile robots. More detailed information can be found in other reviews [1, 3].

16.2.1 Basic Behaviors Found in Small Organisms

Most fundamental forms of olfactory-guided behaviors can be found in microorganisms. Some unicellular eukaryotes and human neutrophil leucocytes swim up to the sources of chemical attractants [1]. They are known to detect the polarity of a concentration gradient by direct comparison of signal intensities at multiple chemoreceptor sites. If similar direct gradient detection is performed using symmetrically placed olfactory organs such as left and right antennae of insects, it is called tropotaxis [1]. Bacteria such as Escherichia coli have a different strategy termed klinokinesis [1]. Since the variation in concentration over their small body length is too small to detect, they compare the stimulus intensity over time. If the detected concentration of an attractant is increasing, a bacterium swims straight. Decrease in concentration means swimming in a wrong direction. However, it can’t tell from the temporal comparison which direction leads to the source. Therefore, a bacterium performs an abrupt turn, and randomly chooses a new direction. Another important class of behavior is klinotaxis in which the concentration gradient is detected by scanning with a single receptor [1]. This yields a much straighter path to the source than klinokinesis [4], but the path may be longer than that of tropotaxis by the length of scanning movements.

16.2.2 Plume Tracking

All three types of behaviors introduced in the previous section lead a searcher to a chemical source provided that smooth and stable concentration gradients are established by molecular diffusion. Although it is true for short-range search in microorganisms, motion of fluid medium (air or water) is almost always more dominant in the 16.2 Olfactory-Guided Behavior of Animals 401 scale of engineering interests than slow molecular diffusion. Therefore, we must face more complex situations. Figure 16.1 shows chemical plumes formed in air and water flows. It is the turbulence of the flows that mainly determines the distributions of chemicals [5]. A chemical substance released from its source trails in the downstream direction, and a number of eddies in the turbulent flow stretch and twist the plume. The result is a complicated, patchy meandering plume. There is no spatially smooth gradient of concentration in these instantaneous images that might guide a searcher to the source. When averaged over several minutes, chemical plumes have continuous concentration gradients. However, this requires too long a time in most of the engineering applications, and is unlikely to be employed by animals. When a stationary sensor is placed in such a chemical plume, a fluctuating signal is obtained [5, 6]. Isolated sharp peaks of concentration are observed when patches of the plume pass over the sensor. As seen in Fig. 16.1, fundamental characteristics of the plumes are the same for both aerial and underwater plumes. Therefore, the discussions on aerial plume tracking can be directly applied to those on underwater plume tracking or vice versa. Animals are able to track the smells of food, mates, nests, etc even in this difficult situation [1, 3]. One of the most intensively studied animals is a male moth tracking sexual pheromone released from a conspecific female [3, 7]. In contrast to the simple chemotactic behavior of bacteria, the fundamental behavioral strategy of moths is upwind flight (anemotaxis) triggered by olfactory cues. When a male moth encounters a patch of a pheromone plume, it turns and progresses in the upwind direction as shown in Fig. 16.2. As long as the moth is flying in the plume, repeated “upwind surges” bring the moth closer to the female. When the male accidentally leaves the plume and the pheromone signal is lost, it starts to fly from side to side across the wind with a gradually broadening scanning area. This behavior is called “casting,” and is effective in relocating the lost plume. Therefore, from the engineering point of

Fig. 16.1 Chemical plumes formed in turbulent flows. (A) Top view of an

aerial plume in a wind tunnel visualized by smoke of TiCl4. (B) Side view of an underwater plume visualized by a dye in fully developed open channel flow of 20 cm depth (photograph courtesy of Drs. Phil Roberts and Don Webster at the Georgia Institute of Technology) 402 16 Machine Olfaction for Mobile Robots

Fig. 16.2 Male moths tracking a sexual pheromone plume released from a female. Male 1 is flying in the plume, and thus repeatedly encounters patches of the plume. This results in iterated upwind surges. Each dot indicates the contact with a patch Male 2 shows casting flight when it has accidentally left the plume. After several scans, the contact with the plume has been regained and it has resumed upwind surges

view, moth’s strategy to achieve a reliable search is twofold: (1) the use of the wind direction combined with the olfactory information and (2) the ability to recover from failures. Extensive work has been also done to reveal the mechanisms underlying the search behaviors of marine animals [8, 9]. There is similarity to a certain extent between the behaviors of terrestrial and marine animals. For example, blue crabs show rheotactic behavior similar to upwind flight of moths [9]; they crawl upstream when they perceive smells of food. However, there is one distinctive difference in the number of sensors used. While a moth uses only a pair of antennae to track a pheromone plume, marine animals seem to make the best use of their chemical sensors, which are spatially distributed over their bodies. A blue crab has chemoreceptors on its eight legs as well as on a pair of antennules, and recent studies suggest all of them are important in tracking odor plumes [10].

16.2.3 Trail Following by Ant

Chemical substances are often used to mark trails or territories [4]. A famous example is an ant laying a pheromone trail on its way back home from a food source. The basic

Fig. 16.3 An ant following a pheromone trail marked on the ground. Concentrations perceived at the left and right antennae are compared and used to turn back to the trail 16.3 Sensors and Signal Processing in Mobile Robots 403 mechanism of trail following by ants is tropotaxis [2] (see Fig. 16.3), as described in Section 16.2.1. Experiments showed that ants do not detect the polarity of the trail. It is said, however, that some animals such as snails and snakes can distinguish one direction from the other [4].

16.3 Sensors and Signal Processing in Mobile Robots

16.3.1 Chemical Sensors

While animals have keen senses for chemical stimuli, sensors for robots with capabilities close to those of animals are not yet available. In the case of gas sensors, a compromise has been made on the rise and decay times. While the response time of an animal’s chemoreceptor is in the order of 100 ms [3, 6], typical gas sensors need several tens of seconds before their responses reach the steady state values. Therefore, the locomotion of robots was slowed down to a few cm/s in most of the studies, such as [11]. When appropriate filters are used to extract rapid changes in concentration from slow sensor responses, the speed of the robots can be increased a few times [12, 13]. Slow sensor response also poses a serious problem in employing a sensor array and a pattern recognition algorithm for odor discrimination. Chemical sensors on mobile robots are exposed to fluctuating concentration in plumes. Since steady-state response is rarely established in this situation, one must use transient sensor response to na- vigate robots. However, the patterns obtained from transient responses are distorted because sensors with different selectivities tend to have different response times. For this reason, e-nose techniques have not been used for mobile robots except for an array of semiconductor gas sensors and a pattern classifier reported by Rozas et al. [14]. The robotic systems developed so far are mostly made to test the ideas in laboratory environments. Therefore, the combinations of target chemicals and sensors were chosen mainly from the ease of handling. The most commonly used combination is alcohol and commercially available tin-oxide gas sensors [11, 13, 15]. QCMs [16, 17] and polymer-based conductometric sensors [18, 19] have been used with camphor, alcohol, and other odorants. Live insect antennae can be also used as odor sensors since they yield voltage differences between their tips and bases according to the intensities of odor stimuli. The measured signal is called EAG (electroantennogram), and robots with antennae cut off from silkworm moths were reported [20–22]. For chemical detection in water, widely used potentiometric sensors, e.g., ion-selective electrodes and ISFETs, suffer from their slow responses. Amperometric microelectrode sensors are promising since fast response comparable to chemoreceptors of animals can be easily achieved [6, 23]. Conductivity sensors can also be used to detect the concentration of ionic solution in fresh water [24]. 404 16 Machine Olfaction for Mobile Robots

16.3.2 Robot Platforms

Most of the mobile chemosensory systems reported so far have used small wheeled robots for their platforms. Legged robots have better maneuverability if they are properly controlled to achieve stable gaits. A six-legged robot mimicking trail following of ants has been reported [25]. Some robots, e.g., the Robolobster [24] and the silkworm moth robot [22], are specifically designed after their model animals. To test the hypotheses on olfactory-guided behaviors of animals, the robots’ sizes and speeds are matched with those of the model animals. The robots are also equipped with chemical sensors that have spatial and temporal resolutions comparable to the chemosensory organs of the model animals. There are several classes of robot configurations as shown in Fig. 16.4. Simple robotic algorithms can be incorporated into a combination of analog and logic circuits shown in Fig. 16.4A [19]. To make the robot perform tropotactic behavior, for example, the logic circuit is wired to turn on the right motor when the left sensor detects an odorant and vice versa. The robot then turns towards the stimulus or move straight if the both sensors are equally stimulated. An on-board microprocessor shown in Fig. 16.4B can accomplish more complicated tasks. As described in the previous section, most of the chemical sensors are slow devices, and because their outputs do not change rapidly, sampling rates of a few Hz are usually sufficient. Therefore, a high-speed processor is not always needed. An 8-bit microprocessor, Motorola 68HC11, with a built-in A/D converter is often used to control a small robot [13, 25]. If more computational power is needed, faster microcomputer boards are available [24]. The flexibility obtained by using micropro- cessors also enables robots to have sensors of other modalities. The sensors that have been incorporated in robots include flow detectors [11, 26, 27] to achieve anemotaxis, a gyro to control turning of the robot [24], and a bump sensor for obstacle avoidance [26]. Another way to accomplish heavy computation is to use a telemetric robot, as shown in Fig. 16.4C. Although an on-board circuitry has to be small enough to fit in a small robot, a fully equipped PC can be used for signal processing in this configuration. Wireless transmitters and receivers are used for the communication between the PC and the robot [27]. Custom-made ASIC chips such as [28] are also promising for signal processing in robotic applications since faster processing can be achieved with smaller circuits.

16.4. Trail Following Robots

16.4.1 Odor Trails to Guide Robots

Automated guided vehicles (AGVs) are a class of industrial mobile robots that follow metal wires buried under the floor and convey parts and materials [25]. The behavior of ants following odor trails implies that a chemical substance can be used as an inex- 16.4 Trail Following Robots 405

Fig. 16.4 Block diagrams of robotic systems. (A) Simplest form of robot. Signals from the left and right sensors are conditioned through the amplifiers and filters. The comparators covert the analog signals into on-off digital values, and the logic circuit is used to map these values to motor commands. (B) Robot controlled by an on-board microprocessor. Signals from sensors are processed in the microprocessor to yield the motor commands. (C) Telemetric robot. The on- board microprocessor acquires sensor signals and transmits them to the PC. After processing the signals, it then sends back motor commands to the on-board processor 406 16 Machine Olfaction for Mobile Robots

pensive alternative to these wires [18]. Odor trails provide higher flexibility since they are easier to lay on the floor. The fundamental constraint is, however, odor trails decay over time as the chemical substance gradually evaporates. Russell proposed other scenarios in which odor trails simplify the tasks to be accomplished by robots [16, 25, 29]. They include: (1) an area coverage task such as cleaning the floor in which odor trails are used to mark the finished area, (2) a cooperative task in which a pathfinder robot lays an odor trail to guide other robots, and (3) an exploring task in which a robot lays an odor trail on its way out and tracks it back to the initial position when the task is accomplished.

16.4.2 Robot Implementations

The most straightforward implementation of ants’ behavior is a robot performing tropotaxis with left and right odor sensors. In the early work of Russell et al., a robot with two QCM sensors was developed [16]. An odor trail is laid by dissolving camphor in an organic solvent and applying the solution to the floor. Although the solvent immediately evaporates, the camphor trail can persist for several hours. The robot successfully traced an odor trail consisting of two straight sections of 50 cm and a sharp turn of 30 degrees between them [16]. There are several variations of this trail following robot. Stella et al. reported a robot equipped with two conducting polymer sensors [18]. Russell later reported a simple robot with a single QCM sensor [25]. In this case, a klinotactic algorithm is employed to follow the edge of a trail. Webb developed a robot with two semiconductor gas sensors (SB-AQ1, Figaro) to investigate the behavioral mechanism of ants [13]. An artificial neural network devised after the tropotactic behavior of ants was employed to control the robot.

16.4.3 Engineering Technologies for Trail-Following Robots

A major disturbance in trail following is external odor confusion [25]. When the left sensor is above the trail and the right sensor is off, the right sensor should ideally show no response. In reality, convection and diffusion bring the odor molecules to the right sensor resulting in a confusing response. This problem can be overcome by using air curtain [25, 29]. Figure 16.5 shows the second generation of air-curtain sensor developed by Russell [25]. Another attempt being made to extend the ability of trail-following robots is to encode useful information into odor trails. Russell proposed several ways of information encoding [25]. A pulse-coded trail as shown in Fig. 16.6 can store information such as the direction of the trail, the identity of the robot laying the trail, and a warning about conditions further along the trail [25]. The information can be retrieved by scanning the trail with a sensor array. This type of system could be also used to obtain odor images evaporating from buried objects such as leaking gas pipes. 16.5 Plume Tracking Robots 407

Fig. 16.5 Air curtain sensor for trail following robots (adapted from [25]). A small fan creates airflow to repel external odor. The part of this airflow moves inward through the QCM sensor to the exhaust, and thus brings the odor to the sensor only from the trail just beneath the sensor

16.5 Plume Tracking Robots

As reviewed in Section 16.2.2, some animals have the excellent ability to locate odor sources by tracking their plumes. In this section, the robotic researches inspired by these animal behaviors are reviewed. The potential applications for the robots that track aerial or underwater plumes include searches for hazardous chemicals, pollutant sources, fire origins, and natural resources. Difficulties in plume tracking come from the random and unstable nature of chemical plumes. While chemical trails marked on the ground never change their shapes, turbulence of flow meanders chemical plumes. They sometimes even change their directions when the direction of air or water flow shifts. Therefore, occasional failures are almost inevitable in the tracking of plumes. As revealed from the following sections, the keys for successful tracking not only lie in the algorithms to track plumes but also in the fail-safe mechanisms to relocate the lost plumes in case of failure. 408 16 Machine Olfaction for Mobile Robots

Fig. 16.6 Robotic system with eight QCM sensors to detect coded trails (adapted from [25]). While the two leftmost sensors are used to trace the continuous guide path, the others are used to detect pulse-coded trails

16.5.1 Chemotactic Robots

The research on plume tracking robots started with purely chemotactic robots. Sandini et al. developed a robot with two semiconductor gas sensors (TGS800, Figaro) for gas leak detection [15]. This robot performs tropotactic search as shown in Fig. 16.7A. A similar tropotactic algorithm was also employed in the robot with two conductometric polymer sensors developed by Kazadi et al. [19]. As mentioned in Section 16.3.2, robots can be used as tools for biologists to investigate mechanisms of animal behaviors. Consi et al. developed an underwater robot based on American lobsters that crawl on the bed of oceans tracking food smells [24, 30]. The robot was equipped with conductivity sensors that mimicked the size and the separation of lobsters’ antennules. Kuwana et al. reported a small robot mimicking a male silkworm moth that walks to a female releasing the sexual pheromone [22]. Two pheromone sensors made of moths’ antennae were used for the robot. All the robots introduced above employ similar tropotactic algorithms. However, there are differences in how these robots react when no chemical signal is perceived. Due to the time-varying nature of a chemical plume, a robot may sometimes leave the plume by chance. Since there is no signal outside the plume, a simple tropotactic robot continues to move straight and never returns. Therefore, it is important to incorporate algorithms to relocate the lost plume. Several algorithms including 16.5 Plume Tracking Robots 409

Fig. 16.7 Chemotactic robots. (A) A robot tracking concentration gradients detected by the comparison of the left and right chemical sensor outputs. (B) Various algorithms to relocate a chemical plume when the robots accidentally lose contact with it. Robot 1 is programmed to back up when neither sensor detects chemical. Robot 2 performs random walk. Robot 3 mimics the behavior of a male silkworm moth. When one of the sensors is stimulated, the robot surges in that direction to track a plume. When the chemical signal is lost, the robot performs zigzag walk and circling to relocate the lost plume

backing up [11, 30], random walk [15], and zigzag walk embedded in a recurrent artificial neural network [21] have been proposed as fail-safe mechanisms in chemotactic search (see Fig. 16.7B). The limit in applying purely chemotactic strategies lies in the structure of chemical plumes. As seen in Fig. 16.1, local and instantaneous gradients fluctuate significantly. Those fluctuations often mislead a chemotactic robot resulting in the circuitous movement of the robot in Fig. 16.7A. The concentration gradient along the plume centerline is extremely small except in the vicinity of the source. Therefore, when a robot has started a search from a distant place, the success rate can be low [27, 30]. One way to overcome this problem would be to use a swarm of cooperative robots [15, 31]. San- dini et al. proposed to use multiple tropotactic robots with a communication link to exchange information among the robots nearby [15]. It was reported that the robots successfully gathered around the source location by programming each robot to be attracted to the robot signaling the higher concentration. Another way to overcome the weakness of chemotaxis is to use flow direction in navigating a robot, which is described in the next section. 410 16 Machine Olfaction for Mobile Robots

16.5.2 Olfactory Triggered Anemotaxis

As we have seen in the behavior of moths, the direction of flow carrying odor molecules is a strong directional cue in searching their source. We have developed a mobile robot equipped with both gas and airflow sensors to incorporate the keys in moths’ behavior into a robotic system [11, 27]. Wind direction with an accuracy of 458 is obtained in this robotic system for a wind velocity of 5–30 cm/s from the response pattern of the four thermistor airflow sensors (F6201-1, Shibaura Electronics). The two orthogonal components of the concentration gradient are also measured as the response differences between the two pairs of diagonally aligned semiconductor gas sensors (TGS822, Figaro). This robot tracks a chemical plume as shown in Fig. 16.8A. While tracking the plume, the gas sensors are used to keep the robot heading towards the plume centerline. Due to the random nature of chemical plumes, however, a fail-safe mechanism to relocate the lost plume is again required to yield a high success rate. Although the algorithms for chemotactic robots shown in

Fig. 16.8 Anemotactic search algorithm triggered by chemical cues. (A) The robot proceeds obliquely upwind to the side with higher concentration. This oblique movement keeps the robot close to the plume centerline. When the robot accidentally leaves the plume, it performs side-by-side scanning similar to moths’ casting. (B) Experi- mental result of the robot tracking a chemical plume at 1 cm/s [27]. A nozzle releasing ethanol vapor at 75 ml/min was placed in a clean room where an air conditioner was producing wind of about 30 cm/s. The solid lines show the track of the robot when moving upwind, and the dotted lines show that when the robot was casting 16.5 Plume Tracking Robots 411

Fig. 16.7B can be also applied here, the robot with airflow sensors can again employ the wind direction as a useful directional cue. When a moth has lost contact with an odor plume, it scans across the wind with gradually increasing width as shown in Fig. 16.2. This casting flight is a reasonable strategy in relocating the lost plume since the possibility of hitting into a plume elon- gated in the wind direction is maximized when a searcher travels across the wind direction [32]. As shown in Fig. 16.8B, this casting behavior was successfully incorporated into the robot. It quickly recovered the plume after a single scan and resumed tracking the plume. Russell et al. reported a mobile robot equipped with a custom-made wind vane [26]. Although the robot performed a similar anemotactic search, a klinotactic strategy with a single QCM gas sensor was employed to adjust the robot position across the wind direction. This simplifies the robot structure, and eliminates the need to match the sensitivities of gas sensors used. However, there is a trade-off between the simplicity and the measurement time. In klinotaxis, the robot needs to scan left and right to make a comparison.

16.5.3 Multiphase Search Algorithm

Olfactory triggered anemotaxis described in the previous section shows its maximum performance in uniform flow fields, which we encounter in wind tunnels or flumes to test robots and animals. However, plume-tracking robots may face more complicated flow fields in real applications. In a domestic or industrial building, for example, the main source of wind is an air conditioner and a robot often encounters winds from multiple air-supply openings simultaneously. One way to tackle a difficult task is to decompose it into easier subtasks. In order to cope with winds from multiple directions, a multiphase algorithm shown in Fig. 16.9 was devised [27]. When a wind from another direction is merging into a side of a chemical plume, the wind direction in this merging area becomes unstable. There- fore, care should be taken to employ the anemotactic strategy. When detected concentration is low, the robot might be in this merging area where unstable winds often direct an anemotactic robot in wrong directions. The robot should search for higher concentration by a chemotactic strategy. Anemotaxis can be employed only when high concentration is detected. In this case, the robot is thought to be in the center of the plume where only the wind from the source direction exists. This change in strategy can be made with a pre-defined threshold in concentration. In Fig. 16.9B, however, the change was made when one strategy makes no significant progress for 60 s. This ensures timely changes in strategies even when the pre-defined threshold is inappropriate. To accomplish a fully autonomous search in real applications, there still remain many questions, including how to locate a plume for the first time in the absence of any chemical signals, and how to decide when the odor source has been located so as to terminate the search. In the multiphase algorithm described above, the robot 412 16 Machine Olfaction for Mobile Robots

Fig. 16.9 Multiphase search algorithm to cope with winds from multiple directions. (A) The robot first tracks concentration gradients to escape from the area of unstable wind. When the robot reaches the center of the plume, high concentration of the target chemical is detected. The robot then tracks the plume in the upwind direction. (B) Experimental result [27]. The multiphase algorithm was tested in the same clean room as in Fig. 16.8B. The starting position of the robot was moved to the side of the ethanol plume where the wind from another direction was merging. Thick lines show the path of the robot tracking the concentration gradient, and thin lines show that in tracking the plume in the upwind direction

sits still until a certain level of gas is detected. This is based on the scenario that the robot is placed in a room as a replacement for conventional gas alarms and that it needs to save its energy until a leakage actually occurs. If the robot is brought to the place where a leakage is detected, more active strategies should be employed. Moving across the wind [26] as shown in Fig. 16.10 would be a choice. This is known to be the most efficient strategy in finding a plume when multiple sources are dispersed in a uniform wind field [32]. As shown in Fig. 16.10, a plume extends downstream from each source to a finite length until the concentration is diluted below the detection limits of the sensors. The robot crossing the flow would eventually hit into one of the plumes in the field although it might have passed by some of them on its way. When the characteristics of the source, such as its shape and size, are known, they can be used for identifying it. Russell et al. proposed the use of a bump sensor for both obstacle avoidance and declaration of the source [26] (see Fig. 16.10). It may not be necessary for a robot to go all way up to the source to declare the source location. When a robot scans the chemical plume on its way to the source, the concentration 16.6Other Technologies in Developing Plume Tracking Systems 413

Fig. 16.10 Multiphase algorithm proposed for a robot equipped with a bump sensor (adapted from [26]). The robot is first made to move across the wind until it hits into a chemical plume. The robot then starts tracking the plume in the upwind direction. When the robot hits an obstacle while tracking the plume, it circles around the obstacle by using the bump sensor. If the target gas is detected at the upstream edge of the obstacle, the robot resumes upwind search. If not, it can be declared that the obstacle is the source of the target gas

change tracing the plume shape is observed. When an appropriate plume model is prepared, the source location can be found by curve-fitting the model to the observed concentration change and extrapolating the curve to the source location [33].

16.6 Other Technologies in Developing Plume Tracking Systems

16.6.1 Olfactory Video Camera

An array of chemoreceptors on the eight legs of a blue crab might be able to detect information that is not accessible with the pair of antennae on a flying moth. The “olfactory video camera” shown in Fig. 16.11 is an engineering realization of such spatially distributed sensor arrays [17, 34]. As described in Section 16.2.2, the sensor responses observed in a chemical plume are highly intermittent, and this intermit- 414 16 Machine Olfaction for Mobile Robots

Fig. 16.11 Schematic diagram of a gas/odor flow imaging system termed “olfactory video camera” [17, 34]. It consists of an array of 5 Â 5 gas sensors with 1 cm spacing, and presents the visualized image of instantaneous concentration distribution over the small array

tency enables to track patches of the plume. When a patch passes over the array, the flow direction and speed can be determined from the visualized image. When the array is moved tracking the visualized plume reversely, it eventually approaches the source. As seen in Fig. 16.10, the source location can be determined to be the point where the target gas is detected on its downstream edge but not on its upstream edge. This can be easily judged from the visualized image when the array is placed over the source location [35].

16.6.2 Odor Compass

Marine crustaceans flick their antennules, and terrestrial vertebrates show sniffing behavior. These actions modulate the reception of chemical signals at the animals’ sensors [36]. An interesting example of this signal modulation is wing fanning of a male silkworm moth tracking a pheromone plume. Mimicking this mechanism, a sensing probe consisting of two semiconductor gas sensors (TGS822, Figaro) and a small fan was devised and termed an “odor compass” [37]. Experiments showed that the effect of the fan is significant in obtaining directional cues (see Fig. 16.12). The direction toward the source can be found by rotating the compass and determining the direction where the two sensor responses match. This sensing mechanism can be extended to a three-dimensional search by adding two vertically aligned gas sensors and rotating the compass head three dimensionally [38]. It was shown from the experiments that this system is effective in searching odor sources around obstacles where complicated three-dimensional fields are formed (see Fig. 16.13) [38, 39]. 16.6Other Technologies in Developing Plume Tracking Systems 415

Fig. 16.12 Mechanism of odor compass consisting of two gas sensors and a small fan [37]. Since the gas concentration gradient along the plume axis is small, no significant difference in the left and right sensor responses is observed when the fan is turned off. When it is turned on, however, the plume is deformed by the airflow and the sensor closer to the source shows a stronger response

Fig. 16.13 Result of plume tracking using a three-dimensional odor compass [39]. A nozzle releasing ethanol vapor at 300 ml/min was successfully located from behind a large obstacle. The compass was iteratively moved in the estimated source direction by 30 cm. (A) Perspective view. (B) Side view 416 16 Machine Olfaction for Mobile Robots

16.7 Concluding Remarks

Various aspects of the mechanisms underlying olfactory-guided behaviors of animals have been transferred to robotic platforms. Animals show a variety of behaviors each of which is optimized for the habitat of that species, and there seems to be no single engineering implementation that can be used in every situation. Future work is needed to establish design strategies that can tell us which type of system is best suited for the problem of current interest and how we can determine its design parameters. Development of chemical sensors for mobile robots is also an important subject for future work. Current chemical sensor technologies were originally developed for stationary sensing systems, and their limitations such as long response times have been strong constraints in the development of mobile biomimetic robots. Chemical sensors tailor-made for mobile robots would further expand the abilities of chemosensory robots and open up new directions in the sensor technologies.

Acknowledgement We gratefully acknowledge that the ideas presented in this article came from the continuing collaborative work with Dr. Takamichi Nakamoto. Enlightening discussions with a chemist (Dr. Jiri Janata), fluid mechanical engineers (Drs. Philip Roberts and Donald Webster), and biologists (Drs. Marc Weissburg, David Dusenbery, and Troy Keller) are also acknowledged. We thank Dr. R. Andrew Russell for giving us the permission to quote his interesting work.

References

1 W. J. Bell, T. R. Tobin. Biol. Rev. 1982, 57, 10 T. A. Keller, M. J. Weissburg. Abstr. Aquatic 219–260. Sciences Meeting Amer. Soc. Limnol. Oceanogr. 2 W. C. Agosta. Chemical Communication: 2000. The Language of Pheromones, Scientific 11 H. Ishida, K. Suetsugu, T. Nakamoto, American Library, New York, 1992. T. Moriizumi. Sensors and Actuators A 1994, 3 E. A. Arbas, M. A. Willis, R. Kanzaki. 45, 153–157. In Biological Neural Networks in Invertebrate 12 T. Nakamoto, T. Yamanaka, H. Ishida, Neuroethology and Robotics (Eds.: R. D. Beer, T. Moriizumi. Meeting Abstr: Electrochem. R. E. Ritzmann, T. McKenna), Academic Soc. 1996, 96–2, 1163. Press, San Diego, 1993, Chapter VIII. 13 B. Webb. Neural Networks 1998, 11, 4 D. B. Dusenbery. Sensory Ecology, 1479–1496. W. H. Freeman and Company, New York, 14 R. Rozas, J. Morales, D. Vega. Fifth Inter- 1992. national Conference on Advanced Robotics 5 J. Murlis, J. S. Elkinton, R. T. Carde´. 1991, 1730–1733. Annu. Rev. Entomol. 1992, 37, 505–532. 15 G. Sandini, G. Lucarini, M. Varoli. Proc. 6 P. A. Moore, J. Atema. Biol. Bull. 1991, 181, 1993 IEEE/RSJ Int. Conf. Intelligent Robots 408–418. and Systems 1993, 429–432. 7 T. D. Wyatt. Nature 1994, 369, 98–99. 16 R. Deveza, D. Thiel, A. Russell, A. Mackay- 8 J. Atema. Biol. Bull. 1996, 191, 129–138. Sim. The International Journal of Robotics 9 M. J. Weissburg, R. K. Zimmer-Faust. Research 1994, 13, 232–239. J. Exp. Biol. 1994, 197, 349–375. 16.7 Concluding Remarks 417

17 T. Nakamoto, T. Tokuhiro, H. Ishida, 28 S. Kawamura, K. Matsuyama, T. Nakamoto, T. Moriizumi. Technical Digest of Transducers T. Moriizumi. Technical Digest of the 17th ’99 1999, 1878–1879. Sensor Symposium 2000, 321–324. 18 E. Stella, F. Musio, L. Vasanelli, A. Distante. 29 R. A. Russell. IEEE Robotics and Automation Proc. 1995 Intelligent Vehicles Symposium Magazine 1995,2,3–9. 1995, 147–151. 30 F. W. Grasso, T. R. Consi, D. C. Mountain, 19 S. Kazadi, R. Goodman, D. Tsikata, D. J. Atema. Robotics and Autonomous Systems Green, H. Lin. Autonomous Robots 2000,9, 2000, 30, 115–131. 175–188. 31 V. Genovese, P. Dario, R. Magni, L. Odetti. 20 Y. Kuwana, I. Shimoyama, H. Miura. Proc. Proc. 1992 IEEE/RSJ Int. Conf. Intelligent 1995 IEEE/RSJ Int. Conf. Intelligent Robots Robots and Systems 1992, 1575–1582. and Systems 1995, 530–535. 32 D. B. Dusenbery. J. Chem. Ecol. 1989, 15, 21 Y. Kuwana, I. Shimoyama. The International 2511–2519. Journal of Robotics Research 1998, 17, 33 H. Ishida, T. Nakamoto, T. Moriizumi. 924–933. Sensors and Actuators B 1998, 49, 52–57. 22 Y. Kuwana, S Nagasawa, I. Shimoyama, 34 H. Ishida, T. Yamanaka, N. Kushida, R. Kanzaki. Biosensors and Bioelectronics T. Nakamoto, T. Moriizumi. Sensors and 1999, 14, 195–202. Actuators B 2000, 65, 14–16. 23 T. Kikas, H. Ishida, P. J. W. Roberts, 35 H. Ishida, T. Nakamoto, T. Moriizumi, D. R. Webster, J. Janata. Electroanalysis 2000, T. Kikas, J. Janata. Biol. Bull. 2001, 200, 222– 12, 974–979. 226. 24 T. R. Consi, J. Atema, C. A. Goudey, J. Cho, 36 M. A. R. Koehl. Mar. Fresh. Behav. Physiol. C. Chryssostomidis. Proc. 1994 Symp. 1996, 27, 127–141. Autonomous Underwater Vehicle Technology 37 T. Nakamoto, H. Ishida, T. Moriizumi. 1994, 450–455. Sensors and Actuators B 1996, 35, 32–36. 25 R. A. Russell. Odour Detection by Mobile 38 H. Ishida, A. Kobayashi, T. Nakamoto, Robots, World Scientific, Singapore, 1999. T. Moriizumi. IEEE Trans. Robot. Autom. 26 R. A. Russell, D. Thiel, R. Deveza, 1999, 15, 251–257. A. Mackay-Sim. Proc. 1995 IEEE Int. Conf. 39 H. Ishida, T. Nakamoto, T. Moriizumi. on Robotics and Automation 1995, 556–561. Sensors Update 1999, 6, 397–418. 27 H. Ishida, Y. Kagawa, T. Nakamoto, T. Moriizumi. Sensors and Actuators B 1996, 33, 115–121. Part D Applications and Case Studies

17 Environmental Monitoring

H. Troy Nagle, Ricardo Gutierrez-Osuna, Bahram G. Kermani, Susan S. Schiffman

Abstract In this chapter, we review some of the previous proof-of-principle work done in this field. Examples of water, land, and air monitoring experiments are examined. Four case studies are then presented. The first three demonstrate the ability of the electronic nose (e-nose) to classify odors from animal confinement facilities (odor source determination, odorant threshold detection, and odor abatement evaluation). The fourth case study demonstrates that the e-nose can differentiate between five types of fungi that commonly diminish indoor air quality in office buildings and industrial plants. Finally, we conclude that environmental monitoring is a promising application area for e-nose technology.

17.1 Introduction

The field of environmental monitoring encompasses a broad range of activities. Con- tamination of the environment can occur not only by dumping wastes in water, land, and air, but also by generating noise in the audio and communications frequency ranges. Sensing systems have been developed for all of these applications. In this chapter, we focus on efforts to employ an electronic nose (e-nose) to monitor airborne volatile organic compounds that are released when waste products are dumped in water, land, or air.

17.1.1 Water

Water quality is threatened when agricultural and industrial concerns allow their waste products to seep into groundwater or to flow into streams or rivers. The e-nose can be used in these applications on samples of the effluent. The headspace of such samples can be tested with an e-nose system, on-line or off-line, to establish the time-course of

emission profiles. Boreholes can also be employed to collect samples to test groundwater contamination. Several research groups have studied the e-nose as an instrument for monitoring water quality. Some teams have utilized metal-oxide sensors for monitoring. Baby et al. [1] used the MOSES II e-nose to measure contaminating residues of insecticides and products from leather manufacture that are often offloaded into streams and rivers. Dewettinck et al. [2] employed an e-nose consisting of 12 metal-oxide sensors to monitor volatile compounds in the effluents of a domestic wastewater treatment plant over a 12-week period. Correlation between the relative overall e-nose output and the parameter ‘volatile suspended solids’ was good, indicating adsorption of volatile organic compounds (VOCs) onto the organic particles. This study also concluded that the e-nose has promise in wastewater monitoring applications. In another study by the same group, Van Hege et al. [3] explored the application of evaporative technology as an alternative desalination technique for wastewater treatment plant effluents. Evaporation completely removed most inorganic and organic contaminants. An e-nose was employed to monitor changes in odor quality and intensity due to volatilization of the VOCs present in the effluent. Conducting polymers have also been used to analyze wastewater. Di Francesco et al. [4] studied the use of an e-nose with conducting-polymer sensors and fuzzy-logic- based pattern recognition algorithms to test wastewater samples. In other work an e-nose with 12 polypyrrole conducting-polymer sensors was used to monitor quies- cent sewage liquors at three wastewater treatment plants over an 8-month period [5–7]. The e-nose was evaluated as a replacement for human panels in monitoring liquid wastewater samples, wastewater odor, and tainting compounds in water sup- plies. The study revealed that a strong linear relationship is expected for site/ source-specific odor samples. The study also showed that low levels of organic pollutants can be detected by monitoring water samples with the e-nose. In addition, the study suggested that it might be feasible to use an e-nose to monitor and/or control the biochemical activities of a wastewater treatment process. More recently, Bourgeois and Stuetz [8] reported the use of a similar sensor array to analyze wastewater samples

sparged with N2 gas in a temperature-controlled flow-cell. The headspace gas was then supplied through a temperature-controlled transfer line to the conducting-polymer sensors. They concluded that an externally generated headspace gas could be used to monitor changes in wastewater quality, and could provide a simple non-invasive technique for on-line monitoring of wastewater. Continuing this avenue of research, Stuetz et al. [9] and Bourgeois et al. [10] examined the use of real-time sensors and array systems for monitoring global organic parameters such as biochemical oxygen demand and total organic carbon. Stuetz et al. [9] and Stuetz [11] compared the odor profiles of sewage liquids with corresponding biochemical oxygen demand and total organic carbon measurements, and determined that a number of different wastewater quality relationships could be formulated from the e-nose analysis of a sewage liquid. They concluded that the organic content of wastewater, as well as the potential of wastewater to produce nuisance odors, could be predicted from a single headspace analysis of a sewage liquid using a sensor array. 17.1 Introduction 421

Di Natale et al. [12] used a sensor array of ion-sensitive electrodes to analyze polluted water. The sensor array was processed using chemometrics, non-linear least squares and neural networks. The devices that use sensor arrays to test liquid samples are called electronic tongues rather than e-noses. See Chapter 11 for more information on electronic-tongue devices. Gardner et al. [13] and Shin et al. [14] developed a system for detecting cyanobacteria (blue-green algae) in potable water. The e-nose system, employing an array of six commercial gas sensors, was able to detect 100 % of the unknown toxic cyanobacteria using a multi-layer perceptron (MLP) neural network. The results showed the potential for a neural network-based e-nose, as opposed to more traditional instruments such as liquid chromatography or optical microscopy, to test the quality of potable water.

17.1.2 Land

Land contamination by toxic and radioactive materials is a chief concern in many countries around the world. Garbage waste dumps are problems everywhere. The e-nose has applications in this arena as well. Borehole samples can be placed in sample containers to generate headspace VOCs. Adding specific reagents to some of these samples can accelerate the generation of VOCs and improve the sensitivity of the e- nose instruments. This is an emerging area for e-nose instrumentation and there should be considerable future growth in this segment of the e-nose market. There have been few research studies in this area. One example of note is Biey and Verstraete [15]. They investigated the use of a 5-W UV lamp, generating ozone for seven hours per day, to reduce the odors produced by the decomposition of kitchen and vegetable waste. An Alpha M.O.S. FOX 3000 e-nose was used to measure odor levels before and after treatment. They concluded that the UV treatment did indeed reduce the odor levels, and thus would be useful in summer, or all year around in warm climates.

17.1.3 Air

Air quality has been the primary target of e-nose research projects in environmental monitoring [16, 17]. The e-nose can monitor odorous emissions at their source, such as paper mills, animal production sites, power-plant stacks, vehicle exhaust pipes, compost facilities, wastewater treatment plants, animal rendering plants, paint shops, printing houses, dry cleaning facilities, and sugar factories. The e-nose also holds promise for monitoring emissions from near-source or remote locations in a populated area. Currently, available sensor arrays have not proven efficient at remo- tely located sites, owing to their lack of adequate sensitivity to many of the offending VOCs in odorant mixtures. However, e-nose measurements made at the source could serve as input to mathematical emission dispersion models that can predict VOC concentrations at remote locations given accurate meteorological data for a specific geo- 422 17 Environmental Monitoring

graphic location [18]. As sensor-array technology improves, the measuring of odorous VOCs at remote locations will become a significant market for hand-held e-nose devices (see Chapter 9). Although in most cases annoying atmospheric emissions do not menace public health, they do greatly reduce the quality of life [4, 19]. Measuring these odors at the site of complaints is very difficult due to the transient nature of the odorous events. The e-nose offers the promise of being able to make accurate and repeatable measurements of odor profiles at sites of complaint. These e-nose measurements can be correlated with those of human panels in order to calibrate the odor quality and perception scales [20] (see Case Study 3 in this chapter). Now we discuss several examples of the application of the e-nose to monitoring air quality. Odor abatement and control is a major issue facing municipal sewage treatment facilities. The odors emitted from these facilities can be monitored by an e-nose. Gostelow et al. [21] reviewed various sensory, analytical, and e-nose methods for monitoring sewage facility emissions. Stuetz et al. [22, 23] employed a Neotronics NOSE to investigate emissions from ten sewage treatment facilities. Odor levels measured by the NOSE unit were compared with those of an independent human panel, measured in odor units per cubic meter. The effect of biofilters was also considered. A linear relationship was observed between the NOSE measurement and the human panel results for data at each independent site. At low odor levels, the results were also extended to the multiple site case. Hydrogen sulfide concentrations, although commonly used as a measure of odor strength, were also compared with the human panel results and were found not to be a reliable marker compound for measuring sewage odor concentrations. The perception of the quality of indoor air by building inhabitants is addressed by Schreiber and Fitzner [24, 25]. Delpha et al. [26, 27] investigated the use of an e-nose using metal-oxide TGS sensors for the detection of a leaking refrigerant gas (Forane R134a) in an air-conditioned atmosphere. First the researchers showed that the time response of the TGS sensors to Forane R134a gas in humidity varying from 0 to 85 % could be represented by a double exponential model. The authors then demonstrated the ability to identify the target gas by discriminant factorial analysis, even for cases in which the relative humidity or the gas temperature were outside the range of the training database. In a similar study, Sarry and Lumbreras [28] investigated the detection of carbon dioxide, Forane R134a, or their mixtures, without a sensor dedicated to carbon dioxide measurement. They used an array of five tin-dioxide sensors. Discriminant factorial analysis was used for processing the data. The authors report a reliable system can be designed for this application. Ramalho [29] analyzed the characteristics of indoor paints and their effect on perceived indoor air quality. Ten different indoor paints were presented to an e-nose and to 13 trained panelists. Significant differences among panelists were found, whereas the sensors displayed little difference. However, some similarities were found between some sensors and individuals. Feldhoff et al. [30] compared the ability of an Alpha M.O.S. FOX 4000 and a LDZ Laboratory Smart Nose GA 200 to differentiate between twenty Diesel fuels from three different refineries. The authors reported that both units were able to correctly identify 17.1 Introduction 423 the production site of the 20 samples. However, the Smart Nose uses a mass spectrometer and its data were easier to obtain and were more reproducible. In a similar study, Lauf and Hoffheins [31] illustrated that a selected array of chemical sensors can produce unique signatures for many aviation and automotive fuels. Patterns for aviation fuel are readily identified by visual inspection. The differences among automotive fuels with different octane ratings are subtle but perceptible. Gasohol mixtures have strikingly different signatures from pure gasoline. The results indicate that an e-nose can distinguish between various classes of petroleum-based fuels. Automotive ventilation may also be monitored and controlled by an e-nose. Menzel and Goschnick [32] investigated methods for improving the time response of an e-nose instrument intended for on-line discrimination applications. Their method combined the classification of the steady-state and transient response via time-series analysis. Rapid signal transients were detected by appropriate digital filters, while steady-state signals were classified by standard statistical methods. To illustrate the method, they investigated automatic control of the ventilation flap of an automobile. Steams of bad air were detected in one to two seconds. The error in the detection of pollutants was reduced from the original 25 % to only 10 % for their new method. E-nose systems have also been studied for detection of hazardous materials and gases. For example, Hopkins and Lewis [33] investigated the use of arrays of carbon-black/organic-polymer composite chemiresistive vapor detectors for detecting nerve agents. Chapter 23 of this handbook is devoted to the detection of explosives. Odorousemission fromanimalproduction facilitieshasbeenextensively studiedover the last few years. We present several case studies in this area later in the chapter. Other research groups have also studied this important problem. Hobbs et al. [20] correlated e- nosemeasurementsofpigmanureodorstothoseofahumanpanel.Fouroftheprinciple odorous compounds in pig manure were selected for the study. Thirty-one different mixtures of hydrogen sulfide, 4-methyl phenol, ammonia, and acetic acid were used to simulate the livestock waste odor. A radial-basis-function neural network was used for signal processing.Predictionsusinga linearregression modelwereon average20 %less than observed values. The authors reported that this approach using the four main odorants is appropriate for determining the odor concentration of pig manure. An e-nose can frequently be employed to identify specific VOCs and mixtures of VOCs. Hudon et al. [34] compared the effectiveness of three different e-nose instruments in measuring the odor intensity of n-butanol, CH3COCH3, and C2H5SH, and binary mixtures of n-butanol and CH3COCH3. Two commercial e-nose systems, the AromaScan A32S (conducting-polymer sensors) and the Alpha M.O.S. Fox 3000 (metal-oxide sensors), and an experimental unit with Taguchi-type tin-oxide sensors were employed. The e-nose measurements were processed using linear regression analysis and neural networks. Very strong correlation (q ¼ :99) was obtained between the sensory data and the two commercial units when using neural network analysis. In a related study, Negri and Reich [35] used an e-nose with commercially available tin- oxide sensors to analyze a mixture of gases containing carbon monoxide, ethanol, methane and/or isobutane. They modeled the theoretical response function of the array and designed a pattern recognition scheme for the simultaneous identification of a given gas and its concentration in the mixture. 424 17 Environmental Monitoring

The growth of bacteria and fungi on organic matter generates a broad range of volatile organic compounds and fixed gases. Wesse´n and Schoeps [36] and Sunesson et al. [37] showed that the presence of certain VOCs could be used as an indicator of the presence and of the identity of microorganisms. Holmberg [38], in a dissertation at Linko¨ping University in Sweden, used an e-nose with 15 sensors to classify five types of bacteria (Escherichia coli, Enterococci sp., Proteus mirabilis, Pseudomonas aeruginosa, and Staphylococcus saprophyticus). The 15 sensors included nine metal-oxide semiconductor field-effect transistors, four Taguchi-type devices, one carbon dioxide sensor, and one oxygen monitor. The volatile compounds generated by the bacteria were sampled from agar plates. The results suggested that this e-nose could successfully classify Escherichia coli and Enterococci sp. but was less successful with the other bacteria. Gardner et al. [39] used an e-nose that contained six commercial metal-oxide sensors, a temperature sensor, and a humidity sensor to predict the class and growth phase of two types of bacteria, Escherichia coli and Staphylococcus aureus. The six sensors were designed to detect hydrocarbons, alcohols, aldehydes/heteroatoms, polar molecules, and non-polar compounds. The best mathematical model correctly identified 100 % of the unknown S. aureus samples and 92 % of the unknown E. coli samples. Other studies have also found that bacteria can be discriminated using an e-nose. In an evaluation of seven bacterial strains, Vernat-Rossi et al. [40] were able to correctly discriminate 98 % of a training set with a cross-validation estimate (test set) of 86 % using six semiconductor gas sensors. Studies at AromaScan PLC [unpublished data from Dr. Krishna Persaud] showed that polymer sensors performed well in discriminating multiple samples of five different types of bacteria. Keshri et al. [41] used an e-nose consisting of 14 polymer sensors to classify six spoilage fungi (four Eurotium sp., a Penicillium sp., and a Wallemia sp.). The headspace was sampled after 24, 48, and 72 hours of growth. The e-nose discriminated the fungi at the 24-hour mark (prior to the visible signs) with an accuracy of 93 %. The best results occurred at the 72-hour mark. The measurement of air quality by an e-nose requires a hand-held unit. Several commercial instruments are available as described in Chapters 7 and 9. Nicolas et al. [42] have also developed a portable prototype e-nose based on tin-oxide sensors for field applications; with this device they generate a warning signal when the malodor level exceeds some given threshold value, identify the source of an odor detected on site, or identify on-line and monitor levels of an odor in the field. As outlined above, the field of environmental monitoring is very broad. In this chapter, we will focus on case studies in livestock odors and microbial contamination. 17.2 Special Considerations for Environmental Monitoring 425

17.2 Special Considerations for Environmental Monitoring

17.2.1 Sample Handling Problems

17.2.1.1 Sample Lifetime If not properly handled (e.g., long exposure to sunlight), some organic samples may disintegrate or undergo certain chemical reactions. Therefore, considerable effort is required in order to maintain samples in their original state prior to their delivery to the sniffing device.

17.2.1.2 Humidity As will be discussed later, it is important that the various odor samples have similar humidity levels. The humidity of the reference sample should also be adjusted to that of the odor samples. This is to ensure minimal response due to humidity when switching from reference to odor inputs. A closed-loop humidity control system for the reference input is offered in some commercial systems for this reason.

17.2.1.3 Extraction of volatiles In cases in which the number of volatile molecules is low, one may be required to boost these numbers via some preconcentration, activation, or agitation method. In order to record a meaningful sensor response, the concentration of volatiles in the sample must be above a minimum threshold. Certain agitation methods may be necessary for liquid samples in order to increase the concentration of volatiles in the headspace. Conversely, in the case of highly volatile molecules (e.g., alcohols), one may need to dilute samples in order to avoid sensor saturation. Chapter 3 covers preconcentration methods.

17.2.1.4 Tubing system The acquisition system is generally equipped with a tubing system that delivers volatile compounds from the sample container to the sensor compartment, and then to the exhaust outlet. The material used in the tubing must be inert to the type of odorants that the device handles. In other words, the tubing material should not modify or adsorb the odor of the samples. Similar requirements exist for the sensor compartment, valves, and so on.

17.2.1.5 Temperature The temperature of the sample, sensor chamber, and sensors must be kept constant to achieve repeatable performance of the e-nose system. A temperature perturbation can cause shift/deformation in the generated patterns, by virtue of changes in concentration or sensor behavior. A constant temperature is usually maintained using a feedback control system. Temperature control is important for all types of sensors. 426 17 Environmental Monitoring

17.2.2 Signal Processing Challenges

In addition to appropriate sample handling, signal-processing algorithms are required to compensate for the variability of conditions in the field. By including temperature and humidity sensors in the e-nose instrument, it may be possible to compensate for these effects by means of signal processing algorithms. Sensor baseline drift and unwanted concentration effects may also be handled by means of preprocessing algorithms (see Chapter 5). Due to the large number of sensors and features (e.g., dynamic response recordings), the e-nose is subject to “the curse of dimensionality.” A large number of dimensions can hinder the true (and useful) information, so the use of dimensionality reduction procedures (e.g. feature selection, principal components) is often required. These signal processing procedures must be carefully chosen to ensure that memory and CPU requirements do not become prohibitive for an economical (e.g., hand-held) device [43].

17.3 Case Study 1: Livestock Odor Classification [44]

17.3.1 Background

Livestock industries are expanding rapidly throughout the world, and this expansion is causing environmental concerns. Modern methods of confining thousands of animals in a single facility have led to increased production and profits while creating concerns about odor and water pollution. Odors associated with livestock operations are generated from a mixture of urine, fresh and decomposing feces, and spilled feed. In swine operations, for example, odors emanate from the ventilation air of confinement buildings, waste storage, and handling systems including lagoons and field applications of waste. Anaerobic microbial decomposition of livestock waste appears to be the source of the most objectionable smells. Odorous compounds identified in livestock wastes include sulfides, disulfides, volatile organic acids, alcohols, aldehydes, amines, fixed gases, nitrogen heterocycles, mercaptans, carbonyls, and esters. Reduction of odors emanating from livestock operations is necessary to improve the relationship between producers and their neighbors. Sensitive measurement techniques are important to characterize and document swine odors, as well as evaluate the effectiveness of methods for reducing odors. At present, olfactometry using human odor panels is the most precise approach for quantifying odors, since the human nose can detect compounds at concentrations that cannot be detected by any other method. Human evaluations, however, can be time-consuming, unrepeatable, and expensive. In addition, odor samples degrade rapidly, and thus human panels must perform evaluations shortly after collection for accurate assessment. Because swine odor abatement research is being conducted 17.3 Case Study 1: Livestock Odor Classification [44] 427 all around the world on a 24-hour basis, odor testing with human panels is often impractical. Rapid, accurate, cost-effective evaluation of techniques to reduce odor production (such as the manipulation of pig diets to reduce excrement odor) is vitally important to the swine industry. For this reason it would be helpful to determine if an e-nose can substitute for human odor panels in evaluating methods for odor reduction.

17.3.2. Description of the problem

The objective of the following study was the classification of various odorant samples related to a hog farm. The main task was to gauge the accuracy and the precision of an e-nose in identifying the source of unknown odor samples.

17.3.3. Methods

Odor samples were collected from three locations at a rural hog farm: lagoon, fan, and downwind ambient air. The samples were presented to an e-nose, and signal-processing algorithms were used to classify the data. A cross-validation method was employed to measure the performance of the system. At each step of this cross-validation method, 70 % of the data was used to train the system, while the other 30 % was used as an unknown sample set. The e-nose used for the experiments of this section was the AromaScan A32S (see Chapter 7). The core of the A32S system is an array of 32 conducting-polymer sensors. Depending on the mode of operation, the sensor compartment is exposed to one of the odorant sample, the reference gas, or the cleansing gas. The reference gas was generated by filtering, dehydrating, and humidifying steps. The humidity of the reference air was set to match that of the odor samples. The cleansing gas (2 % n-butanol bubbler) was used to remove (detach) odorants from the sensors after each data acquisition cycle. Various air-samples from two lagoons, a confinement building exhaust fan, and a downwind site at a hog farm in rural North Carolina were collected using 25-L Tedlar bags. The downwind-air sample was collected 1,500 feet from the swine operation. These bags were filled using a pump device and sealed barrel under negative pressure. The bags were cleaned using a combination of butanol, methanol, nitrogen, and/or dry air, and reused. The most commonly used cleaning technique was flush- ing with nitrogen, then a methanol vapor, followed by clean dry air. A major drawback of this sampling method is the shipping and handling of the filled bags. Since the odors degrade over time, the samples should be processed the same day during which they are collected. Hence, this technique is adequate for sites that are located in close proximity (within 150 miles) to the testing facilities. We have found that holding the bags overnight for processing the following day significantly reduces the odor intensity, and hence the reliability of the sample collection method. 428 17 Environmental Monitoring

17.3.4 Signal Processing Algorithms

The datasets obtained from the e-nose were analyzed using a set of algorithms listed below. More detailed explanations of the various algorithms can be found in Chap- ters 5 and 6. The main steps of signal processing in this case study are outlined as follows:

17.3.4.1 Bias Removal One of the drawbacks of polymer sensors is their inability to return (within a reasonable time frame) to the baseline after washing. The residual signal will result in a gradual shift in the successive data acquisition cycles. The first step of preprocessing was to remove the bias mathematically. In these experiments, the bias was removed by subtracting the response of each sensor at the first time point from all the other subsequent time points in the dynamic response of that sensor.

17.3.4.2 Humidity Another major weakness of some conducting-polymer sensors is their high sensitivity to water molecules. If not controlled, the common-mode response that is caused by humidity could completely overshadow the signal of the odorants. Various approaches have been proposed to counteract humidity and its effects. One is to model the response of the sensors to humidity, and then to subtract it from the composite response. However, due to the low repeatability of the patterns, this was not found to be a suitable approach for the AromaScan A32S polymer sensors. Another approach is to employ the humidity control features of the AromaScan A32S that allow the operator to adjust the humidity of the reference signal to that of the odor sample. We should point out that researchers in this field are developing new types of conducting-polymer sensors that are much less sensitive to changes in sample humidity.

17.3.4.3 Concentration One obvious challenge in sample preparation is the control of the volatile concentration. Within certain ranges, the effect of concentration has been shown to be linear. When comparing samples of the same kind, one must be able to either normalize the effect of concentration, or guarantee that samples contain similar concentrations of the odorant of interest. In the experiments of this study, the response of each sensor at each time point was divided by the average response of all sensors at that time point. When the sensors operate in the linear range, this method has been shown to normalize the response of the sensors with respect to the concentration [44].

17.3.4.4 Dimensionality Reduction In the following experiments, every sample produces 30 32 ¼ 960 data points. Since a single training session may include several dozens of samples, it is evident that the dimensionality could become overwhelming for this problem. Therefore, in lieu of 17.3 Case Study 1: Livestock Odor Classification [44] 429 supplying the time-series data directly into the processing unit, a reduced set of features was extracted prior to the main analysis. Data reduction was done in two stages. In the first stage, a series of bell-shaped curves were used to serve as windowing functions. By using windowing functions, the set of 30 time points of the response of each sensor was reduced to four, the number of windowing functions. The next step of data compression was done by Karhu- nen-Loe´ve (truncated) expansion (KLE), also known as principal components analysis. KLE is known to be the optimal linear method for data compression [45]. Using KLE, a series of features, i.e., the significant eigenvectors, was extracted from the time-windowed traces of each sample. The dimension of the transformed signal was found dynamically by analyzing the relationship between the eigenvalues of the covariance matrix [44]. The set of features extracted from the KLE compression was then directed into an multi-layer perceptron neural network for training and testing. The learning rule of the neural network was based on the Levenberg-Marquardt method [46, 47]. The back-propagation method [48] (with a momentum term and adaptive learning rates) was also used for comparison purposes. A genetic-algorithm-based supervisor was designed to tune the number of neurons in the hidden layer and the learning parameters of the neural network. The genetic algorithm (GA) was also responsible for choosing all or a subset of the windowed values and/or features.

17.3.5. Results

The results are depicted in Fig. 17.1. Aside from the difficulties of sample handling, the results appear to be reasonable. The figure shows the histogram of the performance of 100 cross-validated runs. The y-axis is the number of runs and the x-axis is the correct recognition in percent. Note that 97 of the runs gave a perfect 100 % correct recognition, while the remaining three cases were 97 % correct. The overall correction recognition rate was 99.92 %.

17.3.6. Discussion

Several alternative signal processing methods, e.g., neural networks with back-propagation, with and without the GA supervision, were tried prior to applying the above- mentioned methods. These alternative methods were found to achieve lower performance metrics. The preprocessing steps were found to be necessary for generating repeatable histogram patterns. A neural-network-based classifier with the Leven- berg-Marquardt learning rule was found to be appropriate for this particular pattern-recognition application. Using GAs as a supervisor provided a systematic, reliable, and automated method for feature selection and architectural tuning of the neural network. The final hybrid GA-neural network system proved to serve as an effective signal- processing technique for this application. However, regardless of the efficacy of the 430 17 Environmental Monitoring

Fig. 17.1 Histogram showing test results of 100 runs of training/ testing of hog-farm samples using the hybrid of neural-network and genetic-algorithms in conjunction with the AromaScan A32S. The number of runs is given on the y-axis, and the percent correct recognition is given on the x-axis. On 97 of the runs, there was a perfect 100 % correct recognition, while there was 97 % correct recognition for the remaining three cases

signal-processing method, the quality of the final outcome is a function of the quality of the input data. In general, due to their limited sensitivity, conducting-polymer sensors were found to be more suitable for odor samples containing high concentrations of highly volatile molecules such as those found in fragrances.

17.4 Case Study 2: Swine Odor Detection Thresholds

17.4.1. Description of the Problem

The detection threshold for a specific odorant mixture is related in part to the detection thresholds of its individual components. In this study, we select one of the odorous components of hog slurry – acetic acid – and compare the detection thresholds of a human panel and the AromaScan A32S for this compound. 17.4 Case Study 2: Swine Odor Detection Thresholds 431

17.4.2 Methods

In this experiment, twelve serial dilutions of acetic acid that differed by a factor of three and ranged from 5 % to 0.0000094 % v/v were presented to the human panel at the Taste and Smell Laboratory at Duke University Medical Center and the AromaScan A32S for evaluation. Odorless mineral oil was used as the diluent. The e-nose signals were processed using the same procedure as Case Study 1 above [44, 49]. The techniques used consisted of a preprocessing stage and a data-compression stage. The preprocessing stage involved shifting each sensor’s curve, so that the initial resistance change was adjusted to zero. The data-compression stage consisted of two steps: windowed time integration and Karhunen-Loe´ve expansion (KLE). The windowed time integration multiplied each sensor curve by four bell-shaped kernels and then computed the area beneath the curves. In this way, each odor sample was reduced from 32 45 (sensors x seconds) to 32 4 (sensors x windows) features. Then the KLE was performed to extract the principal components in feature space.

17.4.3 Results

The dilution labels ranged from 13 to 1, for the highest and lowest concentrations, respectively. The resultant two-dimensional KLE scatter plot for the acetic acid dilutions in mineral oil is presented in Fig. 17.2. Note that a detection threshold between labels 9 and 10 can be visually determined.

17.4.4 Discussion

Our results indicate that the e-nose has a detection threshold at a concentration that is a factor of three above that of the human panel. The detection thresholds for the four human subjects were at dilutions 8 or 9 (two subjects at each dilution), whereas the e- nose was between dilution 9 and 10, as can be seen in the figure. Since dilution 10 has an odorant concentration that is three times greater than dilution 9, and dilution 9 has an odorant concentration that is three times greater than dilution 8, on average the human panel’s detection level is at a concentration that is three times lower than that of the e-nose. A factor of three in odorant concentration therefore gives the human panel an advantage over the e-nose in this application. However, the e-nose can be deployed on site and can measure emissions over long time periods, characteristics of a monitoring system that are not practical for human-panel implementation 432 17 Environmental Monitoring

Fig. 17.2 Principal component analysis (PCA) of the e-nose data for dilutions of acetic acid in mineral oil. The two-dimensional scatter plot shows that a detection threshold occurs between labels 9 and 10

17.5 Case Study 3: Biofilter Evaluation [50]

17.5.1 Description of the Problem

The objective of this study was two-fold. First, to develop an experimental procedure to evaluate biofilters for odor remediation in the ventilation exhaust fans of hog confinement buildings. Second, to determine if the AromaScan A32S could be utilized to predict the human panel olfactory ratings of malodors, before and after bioremedia- tion.

17.5.2 Methods

In order to rapidly screen the performance of various odor remediation materials, a bench-top biofilter setup was developed at the NC State University Animal and Poultry Waste Management Center. The biofilter material consisted of earth, wood chips, small twigs, and straw. This material was placed in a one-inch diameter PVC tube, which was cut to a length of 3.9 inches. This length was selected because of the requirement to have the air reside within the filter for 15 seconds, which matches the specifications of field units at this site. The tube was cemented at each end to a PVC fitting which had screw threads and an O-ring to produce an airtight seal with the 17.5 Case Study 3: Biofilter Evaluation [50] 433

Fig. 17.3 Experimental setup for malodor biofiltration assessment. Air from the synthetic hog slurry and the room-air control is filtered and delivered to the human sensory panel and e-nose (AromaScan A32S) for analysis connecting piece. Wire mesh was placed on each end of the cemented tube fitting to prevent the biofilter material from spilling out of the tube. To test this biofilter setup, we conducted an odor remediation experiment with a synthetic slurry following the concoction of Persaud et al. [51]. Serial dilutions (1/ 1, 1/3, 1/9, 1/27 and 1/81) of the headspace above the slurry, as well as serial dilutions of the biofiltered synthetic slurry and biofiltered blank room air (as a control) were presented to both the Duke human panel and the e-nose. The experimental setup is depicted in Fig. 17.3. To measure the human perception to the different odors and dilutions, the panelists were asked to generate scores for intensity, irritation, and pleasantness using the 9- point scale shown in Table 17.1. The e-nose signals were preprocessed by computing the fractional change in resistance of each sensor with respect to its baseline resistance in reference air (steady-state DR/R). The steady-state response of each sensor was extracted to form a 32-dimensional feature vector.

Table 17.1 Hedonic tone odor rating scales

Scale Odor Intensity Irritation Intensity Pleasantness

8 Maximal Maximal Extremely Unpleasant 7 Very Strong Very Strong Very Unpleasant 6 Strong Strong Moderately Unpleasant 5 Moderately Strong Moderately Strong Slightly Unpleasant 4 Moderate Moderate Neutral 3 Moderately Weak Moderately Weak Slightly Pleasant 2 Weak Weak Moderately Pleasant 1 Very Weak Very Weak Very Pleasant 0 None at all None at all Extremely Pleasant 434 17 Environmental Monitoring

17.5.3. Results

The average response of the human panel and the 32 conducing-polymer sensors in the e-nose for each of the 15 dilutions (five dilutions for each of three odor sources) is shown in Fig. 17.4. Note that for the human panel, biofiltering reduced the intensity, irritation, and unpleasantness of the odor. In addition, the panel’s ratings of the biofiltered slurry and blank air were quite similar. In order to establish whether the e-nose could be used to replace a human panel in odor-remediation scenarios, we performed partial-least-squares regression [52] to predict the average response of the human panel from the 32-dimensional average response of the e-nose. To establish the predictive accuracy of this model, we performed cross-validation in which one of the fifteen dilutions was removed from the training data and predicted only after the partial-least-squares model had been obtained. Fig- ure 17.5 shows the performance of the model on test data for these fifteen leave-one- out validation runs. The correlation coefficient (between predictions and true values) on test data for intensity, irritation, and pleasantness are 0.90, 0.94 and 0.86, respectively. Given the notorious cross-sensitivity of conducting polymers to moisture, we decided to analyze the response of the built-in humidity sensor of the AromaScan A32S to the different odors and dilution ratios. The transient response of odor and

Fig. 17.4 Average human and e-nose response versus dilution number in the biofiltration experiment. The labels on the abscissa for the serial dilutions are defined as follows: 5 (1/1 dilution), 4 (1/3 dilution), 3 (1/9 dilution), 2 (1/27 dilution), and 1 (1/81 dilution). The human response sale is defined in Table 17.1. As expected, both human and e-nose (AromaScan A32S) responses decrease with increasing dilution 17.5 Case Study 3: Biofilter Evaluation [50] 435

Fig. 17.5 True vs. predicted human panel ratings for intensity, irritation, and pleasantness using the odor sensor array based on the performance of the model on test data for the fifteen leave-one-out validation runs. q ¼correlation coefficient humidity sensors to the fifteen samples is shown in Fig. 17.6. Two observations can be made. First, looking at the humidity sensor response to the slurry before and after biofiltration, it can be concluded that the biofilter material is increasing the relative humidity of the samples. Second, as a result of serial dilutions, the humidity of the samples is significantly reduced. On the basis of these results, it is necessary to determine if humidity is dominating the e-nose response. A closer look at the data shows one that the response of the sensor array to the synthetic slurry has a unique dynamic signature that is different from the

Fig. 17.6 Transient response of the gas sensor array and the humidity sensor to five serial dilutions per odor using the AromaScan A32S. The waveforms in both the upper and lower portions of the figure show the time response of the odor and humidity sensors for each dilution (labeled in the center of the figure). Note that the humidity sensor response indicates that the biofilter material is increasing the relative humidity of the samples. Serial dilutions with dry air reduce the humidity of the samples 436 17 Environmental Monitoring

Fig. 17.7 True versus predicted human panel ratings using only the humidity sensor. The correlation coefficients between the true and predicted values for intensity, irritation, and pleasantness are reduced compared with those in Fig. 17.5, thus the conducting-polymer sensor array gives much better performance than the humidity sensor alone

exponential decay to the biofiltered samples. This indicates that, in spite of relative humidity changes, the odor sensors are able to detect the synthetic slurry. In addition, if the odor sensors were responding only to the humidity, the largest response of the sensor array would then occur with the 1/1 biofiltered blank since this sample has the highest response on the humidity sensor. To further rule out the possibility that the e-nose is just detecting differences in moisture, it was attempted to predict the human olfactory ratings from the humidity sensor response alone. The results are summarized in Fig. 17.7. The correlation coefficients between these single sensor predictions and true values by the human panel on test data for intensity, irritation, and pleasantness drop down to 0.40, 0.31 and 0.29, respectively. Hence, the conducting-polymer sensor array is giving much better performance, proving that the response of the odor sensors contains information related to the presence of synthetic slurry.

17.5.4 Discussion

The main findings of this study are that the AromaScan A32S can differentiate between different dilutions of the components of swine odor, and between synthetic slurry and biofiltered slurry/blank samples. The sensor array response can be used to predict the intensity and pleasantness olfactory ratings from a human panel. Moist- ure is shown to be a major interferent since biofiltration increases the relative humidity of the samples. However, the signal processing routines were able to mediate this interference. In the future, this interference might be reduced further by performing serial dilutions with a carrier gas having the same relative humidity as the odor samples. 17.6 Case Study 4: Mold Detection [53] 437

17.6 Case Study 4: Mold Detection [53]

17.6.1 Background Microbial contamination of our environment is an area of increasing concern. An e- nose has the potential to identify and classify microorganisms, including bacteria and fungi. When conditions are favorable and a nutrition source is present, microbial organisms such as fungi and bacteria can grow almost anywhere. Microorganisms have been shown to generate VOCs while metabolizing nutrients, and these VOCs have been used as indicators of microbial growth. Colonies of microorganisms not only generate airborne contamination in the form of VOCs, but also generate toxins, conidia (spores), and bacterial cells. When microoganisms infest buildings, they can produce a potentially hazardous environment. Individuals exposed to environments that contain high concentrations of airborne contaminants from microbial organisms report health symptoms including eye and sinus irritation, headaches, nausea, fatigue, congestion, sore throat, and even toxic poisoning. Sick-building syndrom, which includes health symptoms arising from poor indoor air quality, has been correlated with the presence of fungi [54]. A study of two housholds reporting indoor environmental complaints correlated the presence of excessive VOCs with the presence of fungal contamination [55]. Typical signs of microbial contamination include water damage, high levels of humidity, and physical presence. However, these signs are not always present, and therefore cannot be utilized as sole indicators of microbial contamination. Current methods for detecting microbial contamination include air and material sampling with culture analysis, air sampling coupled with gas chromatography/ mass spectrometry, and visual inspection [56, 57]. These methods, however, can be inconclusive as well as time consuming and expensive. Thus, rapid detection of the presence of microbial contamination is needed in order to minimize its impact.

17.6.2 Description of the Problem In this study, we explored the ability of the NC State E-Nose, a prototype electronic system with 15 metal-oxide sensors, to detect fungi at various stages of growth. Fungi that are typically found in indoor air-conditioning systems were chosen for experimen- tation. The purpose of the experiment was to demonstrate that an e-nose system is capable of diagnosing the presence of these fungal types in commercial buildings and residential housing units.

17.6.3 The NC State E-Nose An e-nose instrument was designed and constructed at North Carolina State Univer- sity [44, 49] that uses an array of metal-oxide sensors for measuring odor in air samples 438 17 Environmental Monitoring

(see Fig. 17.8). The e-nose consists of a sampling unit, a sensor array, and a signal processing system. The sampling unit, which consists of a pump and a mass-flow controller, directs the air sample containing the odorant under investigation across the sensor array. The current configuration allows for sampling from a set of 12 odorants, a reference sample (filtered odorless dry ambient air), and a washing agent (ambient air bubbled through a 2 % n-butanol solution). The tubing and sensor chamber are made of stainless steel. The sensor chamber is designed to minimize dead volume (see Fig. 17.9). The sensor array is composed of 15 different metal-oxide sensors. Twelve of the 15 metal-oxide sensors are manufactured by Capteur (Didcot, UK) and include sensors for isopropyl alcohol, toluene, hydrogen sulfide, nitrogen dioxide, chlorine, butane, propane, hydrogen, carbon monoxide, heptane, ozone, and general VOCs. The remaining three metal-oxide sensors are produced by Figaro USA (Glenview, IL) and include methane, a combustible gas, and a general air-contaminant sensor. All of the sensor response patterns are digitized and recorded using a National Instruments Data Acquisition Card controlled by LabVIEWJ.

The solenoid valves are normally closed. Solenoid valve s1 (exhaust) and an appropriate inlet solenoid valve (s2 to s15) are opened at the beginning of each phase and closed afterwards. The mass flow controller must also be set at the beginning of

Fig. 17.8 System configuration for the NC State E-Nose. The exhaust pump pulls air samples through the system. The mass flow controller (MFC) and exhaust pump can be separated from the system by solenoid

valve S1. The system has 14 sample input ports controlled by solenoid valves S2 to S15. Ports S2 and S3 are assigned the washing (cleaning) and reference functions, respectively. Ports S4 through S15 are designated as odor sample handling inputs. The system includes an inline pressure sensor, a combined temperature/humidity sensor, and 15 metal-oxide odor sensors 17.6 Case Study 4: Mold Detection [53] 439

Fig. 17.9 The sensor chamber of the NC State E-nose. (a) airflow pattern; (b) photograph. Commercially available metal-oxide sensors are mounted in a stainless steel chamber. The electrical leads of the sensors are soldered to printed circuit boards with attached ribbon cables that relay the sensor responses to interfacing electronics. From the top of the chamber, air enters a cylindrical tube with holes that ‘jet’ the odor samples directly onto each odor sensor. After passing over the sensors, the air streams merge and exit the chamber each cycle to the appropriate set point (between 0.0 and 1.0 L min1). The operation cycle for the NC State E-Nose consists of three phases: wash, reference, and sample.

Wash phase: solenoid valves s1 and s2 are opened. Room air is passed through a charcoal filter (to remove residual ambient odors) and a bubbler with 2 % diluted n-butanol in distilled water. The resulting gas is used to flush tubing and sensors and remove traces of odorants from previous gas samples.

Reference phase: solenoid valves s1 and s3 are opened. Room air is passed through a charcoal filter (to remove residual odors) and a moisture trap. The resulting odor-free 440 17 Environmental Monitoring

dry air is used as a reference gas to force the sensor resistances back to their baseline values.

Sample phase: solenoid valve s1 and one other valve (s4 to s15) are opened. The odorous sample is passed through the e-nose. Return to Wash phase.

17.6.4 Methods

Five fungi (Aspergillus flavus, Aspergillus niger, Penicillium chrysogenum, Cladosporium cladosporioides, and Stachybotrys chartarum) were incubated at 28 8C on 150-mm diameter Petri dishes containing potato dextrous agar (PDA), a complex media rich in nutrients, and Czapek-Dox agar (CZ), a minimal media. These two types of media were used in order to provide two different growth environments and to produce different growth rates. Twenty-four Petri dishes of each media were inoculated with 0.5 mL of an individual spore suspension containing 10 000 condia mL1 from each fungus, respectively. The suspensions were prepared using a Spencer hemacyt- ometer with improved Neubauer ruling. Using the autosampler functions of the NC State E-Nose, air samples from the headspace of each Petri dish containing one species on each medium were randomly sampled ten times, each after 24 hours and every other day thereafter for two weeks. The headspace above each fungus was sampled through a small hole in the center of the lid of the Petri dish using a PVC tube and an inline 2-lm filter for removing conidia (spores). The data were analyzed with MATLABJ using signal-processing algorithms developed by Kermani [44] and Gutierrez-Osuna [49]. More specifically, the raw data were first compressed using windowing functions that produced a set of four features for each sensor. Linear-discriminant analysis was then applied to the compressed data to maximize class separability. Sixty percent of the compressed data was randomly selected to form a training set for the classification algorithms. K-nearest-neighbors (KNN) and least-squares (LS) techniques were both employed to classify the remaining 40 % of the compressed data [58]. This process was repeated 100 times, and the average score was used as the final classification score.

17.6.5 Results

The data were analyzed using two classification protocols. In the first protocol, the data were grouped into 12 classes: five fungal species grown on PDA and CZ, respectively, plus two controls (the two media PDA and CZ without fungal growth). The results are shown in Table 17.2. After 24 hours of growth, the percent classification was 90 % for KNN, and 76 % for LS. Classification for the 12 classes reached a maximum after five days of growth, with an accuracy of 96 % for KNN and 94 % for LS. After day 5, the percent classification began to decrease slowly. By day 15, the percent classification was reduced to 89 % for KNN and 69 % for LS. 17.7 Future Directions 441

Table 17.2 Percent classification for 12 classes (five fungal species on two different media and two control media) [53]

Classification Method Day of Growth 13579111315

KNN 90 % 91 % 96 % 94 % 89 % 93 % 93 % 89 % LS 76 % 90 % 94 % 90 % 93 % 86 % 80 % 69 %

Table 17.3 Percent classification of seven classes (five fungal species and two control media) [53].

Classification Method Day of Growth 13579111315

KNN 89 % 90 % 94 % 93 % 89 % 94 % 94 % 92 % LS 79 % 88 % 93 % 91 % 95 % 90 % 92 % 86 %

In the second classification protocol, the data were grouped into seven classes: five fungal species (independent of media used for growth) plus two controls (the two media PDA and CZ without fungal growth). In other words, each of the fungi grown in PDA and CZ were combined into a single class. After 24 hours of growth, the percent classification was 89 % for KNN, and 79 % for LS. Classification reached a maximum after five days of growth, with an accuracy of 94 % for KNN and 93 % for LS. After day 5, the percent classification oscillated around an average percent classification of 92 % with a standard deviation of 2 %. The results are shown in Table 17.3.

17.6.6 Discussion

The experiment with five fungi showed that the NC State E-Nose using metal-oxide sensors can detect and classify microorganisms on the basis of volatile emissions. The classification was independent of the media used to grow the fungi. Furthermore, correct classification was achieved early in the experiment at 24 hours of growth. Thus e-nose instruments of this type have the potential to be used for early detection of microbial contamination in office buildings and manufacturing facilities.

17.7 Future Directions

The success of laboratory instruments in classifying environmental odors has been demonstrated by many research groups around the world. This success must now be leveraged to build new portable instruments for field use. These portable units must operate in real time, recording odor concentration profiles at specific time intervals tailored to individual environmental monitoring applications. These devices must 442 17 Environmental Monitoring

be able to detect odors at very low (parts per billion) levels. Hence, more sensitive gas sensors and preconcentration units must be included in instruments that will be used in on-site, real-time environmental measurements. Chapters 7 and 9 have illustrated some progress by the instrument makers towards reaching these goals. Improvements in signal-processing algorithms can offer some assistance. Low-power, embedded mi- croprocessors are continually being improved by the electronics industry. Incorporat- ing more powerful real-time data-processing algorithms onboard these portable instruments will differentiate the different commercial models. If the e-nose manufacturers can ‘break’ into the environmental monitoring market in a significant way, the future of this technology will be guaranteed.

Acknowledgements The authors wish to acknowledge the support of the National Science Foundation, the US Agricultural Research Service, the National Pork Producers Council, the NC State University Animal and Poultry Waste Management Center, and the Center for Indoor Air Research for supporting various portions of the work reported herein.

References

1 R. E. Baby, M. Cabezas E. N. W. de Reca. 7 R. M. Stuetz, S. George, R. A. Fenner, Electronic nose: a useful tool for monitoring S. J. Hall. Monitoring wastewater BOD environmental contamination. SensorActual using a non-specific sensor array. J Chem B-Chem 69 (3): 214–218 OCT 25 2000. Technol Biot 74 (11): 1069–1074 1999. 2 T. Dewettinck, K. Van Hege, W. Verstraete. 8 W. Bourgeois, R. M. Stuetz. Measuring The electronic nose as a rapid sensor for wastewater quality using a sensor array: volatile compounds in treated domestic prospects for real-time monitoring. wastewater. Water Res 35 (10): 2475–2483 Water Sci Technol 41 (12): 107–112 2000. JUL 2001. 9 R. M. Stuetz, R. A. Fenner, S. J. Hall. et al. 3 K. Van Hege, T. Dewettinck, W. Verstraete. Monitoring of wastewater odours using Pre-evaporative Fenton remediation of an electronic nose. Water Sci Technol 41 (6): treated municipal wastewater for reuse 41–47 2000. purposes. Environ Technol 22 (5): 523–532 10 W. Bourgeois, J. E. Burgess, R. M. Stuetz. 2001. On-line monitoring of wastewater quality: 4 F. Di Francesco, B. Lazzerini, F. Marcelloni, a review. J Chem Technol Biot 76 (4): G. Pioggia. An electronic nose for odour 337–348 2001. annoyance assessment. Atmos Environ 35 (7): 11 R. Stuetz. Using sensor arrays for on-line 1225–1234 2001. monitoring of water and wastewater quality. 5 R. A. Fenner, R. M. Stuetz. The application Am Lab 33 (2): 2001. of electronic nose technology to environ- 12 C. Di Natale, A. Macagnano, F. Davide, mental monitoring of water and wastewater A. D’Amico, A. Legin, Y. Vlasov, treatment activities. Water Environ Res 71 (3): A. Rudnitskaya, B. Selezenev. Multi- 282–289 1999. component analysis on polluted waters 6 R. M. Stuetz, R. A. Fenner, G. Engin. As- by means of an electronic tongue. Sensors sessment of odours from sewage treatment Actuat B-Chem 44 (1–3): 423–428 1997. works by an electronic nose, H2S analysis 13 J. W. Gardner, H. W. Shin, E. L. Hines, and olfactometry. Water Res 33 (2): 453–461 C. S. Dow. An electronic nose system for 1999. monitoring the quality of potable water. Sensors Actuat B-Chem 69 (3): 336–34 2000. 17.7 Future Directions 443

14 H. W. Shin, E. Llobet, J. W. Gardner, 26 C. Delpha, M. Siadat, M. Lumbreras. E. L. Hines, C. S. Dow. Classification of Discrimination of a refrigerant gas in a the strain and growth phase of cyanobacteria humidity controlled atmosphere by using in potable water using an electronic nose modelling parameters. Sensors Actuat system. IEE P-Sci MeasTech 147 (4): B-Chem 62 (3): 226–232 2000. 158–164 2000. 27 C. Delpha, M. Siadat, M. Lumbreras. 15 E. M. Biey, W. Verstraete. The use of a An electronic nose for the identification UV lamp for control of odour decomposition of Forane R134a in an air-conditioned of kitchen and vegetable waste. Environ atmosphere. Sensors Actuat B-Chem 69 (3): Technol 20 (3): 331–335 1999. 243–247 2000. 16 P. E. Keller, R. T. Kouzes, L. J. Kangas. 28 F. Sarry, M. Lumbreras. Gas discrimination Three Neural Network Based Sensor in an air-conditioned system. IEEE T Instrum Systems for Environmental Monitoring. MEAS 49 (4): 809–812 2000. IEEE Electro 94 Conference Proceedings, 29 O. Ramalho. Correspondences between Boston, MA, 377–382 1994. olfactometry, analytical and electronic nose 17 C. Mouche. Electronic nose sniffs out, data for 10 indoor paints.Analysis 28 (3): classifies contamination. Pollut Eng 31 (2): 207–215 2000. 31 1999. 30 R. Feldhoff, C. A. Saby, P. Bernadet. Dis- 18 J. Lee, J. Stewart. “Omonos: A Computer crimination of diesel fuels with chemical Model for the Dispersion of Odours in Air” sensors and mass spectrometry based elec- in Clean Air and Environmental Protection, tronic noses. Analyst 124 (8): 1167–1173 Vol 29, No. 5, Published by the National 1999. Society for Clean Air, Brighton, England, 31 R. J. Lauf, B. S. Hoffheins. Analysis of liquid 140–144 1999. fuels using a gas sensor array. Fuel 70 (8): 19 S. S. Schiffman, E. A. Satterly-Miller, 935–940 1991. M. S. Suggs, B. G. Graham. The effect of 32 R. Menzel, J. Goschnick. Gradient gas environmental odors emanating from com- sensor microarrays for on-line process mercial swine operations on the mood of control – a new dynamic classification model nearby residents. Brain Res Bull 37: 369–375 for fast and reliable air quality assessment. 1995. Sensors Actuat B-Chem 68 (1–3): 115–122 20 P. J. Hobbs, T. H. Misselbrook, 2000. M. S. Dhanoa, K.C. Persaud. Development 33 A. R. Hopkins, N. S. Lewis. Detection and of a relationship between olfactory response classification characteristics of arrays of and major odorants from organic wastes. carbon black/organic polymer composite J Scs Food Agr 81 (2): 188–193 2001. chemiresistive vapor detectors for the nerve 21 P. Gostelow, S. A. Parsons, R. M. Stuetz. agent simulants dimethylmethylphospho- Odour measurements for sewage treatment nate and diisopropylmethylphosponate. works. Water Res 35 (3): 579–59 2001. Anal Chem 73 (5): 884–892 2001. 22 R. M. Stuetz, G. Engin, R. A. Fenner. Sewage 34 G. Hudon, C. Guy, J. Hermia. Measurement odour measurements using a sensory panel of odor intensity by an electronic nose. and an electronic nose. Water Scs Technol 38 J Air Waste Manage 50 (10): 1750–1758 (3): 331–335 1998. 2000. 23 R. M. Stuetz, R. A. Fenner, G. Engin. 35 R. M. Negri, S. Reich. Identification of Characterisation of wastewater using an pollutant gases and its concentrations with a electronic nose. Water Res 33 (2): 442–452 multisensor array. Sensors Actuat B-Chem 75 1999. (3): 172–178 2001. 24 F. W. Schreiber, K. Fitzner. Electronic Nose: 36 B. Wesse´n, K.-O. Schoeps. Microbial volatile Investigation of the Perceived Air Quality in organic compounds – what substances can Indoor Environments, Indoor Air 99, Vol. 2, be found in sick buildings? Analyst 121: 624–629 Edinburgh, 8–13.08.1999. 1203–1205 1996. 25 F. W. Schreiber, K. Fitzner. Investigation of 37 A.-L. Sunesson. et al. Identification of the Perceived Air Quality in an Office volatile metabolites from five fungal species Building with an Electronic Nose, Healthy cultivated on two media. Appl Environ Buildings 2000, Helsinki, 6–10.08.2000. Microbiol 61: 2911–2918 1995. 444 17 Environmental Monitoring

38 M. Holmberg. Data Evaluation for an 49 R. Gutierrez-Osuna. Signal processing and Electronic Nose. Dissertation, Depart. Phys. pattern recognition for an electronic nose. Meas. Tech. Linko¨ping University, Sweden, Doctoral Dissertation, Department Electrical 1997. Computer Engineering, North Carolina 39 J. W. Gardner, M. Craven, C. Dow, E. L. State University, USA, 1998. Hines. The prediction of bacteria type and 50 R. Gutierrez-Osuna, S. S. Schiffman, culture growth phase by an electronic nose H. T. Nagle. “Correlation of Sensory with a multilayer perceptron network. Meas Analysis with Electronic Nose Data for Sci Tech 9: 120–127 1998. Swine Odor Remediation Assessment,” in 40 V. Vernat-Rossi, C. Garcia, R. Talon, C. Y. Proceedings of the 3rd European Congress DeLayer, J. L. Berdague. Rapid discrimina- on Odours, Metrology and Electronic Noses, tion of meat products and bacterial strains Paris, France, June 19–21, 2001. using semiconductor gas sensors. Sensors 51 K. C. Persaud, S. M. Khaffaf, O. J. Hobbs, Actuat B-Chem 37: 43–48 1996. R. W. Sneath. Assessment of conducting 41 G. Keshri, N. Mayan, P. Voysey. Use of an polymer odour sensors for agricultural electronic nose for the early detection and malodour measurements, Chemical Senses differentiation of spoilage fungi. Lett Appl 21: 495–505 1996. Microbiol 27: 261–264 1998. 52 P. Geladi, B. R. Kowalski. Partial least- 42 J. Nicolas, A. C. Romain, V. Wiertz, squares regression: A tutorial. Anal Chim J. Maternova, P. Andre. Using the classifi- Acta 185: 1–17, 1986. cation model of an electronic nose to assign 53 S. S. Schiffman, D. W. Wyrick, R. Gutierrez- unknown malodours to environmental Osuna, H. T. Nagle. “Effectiveness of an sources and to monitor them continuously. electronic nose for monitoring bacterial and Sensor Actuat B-Chem 69 (3): 366–371 2000. fungal growth.” in: Gardner JW, Persaud 43 A. Perera, T. Pard, T. Sundic, S. Marco, KC. Electronic Noses and Olfaction 2000, R. Gutierrez-Osuna. “IpNose: Electronic Bristol: Institute of Physics Publishing, nose for distributed air quality monitoring 2000, pp. 173–180. system,” in Proceedings of the 3rd European 54 D. G. Ahearn. et al. Fungal colonization of Congress on Odours, Metrology and Elec- fiberglass insulation in the air distribution of tronic Noses, Paris, France, June 19–21, a multistory office building: VOC production 2001. and possible relationship to sick building 44 B. G. Kermani. On using artificial neural syndrome. J Indust Microbiol 16: 280–285 networks and genetic algorithms to optimize 1996. performance of an electronic nose. Ph.D. 55 G. Stro¨m. et al. Health Implications of Fungi Dissertation, Department of Electrical Eng- in Indoor Environments, Elsevier, Amster- ineering, North Carolina State University, dam, 291–305, 1994. Raleigh, NC, 1996. 56 S. S. Schiffman, J. L. Bennett, J. H. Raymer. 45 K. Fukunaga. Introduction to statistical Quantification of odors and odorants from pattern recognition, 2nd Edition, Academic swine operations in North Carolina. Ag Press Inc., San Diego, CA, 1992. Forest Meteor 108: 213–240, 2001. 46 K. A. Levenberg. A method for the solution 57 A. L. Pasanen. et al. Occurrence and mois- of certain non-linear problems in least ture requirements of microbial growth in squares, Quart Appl Math 2: 164–168, 1944. buildings. Int Biodeter Biodegrad 30: 273– 47 D. Marquardt. An algorithm for least squa- 283 1992. res estimation of non-linear parameters, J 58 R. O. Duda, P. E. Hart. Pattern classification Soc Ind Appl Math 11: 431–441, 1963. and Scene Analysis, Wiley, New York, 1973. 48 D. E. Rumelhart, J. L. McClelland. Parallel Distributed Processing, MIT Press, Cam- bridge, MA, p.318, 1986. 445

18 Medical Diagnostics and Health Monitoring

Krishna C. Persaud, Anna Maria Pisanelli, Phillip Evans

18.1 Introduction

Many diseases and intoxications are accompanied by characteristic odors, and their recognition can provide diagnostic clues, guide the laboratory evaluation, and affect the choice of immediate therapy [1–4]. Common observations are the change in breath odor profile in diabetic patients entering a ketotic state, while the profiles of urinary volatiles from patients with phenylketonuria, maple syrup urine disease, isovaleric acidemia, or trimethylaminuria (fish-odor syndrome) are vastly different from the normal urinary volatiles profile [5]. It is also recognized that many bacteria growing on specific media produce characteristic odorous metabolites, and that these can be used to diagnose which bacteria species are present in a culture [6]. The realization that electronic nose technologies may be a useful diagnostic aid has spurred activity in many research laboratories and companies, one of the earliest clinical trials of the technology reported being detection of infections in leg ulcers in patients in 1995 [7]. This chapter reviews major activity in the field (see Table 18.1), and then focuses on selected investigations in the area of myopathies and in bacterial vaginosis (BV), to provide perspective on measurement and sampling requirements for applications of electronic noses in clinical measurements and diagnosis. Medical and health-monitoring applications are often cited in the electronic nose literature. However, converting these potential markets to commercial reality has yet to be achieved. There are numerous reasons for this, not least being the requirements for robustness when dealing with the health of a patient, mistakes could be costly for all concerned. There is also the dichotomy between the ability to perform the measurement and the need for measurement. An example might be the case of maple syrup urine disease where urine takes on the consistency of maple syrup; this alone is a reasonably good diagnostic marker so the knowledge that the urine has the odor of burnt sugar and fenugreek [8] is probably redundant. Oral malodor has long been cited as a potential application, having the advantage that an incorrect diagnosis is unlikely to lead to death of the patient. This is not to say that serious disease is not detectable by oral malodor [9]. Lung cancer, peritonsillar

Tab. 18.1 Potential electronic nose applications in the medical/healthcare field.

Intended use Author(s)/ Sensors employed Algorithms used Sample handling Findings References

Breath Halimeter Direct sampling monitoring Cell growth [13] Direct headspace Pirouette v2.7 Cultures in vials, Growth phases of mass spectrometry (PCA) headspace by E. coli defined by dedicated changes in volatile autosampler composition Eye infection [14] Polymer/Carbon PCA, FCM, SOM, Cultures in vials, Comparison of data black composite MLP, and RBF handheld sampling processing algorithms found RBF and MLP to be most applicable General medical/ [42] Karlsruhe Linear discriminant Direct Sweat sampling may healthcare microarray analysis (LDA) be useful in diagno- (KAMINA) stic applications Medical [15] 15 metal oxide LDA, least squares Direct sampling Discrimination environmental sensors (LS) and nearest above pure cultures between the fungi monitoring (e.g. sick neighbor neural was achieved along building syndrome) network (KNN) with discrimination between levels of characteristic volatiles Respiratory tract [43] MOSES PCA Headspace Discrimination (e.g. tuberculosis) II þ amperometric sampling achieved of sensors M. tuberculosis from controls Diabetes [9] 2 element MOS Non-supervised Direct sampling Discrimination of fuzzy clustering from patients diabetics from a expired breath normal population Breath alcohol [44] 10 MOSFET and Partial least Forced exhalation Evaluation of the

one IR CO2 sensor squares (PLS) into bags followed requirements of such and artificial by sampling a system for forensic neural network acceptability of breath (ANN) alcohol measurements using an electronic nose-type setup Leg ulcers [7] 20 conducting PCA Sampling of leg Demonstrated polymers ulcer dressings feasibility of the (presence of approach b-haemolytic streptococci) Cultured bacteria [45] 16 Conducting ANN, PCA Headspace from Good discrimination polymers 12 bacteria and achieved (Bloodhound) 1 yeast 18.1 Introduction 447

Tab. 18.1 Continued

Intended use Author(s)/ Sensors employed Algorithms used Sample handling Findings References

Cultured bacteria [46] ANN, Feature Petri dishes of 76 % classification extraction Escherichia coli, Enterococcus sp., Proteus mirabilis, Pseudomonas aeruginosa, Staphylococcus saprophytica Cultured bacteria [47] 6 MOS ANN Headspace from Discrimination and (Neotronics) Escherichia coli Sta- prediction of growth phylococcus aureus phase achieved Estrus in cows [18] Conducting polymer Wavelet analysis Swab in chamber Initial investigation abscess, and cancer of the larynx may all manifest themselves via oral malodor. How- ever, despite a great deal of funding, a successful breath odor device has yet to reach a clinic. Perhaps the principal reason for this is the suite of volatiles produced by the, typically, anaerobic bacteria causing malodor such as hydrogen sulfide, sulfur dioxide and methyl and dimethyl disulfide. The Halimeter system (Halimeter Interscan Inc., Chatsworth Ca, USA) does measure low parts per billion levels of hydrogen sulfide but is prone to several interferences such as ethanol, essential oils, perfumes and mouthwashes. Sulfur compounds have incredibly low human olfactory thresholds meaning that most people would become aware of the odor far quicker than the best of the sensing systems available. Coupled with this are the vast array of variables that need to be compensated for before an accurate measurement may be made; presence of environmental contaminants, patient to patient variability, perfumes, foodstuff in the oral cavity, hunger, tiredness etc. The oral malodor model features some important rules for the investigator into medical and health monitoring applications of electronic nose/sensor systems. A well-defined and controlled symptom is highly desirable. Phenomena such as bad-breath have ill-defined sources and as such are difficult to define sufficiently. This is especially significant when a volatile or combination of volatiles may characterize one or more phenomena. The use of smell in medical diagnostics and the development of systems for evaluation of odor in a medical context have been reviewed by Pavlou and Turner [10]. This article also provides a description of various odors associated with disease such as a stale beer odor on skin associated with tuberculosis and burnt sugar smells in urine associated with maple syrup urine disease. Hanson and Thaler patented a system based on an AromaScan A32S system for the monitoring of patients with lung infections such as pneumonia. The patent also dis- cusses the use of the system in evaluation of fluid samples from the sinus or nose for presence of cerebrospinal fluid [11]. The authors expand the cerebrospinal fluid work further suggesting that electronic nose technology may be used to distinguish cerebrospinal fluid from serum, having applications in the diagnosis of otorrhea or rhi- 448 18 Medical Diagnostics and Health Monitoring

norrhea, and may have further application in the field of otorhinolaryngology [12]. A novel detection means for diabetes detection based upon measurement of breath samples for acetone using a two detector metal oxide system is suggested by Ping [9]. Experiments describing responses before and after eating suggested good correlation between acetone concentration and diabetes. Paulsson et al. describe a breath alcohol analysis using metal oxide field-effect transistor (MOSFET) sensor technology [44]. As part of their evaluation they considered the requirements of applying such a system in the routine use of breath alcohol detection from a forensic standpoint. Changes in the odor of sweat has been proposed as a potential means of disease diagnosis using the KAMINA system [42]. Mantini et al. also present a study of sweat as a potential means of following the menstrual state of women, although the study was merely a demonstration of the idea rather than a clinical study. They also briefly describe an approach to a skin-sampling methodology and the evaluation of urine samples containing blood [48]. Esteves et al. [13] describe an investigation of the growth characteristics of Escher- ichia coli using the Agilent headspace sampling system. The authors present data showing how distinct growth phases may be monitored using principal component analysis (PCA) of selected portions of the mass spectrum acquired. It is suggested that the lower-molecular-weight fragments are more indicative of the growth phase (from cellular metabolism) whilst higher molecular weight fragments are derived from cellular components especially when higher sampling temperatures were used. The application of the Cyrano Sciences handheld electronic nose to the detection of bacteria implicated in eye infections was reported by Boilot et al. [14]. The bacteria investigated were E. coli, Staphylococcus aureus, Haemophilus influenzae, Strep- tococcus pneumoniae, Pseudomonas aeruginosa, and Moraxella catarrhalis. Simple PCA suggested broad discrimination between the six bacteria grown in culture and presented at various colony counts (discrimination based upon bacterial count was not however, reliably achieved). Further off-line analysis was then undertaken using a number of data-processing strategies; (PCA), fuzzy c-means (FCM), self-organizing maps (SOM), multi layer perceptron (MLP), radial basis function neural networks (RBF) and the fuzzy ARTMAP (adaptive resonance theory mapping) paradigm. Com- parisons on the usefulness of all of the approaches were made, with MLP and RBF algorithms being most useful overall. Significant development was cited as being necessary however, before a system could be developed for a truly near-patient system to be developed. Schiffman et al. described the use of a MOS-based system for the discrimination of cultures of common fungi (Aspergillus flavus, A. niger, Penicillium chrysogenum, Cladosporium cladosporoides and Stachybotrys chartarum) that are implicated in sick-building syndrome, toxic poisoning, and allergic reactions [15]. Discrimination of cultured fungi was achieved along with discrimination between volatiles known to be associated with the fungi (ethanol, 3-octanone, 3-octanol, 3-pentanone and 2- methyl-1-propanol). Dodd proposes the use of electronic noses as monitoring tools in conditions such as schizophrenia [17]. This differs considerably from the detection of a pathogenic condition as described previously, with the author suggesting volatiles from autoxidation 18.2 Special Considerations in Medical/Healthcare Applications 449 of arachidonic acid might present a diagnostic marker monitorable via mass spectrometry or electronic noses. Health monitoring is not exclusively used for humans – estrus in cows has also been studied [18] using a modified Osmetech sensor system coupled to a custom built humidity compensation system “the olfactory lens” (a device for measuring dynamic changes in order) using wavelet analysis to process the data.

18.2 Special Considerations in Medical/Healthcare Applications

Medical samples present all of the standard sample presentation problems and more. Chemical and food samples are relatively straight forward to analyze, providing they are not subject to biological change i.e. that there are no overt degradation processes occurring from sample to sample, or a characteristic off-odor or contaminant is present. In a similar field, food spoilage measurements also suffer from many of the effects discussed below. One of the principal difficulties is the variability of the sample. This is especially true if the samples are the patients themselves. Patient to patient variability is a huge factor in any sampling procedure. As described above for oral malodor measurements, any number of environmental and habitual factors can affect the measurement. Any effective electronic nose application must either select out these unaccountable variations or compensate for them by anticipating them. It is easy to envisage that the latter approach is fraught with difficulty although it is given that no measurement is truly free from interferences. Hence, the more commonly encountered broad-selectivity electronic nose model is not the optimum system design. When developing any system and approach, the final application of the system must be considered from the start. An at-patient system must be capable of being exactly that, delivering a reliable and reproducible result within a typical consultation time with the minimum of calibration and user expertise required. The presentation of the sample and its acquisition are critical parts of the process. Developing a system for the discrimination of bacteria in culture, for example, is not a viable end product since standard culture techniques will take the same length of time and produce equally valid results for less resource and probably higher reliability. Enhanced identification through the use of selective media might be a consideration, but this is equally achievable without resort to electronic nose technology, anti- biotic loaded culture plates for resistance checking are a simple example. Additionally, enhanced techniques such as matrix assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) would give superior characterization over a similar time span. Consequently, at-patient or direct patient-derived samples with the minimum of sample preparation are the most desirable approaches since they lend themselves to rapid turnaround, even if the technique is simply a screen to eliminate negatives prior to further investigation. An example where this is an attractive option is in screening for urinary tract infection, where, typically 60–80 % of the presenting po- 450 18 Medical Diagnostics and Health Monitoring

pulation is in fact negative. A technique for pre-screening samples prior to culture automatically saves time and frees resource. Further, in typical populations, 50 to 80 % of the infected population (i.e. between 20–40 % of those initially presenting) is infected with E. coli. A system capable of screening out E. coli positives would offer enormous potential benefit in cost and time saving. Once a technique has been identified, a number of other factors arise. The robustness of the technique and its performance at a clinical level must be considered. To be viable, any clinical/healthcare application would have to at least approach the performance of the current optimum methodology. As an example, a screen for urine infection would be no good no matter how quick it was if it was wrong 50 % of the time. Other factors such as cross-contamination, sample reproducibility, user and patient safety must also be considered.

18.3 Monitoring Metabolic Defects in Humans Using a Conducting Polymer Sensor Array to Measure Odor

18.3.1 Background

The odor of the human body and excreted or secreted products of metabolism is related to many complex factors associated with sex, age, genetics, diet, and metabolic condition. In many cases, bacterial or viral infection, or metabolic diseases modify these odors. Typical examples are bromidrosis in patients affected by rheumatism and uremia and diseases of the respiratory and digestive tract [19]. Some myopathies induce alterations in the metabolic pathway that cause an abnormal secretion of metabolites in blood as ketones and acids [20]. The diagnosis of such genetic diseases is based on gene analysis, muscle biopsy and testing muscle performance. Biochemical tests are carried out by enzyme analysis or by determining metabolites by HPLC gas chromatography coupled to mass spectrometry (GC-MS) or immunological methods [21]. The main objectives of this research were: (a) to determine whether it is possible to use an electronic nose as a diagnostic method for detection and monitoring metabolic diseases such as myopathies, (b) to carry out screening of samples from patients and controls, using GC-MS to identify chemical species that could be used as markers, for which an electronic nose device could be focused. In the course of this research we have been able to achieve the following breakthroughs in understanding how to apply methods based on odor recognition to medical diagnostics. We have identified specific volatile chemical markers in the urine of patients with specific metabolic disorders that are not present in controls, or are present at very different concentrations. We have been able to discriminate populations of diseased persons from controls by their odor fingerprint measured by an electronic nose, using urine samples. We have applied statistical and neural network methods to process data from such systems to enable the future on-line recognition of disease states. 18.3 Monitoring Metabolic Defects in Humans using a Conducting Polymer Sensor Array to Measure Odor 451

18.3.2 Methodology

One useful set of materials that may be utilized as sensors in an electronic nose is that of electrically conducting organic polymers based on heterocyclic molecules such as pyrroles, thiophenes and anilines. These display reversible changes in conductivity when exposed to polar volatile chemicals. Rapid adsorption and desorption kinetics are observed at ambient temperatures. The materials do not display high specificity to individual gases. However, they can be chemically tailored to enhance differences in response to particular classes of polar molecules. For single chemical species, the concentration-response profiles can be fitted to Langmuir type adsorption models. This is advantageous as simple computational methods may be used for information processing [22–24]. Different polymers made from modified monomer units show broad overlapping response profiles to different volatile compounds. Hence, arrays of these sensors should behave very similarly to olfactory sensor arrays in the biological system. Min- iature arrays consisting of up to 48 different conducting polymer materials have now been realized by Osmetech plc (see Fig. 18.1). A microprocessor-driven circuit, measuring changes in resistances of individual sensor elements interrogates the sensor array at user-defined intervals, and data are stored in memory. Each sensor element changes in resistance when exposed to a volatile compound. However, the degree of response to a given substance depends on the type of polymer element used, so that a pattern of resistance changes can be recorded and processed to produce a set of descriptors for that particular substance. The sensor responses are normalized to represent relative changes in resistance and thus approximately concentration-independent patterns can be produced. Taken over the whole array, there are enough statistical differences for many compounds to be differentiated from each other. Functionality of the system depends on devising robust computer programs that will allow the system to operate under adverse conditions whereby background odors may be present,

Fig. 18.1 Osmetech sensor array and electronics 452 18 Medical Diagnostics and Health Monitoring

temperature and humidity may be cycling up and down and sensor-aging effects may also be interfering. For these experiments, a sensor array with 32 different conducting polymers (Os- metech plc) was used for detecting odors from urine. A Hewlett Packard GS-MS (HP5890 GC/HP5971MS) apparatus was used for analyzing the volatiles and Supelco supplied fibers for solid phase micro-extraction (SPME). Odors were commonly monitored by static headspace GC and occasionally by thermal desorption or purge and trap techniques. We opted for use of the SPME method after testing static headspace, purge and trap, and thermal desorption methods for sampling odors. SPME is a powerful technique for introducing analytes into a GC. The technique utilizes a 1 cm length of fused silica coated with an adsorbent. The coated fused silica (SPME fiber) is immersed directly into an aqueous sample or into the headspace above a liquid or solid sample. Organic compounds in the sample are subsequently adsorbed onto the fiber. Finally, the fiber is inserted into a GC injector where the analytes are thermally desorbed and separated on the GC column. This technique is rapid and minimizes any sample manipulation.

18.3.3 Results

Replicate urine samples were taken from ten people affected by different muscular diseases and thirteen from healthy subjects over several days and frozen until they were analyzed. An electronic nose system was used to analyze the headspace from urine samples. To study the individual urine odor of a particular person, it is important to consider their temporary differences, caused by different diet, state of health, physiological condition etc. Thus urine samples were collected over a period of several days. Var- iance between urinary headspace of different individuals is significant, whereas for the same individual the profile over different days remains constant, as measured using the electronic nose system. We analyzed urine headspace in a normal population as well as in patients with myopathies by using the electronic nose and the GC-MS. The patterns obtained from the sensor array were recorded on a computer and stored for further processing. Urine samples collected from normal and diseased populations generated patterns that slightly differed between each person and showed some variation due to the physiological condition and to the diet. In order to process the data we adopted the Sammon map method [25]. The Sammon non-linear mapping algorithm reduces multidimensional pattern space by mapping onto two-dimensional or three- dimensional pattern space based on a distance measure such as the Euclidean distance and produces axes that are meaningful in terms of distances of one cluster from another. By using this method it was possible to differentiate the normal population from that with myopathies. Moreover we obtained subclusters within the population due to slight differences between them. This can be due to the different myopathies or degree of the pathological status. Figure 18.2 shows the population distribution obtained between controls and patients. Each point on the map represents an odor pattern reduced 18.3 Monitoring Metabolic Defects in Humans using a Conducting Polymer Sensor Array to Measure Odor 453

Fig. 18.2 Analysis of urine headspace: the population distribution obtained between controls and myopathic patients. The Sammon map represents in two dimensions the averaged Euclidean distance between urine headspace patterns for each individual tested, each point representing one individual. It is seen that the majority of patients group together, and the controls also group together separately, but there are two patients who group with the controls and one control individual who groups with the patients to two dimensions, and clusters represent how close each odor pattern is to another in the same area (the further away points are from each other, the greater the difference between them). The results obtained from the GS-MS analysis show that the composition of the urine headspace is markedly different within normal and diseased populations. Key volatile components found in the profiles of normal urine were 2-heptanone and 4-heptanone. The amount of these volatiles increases in urine samples from people affected with myopathies. Compounds such as 2(3H)phenanthrene-4-4a-9,10-ter- tahydro-4a-methyl and phenyl-isopropylphenyl ether are present in different quantities only in urine from patients and not in normal controls. We performed the GC-MS analysis of urine in order to validate the results obtained from the electronic nose. The different patterns obtained from the gas sensor apparatus are correlated with the different volatiles detected by the GC-MS, and their quantities. Knowing the composition of urine headspace will allow us to build specific sensors for diagnostic purposes. 454 18 Medical Diagnostics and Health Monitoring

18.4 The Use of an Electronic Nose for the Detection of Bacterial Vaginosis

18.4.1 Background

Bacterial vaginosis (BV) is a particularly ill-defined phenomenon with uncertain symptoms. Numerous reports [28–30] cite as much as 50 % of the affected population being asymptomatic. The consequence of this is that at time of presentation only 50 % of the story is known. The remaining 50 % of the population either go undetected or present during routine examination for another associated or uncorrelated problem. Initial investigations were performed by Chandiok et al. [26] using a standard AromaScan (now Osmetech) system at Withington Hospital, Manchester, UK. The consequences of BV are wide and varied and are not completely understood. This is understandable given the difficulties in getting reliable BV data for a population. The primary challenge facing any prospective diagnostic technique (or aid to diagnosis) is finding a unique indicator against which BV may be detected. Cur- rently, the Amsel test is the benchmark for determining the problem. The criteria for the test rely on at least three out of four conditions being met [27]. These are:

* pH of vaginal fluid > 4.5; * Presence of a typical thin, homogenous vaginal discharge; * Release of strong fishy smell on addition of alkali (10 % KOH) to a sample of vaginal fluid (whiff test); * Clue cells present on microscopic examination of a wet mount of vaginal fluid.

Individually none of these tests are diagnostic. pH variation of the vaginal fluid is nearly always present in BV positive patients but it is a non-specific test and the variation is equally likely to be caused by another infection or problem. Additionally, contamination of the sample by cervical mucus (typical pH 7) can lead to false diagnoses in some cases. pH variation also occurs as part of the natural menstrual cycle. Ethnic background is also a factor affecting vaginal pH and this has been used as a reason for the relatively higher number of black American women who present with symptoms of the disease. According to Hay, pH is highly sensitive (97 %) but very non- specific giving false positives in 47 % of cases [28, 29]. Conversely, discharge is very accurately recognized by clinicians giving false positives at 3 % but only has a specificity of 67 %. Following this, the ‘whiff’ test also gives low false positives (1 %) but is non-specific (43 %). Finally clue cells are typically found in 81 % of positive BV cases whereas 6 % of non-BV cases have positive clue cell tests. Other trials report variation on these figures but all concur with the non-specificity and reliability of any one individual test 30]. BV is commonly thought to arise as a result of fluctuation of the normal vaginal flora. In some cases the flora can fluctuate naturally over the menstrual cycle with no adverse effects. It is thought that one of the primary controlling mechanisms controlling BV-causative bacteria is the presence of adequate colonies of Lactobacillus sp. 18.4 The Use of an Electronic Nose for the Detection of Bacterial Vaginosis 455 that produce hydrogen peroxide which limits the growth of anaerobes associated with BV. The most common organisms associated with BV are: Gardnerella vaginalis, Bac- teroides (Prevotella) spp., Mobiluncus spp. and Mycoplasma hominis. However, the presence or absence of these flora is not reliably diagnostic. Treatment after diagnosis is usually quite effective and usually comprises oral doses of metronidazole. Topical treatments with metronidazole or clindamycin are also common [31]. Originally thought to be a benign infection, recent studies have linked the problem to increased risk of:

* intra-amniotic infection [32] * choroamnionitis [33] * post-caesarean [34] and post-partum endometritis [35] * adverse pregnancy outcome [36] * pre-term labor [36–38]and birth [39] * premature rupture of membranes at term [40] * post-hysterectomy cuff cellulitis [41].

The data presented here are merely an overview and the reader is directed to the literature cited for a more comprehensive discussion of the occurrence, diagnosis, and treatment of this phenomenon.

Fig. 18.3 Example of BV swab in vial and presentation of dual concentric needle sampling system 456 18 Medical Diagnostics and Health Monitoring

18.4.2 Methodology

Patient samples were taken as swabs during normal examination. Swabs were weighed before and after sampling to determine the amount of sample collected during the examination so that variation in sample collection may be evaluated. After collection the stem of the swab was cut off and the vial sealed with a standard septum and crimp top (see Fig. 18.3). After collection the vials may be stored for later analysis or analyzed immediately. The sample is mounted on the carousel of the autosampler system and held at constant temperature until its place in the sequence is reached. The sample is then lowered to a pre-heated platen and its temperature stabilized for a predetermined period before the dual concentric needle is lowered into the vial through the septum and the dynamic headspace extracted using a constant humidified gas flow. The headspace is transferred across the Osmetech sensor array (see Fig. 18.1) where the signal is transduced and recorded for processing.

18.4.3 Results

Results can be produced from the Osmetech Microbial Analyser (OMA) within 20 minutes (as can results from the Amsel test). The microbiology (Nugent score) results can take much longer and in some cases it can be five days before the results are transmitted back to the Genito-Urinary Medicine (GUM) clinic. The results from Table 18.2, which were derived from Fig. 18.4, give an overall sensitivity of 89 % and a specificity of 87 % versus Amsel and Nugent scores with a negative predict value of 96.8 % and a positive predict value of 65 %. Samples projected on the PCA map labeled as Suspect BV were the result of indeterminate microbiology and Amsel results (i.e. the two were not in agreement). As a result of the data available during the conduct of the clinical evaluation it was not possible to follow the suspect patients up to confirm any further clinical diagnosis.

Tab. 18.2 Results from a clinical pre-trial carried out at the department of Genito-Urinary Medicine, Withington Hospital, Manchester, UK using the OMA instrument (BV ¼ bacterial vaginosis, STDs ¼ se- transmitted diseases). The results in this table and Fig. 18.3 are for 89 newly registered non-pregnant females between the ages of 18 and 60.

Total Positives False positives Negatives False negatives

BV 16 15 1 Suspect BV 3 2 1 Yeasts 13 3 10 Negative 48 6 42 STDs 9 9 18.4 The Use of an Electronic Nose for the Detection of Bacterial Vaginosis 457

Fig. 18.4 PCA map of patient swabs from a GUM clinic, positives are based upon combined agreement of Nugent and Amsel scoring systems. Intermediate BV is assigned to tests where the Nugent and Amsel scores do not concur

It is clear from the PCA map presented that sexually transmitted diseases (STDs) are not confused with BV status with all STDs projected in the negative BV sector. As previously stated in the earlier discussions in this chapter it is highly desirable that any clinical electronic nose application should display a high degree of selectivity to the target application. In the case of STDs this is clearly the case, although yeasts did have a tendency to produce false positives (approximately 23 % of yeasts analyzed produced a false positive for BV). However, it is highly likely that the healthcare professional carrying out the test would discriminate between yeast and BV before testing for BV using the OMA system.

18.4.4 Discussion

It is clear from the results presented in Fig. 18.4 and Table 18.1 that BV-positive patients differ from the normal population when the data is processed as a PCA map. The data can be seen to ‘branch’ into two categories away from the defined normal population. These two branches may be described by the use of two standards that are differentiated by means of sensor elements in the Osmetech array responding orthogonally to the test chemicals used. These chemical standards may be used to define a PCA map onto which the experimental data is processed. The threshold between positive and negative results is subsequently defined by means of experimentally defined parameters in the first instance, and then by adjustment of the chemical standards to reflect the threshold giving the clearest distinction between BV positive and negative patients. 458 18 Medical Diagnostics and Health Monitoring

The use of standardized test chemicals is critical to the success of the procedure. The standards may be used to produce a master projection against which all subsequent test chemical data may be projected. Subsequent system calibrations or checks may then be projected against this master such that the current system performance may be compared against the original blueprint. Significant change to the mapping of the standards will flag substandard performance. Failure to pass this system check procedure would prevent further use of the system until a suitable remedy has taken place. This imparts an inherent strength to the system allowing greater faith in the accuracy of any prediction made. Hence it can be seen that through the use of surrogate test chemicals a medical device can be successfully used without resort to complicated drift correction and standardization algorithms.

18.4.5 Conclusion

The OMA system in this case offers clear potential in the rapid diagnosis of BV. It can clearly compete with the established means of detection and in the case of Nugent scoring is a much more rapid technique. With further development it should prove superior to the existing Amsel technique offering the advantages of ease of use and removal of doubt from interpretative testing such as sniffing (the whiff test) and visual inspection (examination for the presence of clue cells).

18.5 Conclusion

It will be apparent from the information presented in this chapter that enormous potential exists for the application of electronic nose technology in medical applications. However, the field is still in the research and development stage, where clini- cally proven robust applications are still to come. There is now rapid growth in capability of the technology and it is clear that many future diagnostic tools for selected applications will be available for physicians to utilize. Indeed, Osmetech plc has submitted an application for approval of its urinary tract infection technology to the FDA after a series of successful clinical trials.

Acknowledgements This work was in part supported by Osmetech plc, Crewe, UK. AMP was funded by the Wellcome Trust for work on myopathies. We thank Dr. Ros Quinlivan, of Oswestry Hospital for co-operation and help with patient samples, Prof. Robert Beynon and Dr. Duncan Robertson for much help with GC-MS analysis. 18.4 The Use of an Electronic Nose for the Detection of Bacterial Vaginosis 459

References

1 G. F. Hayden. Postgraduate Medicine 1980, 15 S. S. Schiffman, D. W. Wyrick, G. A. Payne, 67(4), 110–5, 118. G. O’Brian, H. T. Nagle. Detecting Microbial 2 W. Z. Stitt, A. Goldsmith. A. Archives Of Contamination using an Electronic Nose, in Dermatology 1995, 131(9), 997–999. ISOEN200 abstracts, Persaud, K. C.; Gard- 3 A. Zlatkis, R. S. Brazell, C. F. Poole. Clinical ner, J. W., editors; ECRO, Indigo Lithoprint: Chemistry 1981, 27(6), 789–797. Manchester, UK, 2000. 4 A. Zlatkis, C. F. Poole, R. Brazell, K. Y. Lee, 16 R. T. Marsili. Journal Of Agricultural And F. Hsu, S. Singhawangcha. Analyst 1981, Food Chemistry 1999, 47(2), 648–654. 106(1260), 352–360. 17 G. H. Dodd. Prostaglandins, Leukotrienes 5 D. G. Burke, B. Halpern, D. Malegan, Essential Fatty Acids 1996, 55(1 þ 2), 95–99. E. McCairns, D. Danks, P. Schlesinger, 18 T. T. Mottram, R. M. Lark, A. J. P. Lane, B. Wilken. Clinical Chemistry 1983, 29(10), D. C. Wathes, K. C. Persaud, M. Swan, 1834–1838. J. M. Cooper. Techniques to Allow the 6 P. Grametbauer, S. Kartusek, O. Hausner. Detection of Oestrus in Dairy Cows with an Ceskoslovenska Epidemiologie, Mikrobiologie, Electronic Nose, in Electronic Nose and Imunologie 1988, 37(4), 216–223. Olfaction 2000, Gardner, J. W.; Persaud, 7 A. D. Parry, P. R. Chadwick, D. Simon, K. C., editors; IOP Publishing: Bristol, UK, B. Oppenheim, C. N. McCollum. Journal 2000; pp. 201–208. Of Wound Care 1995, 4(9), 404–406. 19 M. Inaba, Y. Inaba. Human Body Odor. 8 K. Monastiri, K. Limame, N. Kaabachi, Etiology, Treatment and Related Factors.; H. Kharrat, S. Bousnina, H. Pousse, Springer Verlag: Berlin, 1992. M. Radhouane, M. N. Gueddiche, 20 H. Chen, F. Aiello. Amer. J. of Med Genetics N. Snoussi. Journal Of Inherited Metabolic 1993, 45, 335–339. Disease 1997, 20(4), 614–615. 21 M. A. Hollinger, B. Sheikholislam. The 9 W. Ping, T. Yi, H. B. Xie, F. R. Shen. Bio- Journal of International Medical Research sensors & Bioelectronics 1997, 12(9–10), 1991, 19, 63–66. 1031–1036. 22 P. Pelosi, K. C. Persaud. Gas sensors: 10 A. K. Pavlou, A. P. F. Turner. Clinical Towards an artificial nose. In: Sensors and Chemistry and Laboratory Medicine 2000, Sensory Systems for Advanced Robots., 38(2), 99–112. in NATO ASI Series F: Computer and Systems 11 C. W. Hanson, R. Thaler. 09965386 WO, Science, Dario P, editor; Springer-Verlag: 1999. Berlin, 1988; pp. 361–382. 12 E. R. Thaler, F. C. Bruney, D. W. Kennedy, 23 K. C. Persaud. Analytical Proceedings C. W. Hanson. Archives of Otolaryngology (London) 1991, 28(10), 339–341. Head and Neck Surgery 2000, 126(1), 71–74. 24 K. C. Persaud. Trends in Analytical Chemistry 13 R. Esteves de Matos, D. J. Mason, C. S. Dow, 1992, 11(2), 61–67. J. W. Gardner. Investigation of the Growth 25 J. W. Sammon Jr.. IEEE Transactions on Characteristics of E. coli using Headspace Computers 1969, 5(C-18), 401–409. Analysis, in Electronic Nose and Olfaction 26 S. Chandiok, B. A. Crawley, B. A. Oppen- 2000, Gardner, J. W.; Persaud, K. C., editors; heim, P. R. Chadwick, S. Higgins, K. C. IOP Publishing: Bristol, UK, 2000; Persaud. Journal Of Clinical Pathology 1997, pp. 181–188. 50(9), 790–791. 14 P. Boilot, E. L. Hines, S. John, J. Mitchell, 27 R. Amsel, P. A. Totten, C. A. Spiegel, F. Lopez, J. W. Gardner, E. Llobet, M. Hero, K. C. Chen, D. Eschenbach, K. K. Holmes. C. Fink, M. A. Gonogora. Detection of American Journal Of Medicine 1983, 74(1), Bacteria Causing Eye Infections using a 14–22. Neural Network Based Electronic Nose 28 P. E. Hay, D. Taylor-Robinson, R. F. Lamont. System, in Electronic Nose and Olfaction 2000, British Journal Of Obstetrics And Gynaecology Gardner, J. W.; Persaud, K. C., editors; 1992, 99(1), 63–66. IOP Publishing: Bristol, UK, 2000; 29 P. E. Hay. Dermatologic Clinics 1998 16(4), pp. 189–196. 769–773. 460 18 Medical Diagnostics and Health Monitoring

30 J. L. Thomason, S. M. Gelbart, 42 U. Kruger, R. Ko¨rber, J. Ziegler, R. J. Anderson, A. K. Walt, P. J. Osypowski, J. Goschnick. Prospective experiments F. F. Broekhuizen. American Journal Of to determine sweat odour with a gradient Obstetrics And Gynecology 162(1), 155–160. microarray, in ISOEN 2000 Abstracts, 31 P. E. Hay. Journal of Antimicrobial Chemo- Persaud, K. C.; Gardner, J. W., editors; therapy 1998, 41(1), 6–9. ECRO Indigo Lithoprint: Manchester, 2000; 32 D. H. Watts, M. A. Krohn, S. L. Hillier, pp. 47–48. D. A. Eschenbach. Obstetrics And Gynecology 43 J. R. Stetter, W. R. Penrose, C. McEntegart, 1990, 75(1), 52–58. R. Roberts. Prospects for infectious disease 33 D. H. Watts, D. A. Eschenbach, G. E. Kenny. diagnosis with sensor arrays, in ISOEN 2000 Obstetrics And Gynecology 1989, 73, 52–60. Abstracts, Persaud, K. C.; Gardner, J. W., 34 M. G. Gravett, H. P. Nelson, T. DeRouen, editors; ECRO Indigo Lithoprint: Manche- C. Critchlow, D. A. Eschenbach, ster, 2000; pp. 101–104. K. K. Holmes. JAMA 1986, 256(14), 44 N. Paulsson, E. Larsson, F. Winquist. Sensors 1899–1903. And Actuators A-Physical 2000, 84(3), 35 S. Faro. Journal Of Reproductive Medicine 187–197. 1989, 34(8 Suppl), 602–604. 45 T. D. Gibson, O. Prosser, J. N. Hulbert, 36 M. G. Gravett, D. Hummel, D. A. Eschen- R. W. Marshall, P. Corcoran, P. Lowery, bach, K. K. Holmes,. Obstetrics And Gyne- E. A. Ruck-Keene, S. Heron. Sensors And cology 1986, 67(2), 229–237. Actuators B-ChemicalK 1997, 44(1–3), 37 J. A. McGregor, J. I. French, R. Richter, 413–422. A. Franco-Buff, A. Johnson, S. Hillier, 46 M. Holmberg, F. Gustafsson, F. N. Judson, J. K. Todd. American Journal E. G. Hornsten, F. Winquist, L. E. Nilsson, Of Obstetrics And Gynecology 1990, L. Ljung, I. Lundstrom. Biotechnology 163(5 Pt 1), 1465–1473. Techniques 1998, 12(4), 319–324. 38 J. A. McGregor, J. I. French. Obstetrical And 47 J. W. Gardner, M. Craven, C. Dow, Gynecological Survey 2000, 55(5 Suppl 1), E. L. Hines. Measurement Science & S1–19. Technology 1998, 9(1), 120–127. 39 D. E. Soper, R. C. Bump, W. G. Hurt. 48 A. Mantini, C. DiNatale, A. Macagnano, American Journal Of Obstetrics And Gyneco- R. Paolese, A. Finazzi-Agro, A. D’Amico. logy 1990, 163(3), 1016–1021. Critical Reviews in Biomedical Engineering 40 C. A. Spiegel. Clinical Microbiology Reviews 2000, 28(3–4), 481–485. 1991, 4(4), 485–502. 41 R. L. Cook, G. Reid, D. G. Pond, C. A. Schmitt, J. D. Sobel. Journal Of Infectious Diseases 1989, 160(3), 490–496. 461

19 Recognition of Natural Products

Olivia Deffenderfer, Saskia Feast, Franc¸ois-Xavier Garneau

Abstract The application of sensor-array analysis to natural products is still in its infancy. This chapter seeks to provide an overview of the work that has been accomplished on natural products, and to discuss various sampling and instrument setup considerations that apply in this arena. Two examples of the application of a polymer-composite sensor-array-based electronic nose to the identification of natural products are described. In one study, the CyranoseTM 320 accomplishes the sorting of wood species, jack pine, balsam fir, and black spruce, used in the lumber industry. In the second study the volatile natural compounds from essential oils are used to distinguish closely related species of plants.

19.1 Introduction

Electronic noses provide a powerful modern analytical technique that addresses many safety, quality, and process challenges facing manufacturers. Since their introduction in the early 1990s there have been many advances in sensor technology and data processing procedures used in electronic noses, coupled with a much greater understanding of the appropriate applications for this technology. This chapter provides an overview of this modern analytical tool for applications in natural products. Many of the natural products we shall discuss are also used in the food industry and are covered from a food quality perspective in Chapter 21. Using an electronic nose in natural product applications can be challenging. Sam- pling, sensor technology, sensitivity, and the inherent variability of natural products are some of the concerns.

19.2 Recent Literature Review

Electronic noses have been used for many applications from identifying solvents to natural products. A summary of the recent publications on the applications of electronic noses to natural products is included in Table 19.1. It is clear that most of the work to date on natural products has focused on those that we consume. The applications include discrimination of spirits, beverage quality, fruit ripeness and quality, grain quality, meat and fish freshness, and oil quality; all types of sensors and a variety of data processing tools are used. The electronic nose sensor technology used most often to sniff natural products, were metal oxide sensors (MOS) and conducting polymers (CP), or combinations of different sensing technology. Quartz crystal microbalance (QMB), surface acoustic wave (SAW), and mass spec- trometery (MS)-based electronic noses have also been tested. The most common data analysis tool used was either principal component analysis (PCA) or cluster analysis to easily visualize the differences between samples. Neural networks (NN) and factor or discriminant analysis (DA) as well as regression techniques were used to test models. One of the main differences between the various studies is the sampling technique. Though the general process of sampling, such as placing the samples in a sealed container, allowing headspace to equilibrate, and presenting the sample to the electronic nose was similar for many applications, the method in which this was done varied greatly.

19.3 Sampling Techniques

One of the major components of successfully using an electronic nose is sample preparation. Sample containment, treatment, conditioning, storage, and seasonal variations all impact the results of experiments performed with electronic noses.

19.3.1 Sample Containment

Typically all the samples need to be contained. These containers vary from simple vials or jars to more sophisticated headspace vials for auto samplers. Electronic noses sample headspace, hence knowledge of headspace generation and consistency is necessary to develop the methods. 19.3 Sampling Techniques 463

Tab. 19.1 Review of recent literature on electronic-nose applications in natural products

Application Sensor Sampling Data Analysis Findings

Toasting level 6 MOS Headspace sample taken PCA, An electronic nose would be of oak wood from above hot barrel discriminant useful in process barrels [1] immediately after toasting. function monitoring of the toasting analysis level of oak wood barrels. (DFA), NN Fermentation- eNOSE 4000 10 mL samples were DA Media spoilage, contamina- bioprocess (Neotronics) placed in 500 mL glass tion, and microbial conta- monitoring [2] 12 CP sample vessels and mination could be detected tested at 30 8C. earlier than other conventional methods using an electronic nose. Sterilization level and inoculation level could not be discriminated. Freshness 6 MOS 10 s baseline, 50 s PCA Sensitivity decreased with of soybean sample, 40 8Cs higher temperatures. The curd [3] ample extraction. electronic nose was able to predict freshness of soybean curd over time. Cheese eNOSE 5000 Various electronic noses CDA MOS discriminated well but ripening [4, 5] (12 CP 8 were used to test the ripe- were ‘poisoned’, CP and MS MOS); ning of four Swiss Emmen- poor sensitivity resulting in 6 QMB; tal cheeses over a period of poor discrimination, QMB 10 MOS- one year. Static heasdspace no discrimination, MOS- FET þ measurements were taken FET alone gave poor discri- 5 MOS; in first study. SPME was mination but with MOS was Smart used for pre-concentration good system. MS with Nose (MS) in second study. SPME was best method in discriminating cheeses because of repeatability, simplicity, autosampler capability. Milk spoilage 14 CP Samples allowed to BP-NN, Study shows promise in (yeast/bacteria) (Bloodhound) equilibrate for 30 min. DFA, PCA, using an electronic nose for [6] A charcoal filter was used canonical detecting milk spoilage. and the samples were analysis (CA) ‘bubbled’. Espresso Pico-1 Coffee was ground and PCA, ANN There was noticeable drift (seven blends) (five thin- static headspace was that needed to be corrected. [7] film MOS) sampled. 95 % correct predictions when two similar classes were combined as one class. Espresso Four thin- Espresso beans and PCA, MLP Whole beans: 100 % beans/ground/ film tin ground beans were placed ANN, data is correct classification with liquid [8] oxide in 20-mL vials. Liquid drift two sensors. Ground coffee was extracted at highcorrected coffee: 87.5 % correct clas- pressure then placed in sification. vials. Samples equilibrated Liquid coffee: unsuccessful. in vial at 50 8C for 30 min before sampling. 464 19 Recognition of Natural Products

Tab. 19.1 Continued

Application Sensor Sampling Data Analysis Findings

Coffee [9] 12 MOS Static headspace sampling Fuzzy 97 % of samples were of roughly 30 samples of ARTMAP accurately classified as a three roasted coffees result of data processing. (data from 1992 article). Vanillin Ion-trap MS Juice samples were PCA, DFA Vanillin limit of detection fortified grape- chemical spiked with 40 to was 40 ppm with classifica- fruit Juice [10] sensor 2000 ppm vanillin. tion possible at 100 ppm.

Fruit ripeness Tin oxide Peaches, pears and apples NN A sealed chamber was used monitoring [11] placed into a plastic box. to increase signal. Peach and 150 mL headspace was pear ripeness could correctly pulled out with a gas tight be determined more than syringe after 1 hr equilibra- 92 % of the time. Apple ri- ting. Sensors were allowed peness could not be deter- to stabilize for 10 min. They mined well. were purged with synthetic dry air. Fruit quality Thickness Slices of peaches and PCA and Discrimination evident be- [12] shear mode nectarines in sealed glass Learning tween fruits that had been quartz bottles and allowed to Vector classified by a sensory panel. resonators equilibrate for Quantization coated with 10 min at 30 8C. neural network pyrrolic macrocycle Tomato aroma e-NOSE 4000 Ripe tomatoes were stored MVDA The electronic nose was able [13] (12 CP) at 5, 10, 12.5, and 20 8C (CDA) to detect differences be- and tested over 12 days. 20 g tween ripe tomatoes stored of frozen tomato puree was in different conditions. The placed in sealed 113-mL results from the electronic cups and thawed in 25 8C nose corresponded with water bath. Then the sample sensory panel results. was placed into the electronic nose sampling glass. The electronic nose was purged for 4 min, allowing headspace to equilibrate. Soft-rot Two MOS Ambient conditions were Threshold One tuber with soft rot in a detection in and three 4 8C and 85 % RH. 1 Kg te- storage crate of 100 kg good potato tubers MOS sted in Quickfit jar, 25 Kg in tubers could be detected. An [14] (two experi- paper sack with diseased inoculated tuber, not sho- ments) tuber at bottom of sack, wing signs of soft rot, could 100 Kg tested in storage also be detected within 10 kg crate. of good tubers. 19.3 Sampling Techniques 465

Tab. 19.1 Continued

Application Sensor Sampling Data Analysis Findings

Oatmeal Fox 3000 1 g of oatmeal was PCA, SIMCA Hexanal is main rancidity oxidation [15] (Alpha MOS) placed in 10 mL vial and marker. Small variations in incubated at 100 8C for volatile profile were seen 30 min. Compressed air among samples analyzed was used as carrier gas. with an electronic nose. Triplicate/quadruplicate After six weeks of storage, analyses performed for differences could be seen each sample. Oatmeal had between different pack- been packaged in four aging. Two to four weeks different pouches, some was not long enough. designed to prevent/delay rancidity. Barley grain 10 MOSFET, 10 samples with normal PCA, PLS, SIMCA used to classify if quality [16] six MOS, one odor and 30 with off PLS-DA, samples had off odor. E-

CO2 monitor odor. 3 33 g samples of SIMCA nose: 3/40 misclassified, each class were heated to GC-MS: 6/40 misclassified. 50 8C. Baseline and purge PLS used to predict ergo- with zero air. sterol with high confidence and CFU level, which could not be predicted well from naturally infected grain. Cereal quality BH114, Cultures grown for 48, PCA, DA, Classification of grain [17] Blood hound, 72, and 96 hours on CA quality may be a possibility 14 surface- wheat meal agar. Single using electronic-nose tech- responsive replicate petri plate cultures nology. May be a simple and polymer placed in 500-mL sampling fast way to detect and diffe- arrays bags filled with 300 mL rentiate between strains and sterile air. Samples equili- species of fungus. brated for 1 hr at 25 8C. Sampled in a 25 8C constant temperature room. Wheat 16 40 g of grain heated to Nearest (k-NN) classified 68 % cor- classification electro- 60 8C in sealed glass neighbor rectly and NN classified by grade [18] chemical container. 10 L of air (k-NN), NN 65 % correctly. After data circulated through contai- correction for instrument

ner, ice trap and liquid N2 changes NN improved to trap. Volatiles from traps 83 %. NN outperformed evaporated into air and k-NN. saved into tedlar bags. Wheat quality CP array Wheat samples were RBF-ANN 92.3 % correct classification [19] made artificially moldy (92 samples (40 samples) with no bad in the laboratory. in training) samples misclassified as good. 466 19 Recognition of Natural Products

Tab. 19.1 Continued

Application Sensor Sampling Data Analysis Findings

Rice quality 10 MOSFET Rice varieties from two PCA Differences related to the [20] and 12 MOS crop years were studied. rice variety and age were 5 g placed into 20-mL observed, but the varietal vials sealed with Teflon differences were small in lined septa and caps. comparison to differences in Sample kept at room age. The electronic nose may temperature prior to be used to monitor aging or analysis and at 50 8C shelf-life of rice. during sampling. Capelin FreshSense Headspace gas above PLS1, The total volatile base value spoilage for (Nine electro- capelin was sampled at saturated of capelin stored under dif- fishmeal chemical gas 0and58C in storage generalized ferent conditions could be production sensors) conditions. 1 kg of capelin linear model predicted with an electronic [21] was placed in 5.2 L contai- nose. ner. Sensors reached steady state within 10 min. Mahi-mahi AromaScan The fish was stored at MDA using The quality changes in ma- freshness [22] (32 CP) 1.7, 7.2, 12.8 8C for 0, 1, AromaScan hi-mahi using the electronic 3, 5 days and analyzed A32S nose correlated with sensory with AromaScan. 10 g of Windows panel results and microbio- fish was placed in a bag. software logical analysis. The elec- The bag was evacuated v. 1.3 tronic nose was also able to and filled with carbon-filte- predict different grades of red air and allowed to equi- mahi-mahi stored at 7.2 8C. librate for 10 min at 35 8C. The baseline was dried with Silica gel. Carbon-filtered ambient air was reference air. Sensors were purge- dwith headspace from 2 % 2-propanol and allowed to react with reference air for 2.5 min before next sample Chicken 8 MOS The chicken was placed in NN The electronic nose could freshness [23] glass sampling containers. predict freshness within 40 min of actual time using one sensor and 20 min using eight sensors Minced-meat HP 4440 Minced beef was stored PCA The electronic nose was able rancidity [24] at 4 8C with lighting and to measure the development storage equivalent to a of rancidity in minced beef retail store. over 17 days. Swine products FOX 2000 Samples of subcutaneous LDA, Swine products could be [25] six MOS adipose tissue were min- SIMCA classified with an electronic ced and frozen. For testing, nose based on what the 0.5 g was placed in a 10-mL swine were fed. (feed, glass vial. Synthetic air in- feed þ acorn, acorns alone). jected to remove ambient air and the sample equilibrated at 35 8C for 7 min. 19.3 Sampling Techniques 467

Tab. 19.1 Continued

Application Sensor Sampling Data Analysis Findings

Olive oil Eight CP 2 mL of oil was placed in PCA Five oil qualities could be quality [26] 10 mL vials for static head- discriminated with 90 % space sampling. The samp- confidence interval. Five les equilibrated at 50 8C for different oils could be dis- 9 min prior to testing. criminated with 90 % confidence interval. Frying fat Four MOS The fats were aged in Line plots Fat is deteriorated if level of quality [27] air at 180 8C. Vapor were used to polar compounds exceeds passed through a GC compare 24–27 %. The results show separation column, then results of good correlation with the the flow of gas was split MOS sensors Food oil sensor. Water in- between FID and MOS to reference fluence could be removed chamber. Fat was hot food oil and there was no interfe- during sampling. sensor. rence from different foods cooked in oil. Corn oils [28] AromaScan Samples were stored in PCA The electronic nose was (32 CP) 50 mL beakers at 60 8Cin successful in detecting off- dark. Testing occurred on odors that were produced by days 0, 4, 8. The total sample oxidation. time was 200 s with a 30 s purge with 2 % IPA vapor followed by a 30 s purge with water vapor. Maize corn oil MOSES II: Sampling parameters not PCA The limit of detection was rancidity [29] eight MOS, outlined in article but are the 1 ppm of aldehyde in oil. eight QMBs same as the parameters used in similar GC/MS headspace analysis. Tansy essential 32 CP 0.5–1 mL of oil was placed PCA Good discrimination was oil [30] in 8-mL glass vials with seen between three chemi- Teflon septum cap. Samples cal varieties of Tansy es- were left for 1 hr at room sential oil using an temperature. Each sample electronic nose. was sampled five times in random order. Golden Rod es- 32 CP Refer to Tansy Oil PCA In less than 30 s per sample, sential oil [31] sample preparation [31] essential oils of three Gol- den Rod species could be discriminated using an electronic nose. Wood chip sor- 32 CP Pieces of wood were broken PCA An electronic nose was rating [32] and placed into 250 mL pidly able to discriminate sealed glass jars. The and identify black spruce, samples were tested at room balsam fir, and jack pine. temperature. Each sample was sampled five times in random order. 468 19 Recognition of Natural Products

19.3.2 Sample Treatments

Although an electronic nose may be a ‘point and sniff’ device for certain applications, additional sample treatment is often required for natural-product applications. Heat- ing, preconcentration, and grinding are methods used to increase the volatiles in the headspace. Cooling can be used to prevent or slow spoilage over time. Removing a base component can improve sensitivity to slight differences in samples. In the following paragraphs, applications using these sample treatments are discussed.

19.3.2.1 Heating A natural-product application for electronic noses is determining oil quality, which is often done organoleptically. Cooking oils tend to have little or no odor, are not volatile, and have a low vapor pressure; it is therefore difficult to use electronic noses to detect oil. However, off odors in oil can be volatile. In several studies, electronic noses have been used to detect the rancidity of oil. Shen determined that an array of 32 CP sensors could detect odors produced by oxidation of corn oil [28] and Frank determined that an array of eight MOS and eight QMB could detect as little as 1 ppm of aldehyde in corn oil [29]. In another study, discrimination of flat, musty, rancid, fusty, and muddy olive oil could be determined with 90 % confidence using conducting polymer sensors and PCA [27]. All studies were conducted in sealed containers and the samples were heated to a minimum of 50 8C.

19.3.2.2 Cooling Because meat can spoil rapidly it is essential to keep the samples cool. Process-line monitoring would also require the sensors to perform at cool temperatures. Several studies have been done on fish freshness, while keeping the fish in cool conditions. An electronic nose using an array of electrochemical gas sensors, FreshSense, has been specifically designed to detect the volatiles resulting from the spoilage of fish. These studies were usually done at normal storage conditions, between 0–7.2 8C. In another study, an electronic nose with CP sensors was used to evaluate the freshness of mahi- mahi fillets [22]. The electronic nose results correlated with sensory panel results as well as microbiological analysis, and were successfully used to predict different sensory grades of mahi-mahi stored at 7.2 8C.

19.3.2.3 Removal of Base Component Another sampling technique was used to discriminate different brands of beer. Etha- nol is present in beer in high concentrations masking slight differences between beers. In this case, the ethanol was ‘pre-separated’ from the beer. The remaining components were presented to an electronic nose with eight QMB sensors resulting in good discrimination between brands using PCA [33]. 19.3 Sampling Techniques 469

19.3.2.4 Preconcentration Preconcentration is a technique used to concentrate volatiles prior to testing, and is most commonly used for gas chromatography (GC)-MS headspace analysis. Types of preconcentration include solid-phase micro extraction (SPME), direct thermal desorption, purge-trap, and cyrotrapping. Preconcentration using SPME was typically performed for the electronic nose applications studied. Using this technique, there was improvement in the ability to discriminate cheeses. Schaller examined a variety of electronic nose technologies to test the ripening of Swiss Emmental cheese over the period of a year [4]. A MOS sensor array alone and a MOS-field-effect transistor (MOS-FET) plus MOS sensor array resulted in a good assessment of cheese ripeness. However, the MOS sensors were ‘poisoned’ over time by the vapor. No discrimination was seen using a QMB array or CP array. The MS-based electronic nose was not sensitive enough. However, when the cheese vapor was pre-concentrated on an SPME fiber, good discrimination was seen using the MS electronic nose [5]. Another example where preconcentration is used is in the discrimination of similar wines. Good discrimination between different types of alcoholic beverages such as beer, wine, spir- it, and samshu could be obtained with a relatively simple sampling method and eight SAW sensors [34]. Predictions of unknown samples using a back-propagated ANN were also successful. However, discrimination between similar alcohols, such as organoleptically similar wines [35] or beer [34] required sampling technique improvement. Wines from the same region with a similar taste were discriminated using SPME fiber to concentrate the headspace before being presented to an electronic nose with 12 CPs [35].

19.3.2.5 Grinding Grinding or crushing a solid sample creates more surface area, therefore a greater concentration of volatiles can be released into the headspace. This will reduce problems created by headspace depletion and low volatile solids. In one study, different brands of espresso were classified by looking at whole beans, ground coffee, and brewed coffee. Classification by espresso brand was 100 % correct for whole bean samples and 87.5 % correct for ground coffee using a NN. Classification was not successful when brewed coffee was sampled [8]. Two other studies showed similar results with 95 % [7] to 97 % [9] prediction accuracy for roasted coffee samples. In this case grinding the coffee did not enhance prediction ability over using whole beans. How- ever, grinding coffee is a better sample preparation method than brewing coffee. This example illustrates the importance of finding the best sample preparation technique for the application.

19.3.3 Instrument and Sample Conditioning

Instrument and sample conditioning are also important parts of the sampling technique when using an electronic nose. This section refers to the pathway between the sample and the sensors. Modification of the baseline, purge technique, and temperature control in the instrument are discussed. 470 19 Recognition of Natural Products

19.3.3.1 Modifying Baseline Some electronic-nose systems use gas cylinders to supply a constant baseline. How- ever, many electronic noses, including portable ones, draw the baseline from the ambient air. Many modifications can be made to the baseline measurement including drying, humidifying, and filtering. A dry baseline is important when sampling very dry products, such as dried spices, with a sensor array that responds to ambient moisture in the baseline. For example, the baseline air was dried with Drierite (calcium carbonate) when discriminating between two types of whole dried black peppercorns from different origins. Using a dry baseline improved the response allowing discrimination and identification of unknown samples over a period of 13 days [36]. A humidified baseline can improve the sensitivity of an electronic nose to similar compounds in aqueous solutions such as beverages. Filtering the baseline can be especially important when using a portable electronic nose in the field. A charcoal filter cleans the baseline air that may be contaminated by factory, fuel, or other strong odors, preventing the sensors from responding to the baseline vapor.

19.3.3.2 Purge Technique Following a sample, the sensors need to be cleaned to return back to baseline prior to the next sample. This is imperative in order to prevent cross contamination of samples or carryover. Different methods are used to wash or purge the sensors after sampling. Often ambient or dry air is passed over the sensors for a period of time to clean the sensors of any remaining sample vapor. However, in determining mahi-mahi freshness [22] and corn-oil freshness [28], the sensors were purged with 2 % isopropyl alcohol in water vapor followed by a second purge of only water vapor.

19.3.3.3 Temperature Control In some electronic noses the sampling pathway before the vapor reaches the sensors is heated. This ensures that the sample temperature is always consistent regardless of ambient temperature.

19.3.4 Sample Storage

A great challenge of working with natural products is that they change over time. By understanding the mechanism of change in natural products, for example spoiling or ripening, sample quality can be maintained over time. An electronic nose can be used to track the quality of natural products, such as grain, over time. Grain quality parameters including rancidity and the presence of microorganisms have been studied with various electronic noses (see Chapter 21, reference 5). In one study an electronic nose trained on wheat made artificially moldy was used to identify commercial wheat samples (of which 24 where good and 17 bad) with a 92.3 % correct prediction rate [19]. Importantly, no bad samples were misclassified as good. Another example of the effect of storage on natural products is shown in Maul’s study of tomato flavor and aroma [13]. Tomatoes stored at lower temperatures had 19.4 Case Study: The Rapid Detection of Natural Products as a Means of Identifying Plant Species 471 a less flavorful aroma than tomatoes stored at higher temperatures. The electronic nose used for this study was able to classify ripe tomatoes based on storage conditions. Consideration of the variation of the quality of natural products as a result of storage is therefore necessary in developing methods to use an electronic nose for natural-product applications.

19.3.5 Seasonal Variations

Electronic noses have been used to study the quality or ripeness of fruits. Brezmes studied the ripeness of peaches, pears, and apples, using whole fruit, an array of metal-oxide sensors and a NN [11]. Over 92 % of the time, peach and pear ripeness could be determined. Unfortunately, the same results were not seen for apples. DiNatale was able to discriminate the quality of sliced peaches and nectarines based on sensory markers, such as size and color, and QMB array [12]. Maul used an electronic nose with CPs to detect differences in ripe tomatoes stored in different conditions [13]. Though electronic noses potentially can be used to monitor the quality of some fruits over one season, seasonal variations need to be addressed before there is widespread use of the electronic nose in fruit quality monitoring. In another example, the seasonal variations over two crop years of different varieties of rice were found to be greater than the differences in the rice varieties [20]. It was suggested that the electronic nose might be more useful for shelf-life studies of rice than for determining the variety of rice.

19.3.6 Inherent Variability of Natural Products

Natural products vary from season to season, by country of origin, and by species. Even two plants growing next to each other are different. Like humans, each plant and animal and therefore natural product, is unique although the major characteristics are similar. Due to this inherent variability it is critical that a large enough data set be taken to capture as much variability as possible resulting in a more robust model.

19.4 Case Study: The Rapid Detection of Natural Products as a Means of Identifying Plant Species

Natural products have often been used to characterize and differentiate plants. One example is the sorting of wood of different species of trees in the lumber industry by the detection of species-specific marker compounds. Volatile natural compounds have also been used to distinguish closely related species of plants or chemical varieties (chemotypes) of a particular species of plant by GC analysis of their essential oils. We have applied electronic-nose technology in both of these cases. 472 19 Recognition of Natural Products

19.4.1 Wood Chip Sorting

The pulp and paper industries in eastern Canada have a need to differentiating black spruce, balsam fir, and jack pine because their proportions in wood chips affect the quality of the pulp and paper produced. A prerequisite to determining their proportions is to be able to rapidly identify the wood of the three conifers. Several attempts have been made and the few that have succeeded were mainly directed to the sorting of lumber. The methods developed involved recognition of the heartwood of the three species by spectroscopic and/or visual differentiation [37]. These methods failed to distinguish the sapwood of these conifers which makes up the major proportion of the wood chips used by paper mills. Pichette et al. [38] were able to distinguish the three woods using a combination of marker compounds and GC profiles (fingerprints) of the hexane extracts, however the method is too slow to be of any use to paper mills. The rapid sapwood differentiation of these conifers has now been achieved using an electronic nose based on sensor-array technology. In addition, the heartwood of the three trees was also differentiated in the same manner.

19.4.2 Experimental Procedure

Pieces of the sapwood measuring 3 5 cm from seven jack pine, eight balsam fir, and eight black spruce trees were sampled using the CyranoseTM 320. The wood chips were placed in 23 250-mL glass jars, randomly ordered, and kept at room temperature. The samples were sealed with a Teflon-lined lid for storage. The lid was removed for testing and replaced with a two-port Teflon covering. One port was fitted to the snout of the CyranoseTM and used for sampling while the other port was open to the atmosphere. The headspace of each jar was sampled five times in succession using the sampling conditions listed in Table 19.2. A total of 115 smell prints were acquired from the 23 logs. The smell prints were analyzed by PCA and eight smell prints that were identified as outliers with 95 % confidence were removed. Canonical analysis was then applied to the data. The canonical plot (Figure 19.1) shows separation between the different woods. The samples were correctly classified 95 % of the time as shown in Table 19.3.

Tab. 19.2 Cyranose 320 Sampling Conditions for wood chips

Baseline Time 15 s Sample Time 25 s Purge Time 60 s Sample Flow Rate 75 mL min1 Sample Temperature Room temperature Sensor Temperature 41 8C 19.4 Case Study: The Rapid Detection of Natural Products as a Means of Identifying Plant Species 473

Fig. 19.1 Canonical plot projections of the 114 smell prints of wood chips from fir, spruce, and pine

Tab. 19.3 Number of correct identifications for wood chips sampled with a CyranoseTM 320. The value in parentheses is the percentage correct

Identified as Fir Identified as Pine Identified as Spruce

Fir 36 (100) 0 (0) 0 (0) Pine 0 (0) 30 (88) 4 (12) Spruce 0 (0) 1 (3) 36 (97)

Tab. 19.4 GC Temperature setting for study of ground sapwood.

Temperature 8C Time

Injection Temperature 280 Detector Temperature 320 Temperature Program Step 1 60 2 min Step 2 (ramping) 220 5 8Cmin1 Step 3 220 5 min Step 4 (ramping) 320 10 8Cmin1 Step 5 320 40 min

19.4.3 SPME-GC Analysis of the Sapwood of the Conifers Used in Pulp and Paper Industries

An electronic nose essentially analyzes the headspace of a sample, and SPME-GC analysis can indicate whether a difference exists in the headspace between different materials. SPME-GC analyses were carried out on samples of ground sapwood from individual trees of balsam fir, jack pine, and black spruce. One gram of wood was placed in headspace vials and heated for 3 minutes at 70 8C. Then a polyacylate SPME fiber (85lm) was inserted into the sample vial for 5 minutes at 70 8C. The fiber was desorbed for 2 minutes at 280 8C in the injection port of the GC. GC analysis was performed using a non-polar DB-5 capillary column (25 mm 0.25 mm 0.25 lm) using the time settings listed in Table 19.4. 474 19 Recognition of Natural Products

Fig. 19.2 GC profiles of the headspace of balsam fir, jack pine, and black spruce sapwoods obtained by SPME 19.5 Case Study: Differentiation of Essential Oil-Bearing Plants 475

The GC profiles, also referred to as fingerprints, are shown in Fig. 19.2 and represent the average of the individual tree profiles obtained for each of the three species of conifers studied. As can be seen, the differences observed in the three GC profiles correlate to the clusters shown in the PCA plots obtained (Fig. 19.1) using the Cyra- noseTM 320.

19.4.4 Conclusion: Wood Chip Sorting

This procedure, if extended to a chip-by-chip analysis of samples representative of a pile of sawmill wood chips, should lead to a means of determining the proportions of the three conifers present in the mixture.

19.5 Case Study: Differentiation of Essential Oil-Bearing Plants

19.5.1 Golden Rod Essential Oils

The essential oils of three species of Golden Rod, Solidago canadensis, S. rugosa and S. graminifolia, were analyzed by GC using a non-polar and a polar capillary column and by GC-mass spectrometry. As can be seen from the results shown in Table 19.5, the chemical compositions are quite different. The major constituents of S. canadensis are a-pinene (26.9 %) and myrcene (28.3 %). Sabinene (10.1 %), limonene (14.8 %) and b- Phellandrene (18.9 %) are the major components of S. graminifolia whereas a-pinene is by far the most important constituent of S. rugosa at 49.4 %. Other differences are also noticeable in the percentage composition and the presence or absence of certain minor compounds. Approximately two hours of experimental work were required to perform these analyses. The essential oils of these three species of Golden Rod were also analyzed using the CyranoseTM 320 unit. A 0.5–1 mL sample of the essential oil from each of the three species of Golden Rod was placed in an 8-mL glass bottle fitted with a Teflon-faced rubber-lined cap. A small hole in the cap was covered and the oil was allowed to stand for 1 hr at room temperature. The headspace of each of the oils was then sampled five times in a random order. The sampling conditions are shown in Ta- ble 19.6. A total of 15 smell prints were acquired from the three essential oils. Each print required less than 30 s. The 15 smell prints were analyzed by PCA and the plot projections (Fig. 19.3) show a clear distinction of the essential oils of the three species of this plant. 476 19 Recognition of Natural Products

Tab. 19.5 Percentage composition of essential oils of three species of Solidago

Compounds R.I. (DB-5)a) S. canadensis S. graminifolia S. rugosa

a-pinene 941 26.9 1.8 49.4 camphene 954 0.8 0.7 0.4 sabinene 977 0.8 10.1 13.5 b-pinene 978 4.2 5.8 5.8 myrcene 993 28.3 4.7 3.5 a-phellandrene 1002 1.3 3.1 limonene 1033 11.1 14.8 3.1 b-phellandrene 1033 1.2 18.9 14.4 (E)-b-ocimene 1058 0.7 3.9 bornyl acetate 1295 3.5 3.5 1.0 b-elemene 1389 1.1 a-gurjunene 1402 4.8 a›-caryophyllene 1459 0.9 0.4 c-gurjunene 1472 1.1 germacrene D 1488 5.5 3.1 3.4 germacrene A 1512 2.8 d-cadinene 1531 1.2 cubenol 1632 1.2 m/e: 1799 6.4 105,147,161,148,218

a) R.I.: retention indices; DB-5: non-polar capillary column.

Fig. 19.3 PCA plot projections of 15 smell prints of essential oils of Solidago. Squares are S. graminifolia; circles are S. rugosa; triangles are S. Canadensis 19.5 Case Study: Differentiation of Essential Oil-Bearing Plants 477

Tab. 19.6 Cyranose 320 sampling conditions for essential oils

Baseline Time 2 s Sample Time 2 s Purge Time 20 s Sample Flow 120 mL min1 Sample Temperature Room temperature Sensors Temperature 35 8C

19.5.2 Essential Oils of Tansy

Several different chemical varieties of Tansy (Tanacetum vulgare) have been reported [31]. The chemical compositions of the three varieties observed close to Chicoutimi in the Saguenay Region of northern Quebec, Canada, are shown in Table 19.7. The three chemotypes are characterized by the predominance of either b-thujone (75.3 %) or chrysanthenone (54.8 %) or similar amounts of 1,8-cineol (16.9 %), camphor (17.5 %), and borneol (19.3 %). The three varieties of Tansy essential oil were also analyzed by the Cyrano Sciences electronic nose. The sampling conditions and the procedure used were the same as those described above for the Golden Rod essential oils. A total of 15 smell prints were

Tab. 19.7 Percentage composition of essential oils of three chemotypes of Tanacetum vulgare (T.v.)

Compounds R.I. (DB-5)a) T.v. 538b) T.v. 540c) T.v. 541d) a-pinene 941 0.5 4.8 3.8 camphene 954 0.5 1.1 7.0 sabinene 977 2.5 2.4 5.3 b-pinene 978 0.5 4.4 2.3 myrcene 993 0.3 a-phellandrene 1002 1.3 3.1 para-cymene 1028 0.40 0.9 0.9 1,8-cineol 1034 4.1 6.8 16.9 c-terpinene 1068 0.3 0.4 0.6 linalool 1112 0.9 1.1 0.2 a-thujone 1117 0.2 0.8 0.3 b-thujone 1123 75.3 3.1 chrysanthenone 1130 3.3 54.8 5.9 camphor 1146 0.9 1.3 17.5 pinocarvone 1163 0.2 1.2 0.9 borneol 1166 2.3 4.3 19.3 bornyl acetate 1295 0.7 1.0 7.6 germacrene D 1488 2.3 2.7 3.0 a) R.I.: retention indices; DB-5: non-polar capillary column. b) Tanacetum vulgare, b-thujone chemotype. c) Tanacetum vulgare, chrysanthenone chemotype. d) Tanacetum vulgare, camphor, borneol and cineol chemotype. 478 19 Recognition of Natural Products

Fig. 19.4 PCA plot projections of 15 smell prints of essential oils of Tanacetum vulgare. Squares are camphor chemotype; circles are chrysanthenone chemotype; triangles are b-thujone chemotype

acquired from the three essential oils. The 15 smell prints were analyzed by PCA and the plot projections (Fig. 19.4) show a clear distinction of the essential oils of the three varieties of this plant.

19.5.3 Conclusion: Essential Oils

These results show promise for the rapid identification of essential oils from different species of plants, and of oils from different chemical varieties of a specific species of a plant. This is particularly important when one considers that the various applications of essential oils require consistency in their chemical composition. An extension of this method would be a plant-by-plant identification in the field by sampling the headspace volatile compounds using this electronic nose technology.

19.6 Conclusion and Future Outlook

The application of electronic noses to the classification and identification of natural products provides a large potential market. There are opportunities to classify plant species by aroma, identify and sort raw materials, check for consistency among natural oils used in perfumes and as flavors. With careful sample preparation and control, electronic noses can be usefully applied to the recognition of natural products. In order for the full potential of the electronic nose to be realised in this field, we need to develop library-type applications whereby the instrument could be taught the patterns of a species and a database developed that spans the seasons. For this to become reality very 19.6 Conclusion and Future Outlook 479 stable systems, or systems that are readily calibrated, are required. Both of these solutions are being developed. Noses using mass spectometry are more stable but ulimt- ately they are too expensive for widespread implementation and not yet versatile for point-of-need deployment. In addition, software capable of handling hundreds of response patterns needs to be provided.

Acknowledgments The valuable contributions to these studies by Steve Hobbs, Bernard Riedl, Andre Pichette and Helene Gagnon are gratefully appreciated. We also thank Guy Collin for the reproduction of the Tansy essential oil percent composition Table from his publication [31].

References

1 P. Chatonnet. American Journal of Enology S. Campini, A. D’Amico. Sensors and and Viticulture 1999, 50, 479–494. Actuators B 2001, 77, 561–566. 2 P. Namdev, Y. Alroy, V. Singh. Biotechnology 13 F. Maul, S. Sargent, C. Sims, E. Baldwin, Progress 1998, 14, 75–78. M. Balaban, D. Huber. Journal of Food 3 E. Y. Park, K. Y. Han, S. H. Ho, S. S. Kim, Science 2000, 65, 1228–1237. B. S. Noh. 2000 IFT Annual Meeting: 14 B. Costello, R. Ewen, H. Gunson, Prediction of Freshness for Soybean Curd by N. Ratcliffe, P. Spencer-Phillips. Measure- the Electronic Nose, June 10–14, 2000, ment Science and Technology 2000, 11, Dallas, TX. 1685–1691. 4 E. Schaller, J. Bosset, F. Escher. Chemical 15 C. Wessling, T. Nielsen, J. Giacin. Journal Sensors, Biosensors, Bioarrays 1999, 53, of the Science of Food and Agriculture 2001, 81, 98–102. 194–201. 5 E. Schaller, S. Zenhausern, T. Zesiger, 16 B. T. Olsson J, T. Lundstedt, J. Schnurer. J. Bosset, F. Escher. ANALUSIS 2000, 28, International Journal of Food Microbiology 743–749. 2000, 59, 167–178. 6 N. Magan, A. Pavlou, I. Chrysanthakis. 17 G. Keshri, N. Magan. Journal of Applied Sensors and Actuators B 2001, 72, 28–34. Microbiology 2000, 89, 825–833. 7 G. Sberveglieri. The Knowledge Founda- 18 J. R. Stetter, M. W. Findlay, K. M. Schroeder, tion’s Conference on Electronic Nose C. Yue, W. R. Penrose. Analytica Chimica Technologies: Advances in Engineering, Acta 1993, 284, 1–11. Integrating and Commercial Novel Sensor 19 P. Evans, K. Persaud, A. McNeish, R. Sneath, Technologies, October 26–27, 2000, N. Hobson, N. Magan. Sensors and Actuators San Diego, CA B 2000, 69, 348–358. 8 M. Pardo, G. Niederjaufner, G. Benussi, 20 A. Kramer, C. Grimm, E. T. Champagne. E. Comini, et al. Sensors and Actuators B 2000 IFT Annual Meeting. 2000, 69, 397–403. 21 G. Olafsdottir, A. Hognadottir, E. Martins- 9 E. Llobet, E. L. Hines, J. W. Gardner, dottir, H. Jonsdottir. Journal of Agricultural P. N. Bartlett, T. T. Mottram. Sensors and and Food Chemistry 2000, 48, 2353–2359. Actuators B 1999, 61, 183–190. 22 W. Du, T. Huang, J. Kim, M. Marshall, 10 K. Goodner, R. Rouseff. Journal of Agricul- C. Wei. Journal of Agricultural and Food tural and Food Chemistry 2001, 49, 250–253. Chemistry 2001, 49, 527–534. 11 J. Brezmes, E. Llobet, X. Vilanova, G. Saiz, 23 A. Galdikas, A. Mironas, D. Senuliene, X. Correig. Sensors and Actuators B 2000, 69, V. Strazdiene, A. Setkus, D. Zelenin. Sensors 223–229. and Actuators B 2000, 69, 258–265. 12 C. DiNatale, A. Macagnano, E. Martinelli, 24 V. P. Shiers, A. D. Squibb. Leatherhead E. Proietti, R. Paolesse, L. Castellari, Technical Note 1999, 130, 1–18. 480 19 Recognition of Natural Products

25 I. Gonzalez-Martin, J. Perez-Pavon, 32 S. Hobbs, F. Garneau, B. Riedl. “Distin- C. Gonzalez-Perez, J. Hernandez-Mendez, guishing Spruce, Fir & Pine woods for the N. Alvarez-Garcia. Analytica Chimica Acta pulp and paper industry”, Cyrano Sciences 2000, 424, 279–287. Application Note, March 2000. 26 A. Guadarrama, M. L. Rodriguez-Mendez, 33 I. Heberle, A. Liebminger, U. Weimar, J. A. de Saja, J. L. Rios, J. M. Olias. Sensors W. Go¨pel. Sensors and Actuators B 2000, 68, and Actuators B 2000, 69, 276–282. 53–57. 27 M. Muhl, H. Demisch, F. Becker, C. Kohl. 34 Y. Yang, P. Yang, X. Wang. Sensors and European Journal of Lipid Science and Tech- Actuators B 2000, 66, 167–170. nology 2000, 102, 581–585. 35 Guadarrama, J. A. Fernandez, M. Iniguez, 28 N. Shen, S. Duvick, P. White, L. Pollak. J. Souto, J. A. de Saja. Sensors and Actuators B Journal of the American Oil Chemists Society 2001, 77, 401–408. 1999, 76, 1425–1429. 36 O. Koper, T. Zhang. ‘Discrimination of 29 M. Frank, T. Hermle, H. Ulmer, J. Mitrovics. Black Peppers’, http://cyranosciences.com/ Sensors and Actuators B 2000, 65, 88–90. applications/F_PepperIdentification_14.pdf, 30 G. J. Collin, H. Gagnon. Personal Com- 2000. munication. 37 A. H. Lawrence, R. J. Barbour, R. Sutcliffe. 31 G. J. Collin, H. Deslauriers, N. Pageau, Analytical Chemistry 1991, 63, 1217. M. Gagnon. Journal of Essential Oil Research 38 A. Pichette, F.-X. Garneau, F.-I. Jean, B. 1993, 5, 629. Riedl, M. Girard. Journal of Wood Chemistry Technology 1998, 18(4) 427. 481

20 Process Monitoring

Thomas Bachinger and John-Erik Haugen

Abstract Electronic noses have the potential to prepare new ground for non-invasive on-line monitoring of biological processes. In this article we outline their applicability for process monitoring on the basis of selected examples in the areas of food- and biotechnology. Specific case studies on bioprocess monitoring are presented showing that an investigation of the odor of cell cultures can provide the bioprocess operator with valuable information on cell and process state changes. The second application presented outlines the use of electronic noses at-line for monitoring industrial processes in the food and feed industry. For both applications we show that the implementation of electronic noses represents a cost-effective tool for rapid assessment of the chemical and microbial status of raw materials, process streams and end products. Extensive and costly rework or disposal of products that do not fulfill their specifications could be prevented.

20.1 Introduction

The quality of biological products has today become of increasing concern to society. Based on concerns like the potential threat of BSE in food products or the cross transferability of viruses between vertebrates this is especially true for biopharmaceutical products, which is also expressed in the existing vast amount of public regulations. This draws attention to the importance of the monitoring of batch processes to ensure their safe operation and to assure that they produce consistent high-quality products. Most biological processes that can be found in the food and biotechnology industries are probably suited for the application of electronic noses. This is because they involve high concentrations of aromatic compounds or microorganisms producing a wide range of volatiles. However, the demands put on real-time monitoring methods by such processes are high regarding information content, system integration and stability. One reason is that traditional chemical and biological plants are complex non-linear dynamical systems with multiple input and output variables. Often they

Fig. 20.1 Schematics on the integration of an electronic nose into a biological process. The reference gas, which can be the same as the process gas, is humidified before reaching the gas sensors. The sampling interface protects for liquid entry and compensates for flow

variations. VS, Sample gas valve. VR, Reference gas valve

are also composed of numerous sub-processes closely integrated with interconnected mass and energy balances. Figure 20.1 illustrates the simplicity of system integration for electronic noses into such processes. The emission from the process is sampled on-line and analyzed by the gas sensor array in regular time intervals. Besides the ease of system integration several other advantages are obvious: eventual process barriers, e.g. the sterile barrier in a bioprocess, are not violated and system maintenance as well as operator interference are minimized due to a possible high degree of automation. In this review we will focus on two application areas: the on-line monitoring of bioprocesses and the at-line monitoring of food processes. A short introduction to these fields will be followed by a review of previous works. On the basis of recent results we will then outline the capacity of the electronic nose for process monitoring.

20.1.1 On-line Bioprocess Monitoring

In a typical bioprocess cells are grown under sterile conditions in tanks on liquid media that provide, for example, essential nutrients, and vitamins. The products from bioprocesses range from enzymes and single cell protein to biopharmaceuti- cals, which naturally all impose high demands on product quality and safety. Today, most bioprocesses still operate at relatively low yields despite the fact that microbial transformations often reach yields close to the theoretical maximum. One of the reasons is that sensors that acquire real-time information about the 20.2 Previous Work 483 cells’ state and their interaction with the bioreactor environment are rarely available. Consequently, the implementation of sophisticated process control is prevented. Since the experienced operators have long used the odor from bioprocesses for state identification, it could be expected that relevant information can be extracted from the bioprocess off-gas. The application of non-invasive on-line monitoring methods like electronic noses could therefore certainly contribute to improve the quality of bioprocess products.

20.1.2 At-line Food Process Monitoring

A typical production line in the food industry is characterized by several production steps/stages on the way from raw material to final product. In order to keep product quality high throughout the whole production line there may be quality properties that are not measurable on-line and therefore would require at-line sampling and off-line analysis at the production line or in the QC laboratory of the factory. The properties to be investigated off-/at-line will therefore not necessarily coincide with the on-line requirements to quality control analysis. In the case of at-line gas sensor array applications for food process monitoring such properties may be related to the food chemistry of the product and can be measured directly or indirectly by analyzing the vapor phase of the product at different production stages. They may represent product properties related to, for instance, odors, flavor, rancidity, and spoilage. Since such properties are of importance for a variety of processed food products, electronic nose technology should have a wide application range in the food industries.

20.2 Previous Work

A large number of investigations on biological activity monitoring using electronic nose technology can be found in the literature. Examples range from the classification of microbial strains [1, 2] and grains [3] to bacterial contamination of meat [4] and medical applications like the diagnosis of diabetes [5]. However, only a few are directly related to at-line or even on-line process monitoring (see Tab. 20.1).

20.2.1 Quantitative Bioprocess Monitoring

The applicability of electronic noses to bioprocess monitoring has only recently been presented [6–8]. The main focus was initially on using multivariate methods to relate the gas sensor responses to key metabolite concentrations or cell growth. This is because such variables can be expected to be directly associated with the aroma from the cell culture. For instance, the concentration of ethanol and the cell growth could be 484 20 Process Monitoring

Tab. 20.1 Listing of process monitoring applications presented in the literature. BM, bioprocess monitoring. FM, food process monitoring

Application Sensors Algorithms Comment

BM – estimation MOSFET, MOS, IR Forward selection [28] for ANN models for glucose [11, 21], acetate of key metabolites signal parameter selection [11], ethanol [11, 21], acetaldehyde [11], [11, 21] applied [11] glycerol [11] in S. cerevisiae batch processes BM – estimation MOSFET, MOS, IR Forward selection [11, 12]. ANN models established in E. coli batch of cell growth Component correction [27] [10]; CHO perfusion [12]; S. cerevisiae [10–12] for drift compensation [12]. batch [11] processes BM – estimation MOSFET, MOS, IR Forward selection ANN model for rFVIII estimation in of product and component long-term perfusion CHO process concentration [12] correction BM – quality of MOSFET, MOS, IR Forward selection Discrimination of casein hydrolysate bioprocess media [19]. [19], PCA [17, 19], for E. coli growth [17]. Prediction of [17, 19] ANN [19] fermentability of lignocellulose media CP [17, 19] for S. cerevisiae [19] BM – estimation MOSFET, MOS, IR Forward selection, PLS Preculture quality and state estimation of preculture quality for a rec. E. coli strain [20] BM – process and MOSFET, MOS, IR PCA [9, 13, 14] Process state visualization in rec. E. coli cell state determination fed-batch [13], and S. cerevisiae large-scale [9, 13, 14] processes [9]. Cell transition state visualization in perfusion CHO cell process [14] BM – cell physiology MOSFET, MOS, IR Forward selection, Semi-quantitative estimation of physio- prediction [15] PLS logical state variables in E. coli and S. cerevisiae processes BM – observation MOSFET, MOS, IR – Visualization of cell stress caused by of metabolic burden strong overexpression of rec. protein in E. [18] coli BM – detection of MOSFET, MOS, IR – Identification of Micrococcus sp. infection infection [14, 16, 17] [14, 16]. in 500 L CHO process [14]. Identification CP [17] of B. cereus, P. aeruginosa in 2 L CHO process [16]. Shake flask tests with E. coli [17]

FM – aroma quality MOS, Electro-chemical ANN Identification of off-flavor in Serrano of cured ham [22] type dry cured hams FM – quality control MOS Critical level of accumulated Control of drying process of Iberian of drying process in sensor response hams in chambers ham production [23] FM – quality control Ion mobility Critical level of sensor Identification of spoiled sugar beet of sugar beet [24] spectrometry, MOS response FM - sorting of fresh MOS PCA Identification of grape juices with fruit juices [25] off-flavor FM – off-flavors in MOSFET, MOS, IR DPLSR, ANN Identification of feed off-flavor in cow’s milk [26] cow’s milk 20.2 Previous Work 485 estimated with an accuracy of about 10 % in a 200 m3 Saccharomyces cerevisiae fermentation process using artificial neural network (ANN) technology [9]. Improvement of the electronic nose system and the sampling method allowed the estimation of cell growth (biomass) in a 2 L Escherichia coli batch process to as high as 1.46 % accuracy [10]. Such cell growth estimates correlate almost perfectly with the accuracy of standard reference methods (see Section 20.4, study 1 for details). The same improved system was also used to measure cell growth and metabolites like ethanol, glucose, or acetate in a 2.5 L Saccharomyces cerevisiae batch process [11]. Consistently high accuracies between 2.4 and 5 % for the process variables were validated by adding a total of 6 batches at extended batch duration of 35 h to the artificial neural network (ANN) training set. In a different study, the viable cell count was estimated instead of biomass in a five-week production-scale perfusion process. There, it was shown that the viable cell count of Chinese Hamster Ovary (CHO) cells can be estimated accurately at 10 % despite the typical low cell concentration of such processes ( 106 cells mL1) [12]. The successful measurement of glucose in the studies described above is of course not related to a direct measurement of glucose in the process off-gas, since glucose is non-volatile. However, the presented results suggest that it is possible to predict such metabolites because they are correlated with other volatile compounds from the process via stoichiometric or other complex correlations. In the same context, it was possible to measure the product concentration in the above-described CHO cell process [12]. The therapeutic high molecular weight protein ‘human blood coagulation factor VIII’ could be estimated accurately to about the same value as the viable cell count, despite the fact that it is non-volatile.

20.2.2 Qualitative Bioprocess Monitoring

To improve current control strategies in bioprocesses it is often not necessary to measure all key metabolite concentrations accurately. Instead it can be of great advantage if the sensor signal changes in time can reveal simple process state deviations or metabolic changes of the cells. With such on-line information available the operator could react faster to process faults or unfavorable conditions in the bioreactor. This principle was described when following simple process phases in a small-scale Escherichia coli fed-batch fermentation producing recombinant human growth hor- mone [13], as well as in a fed-batch bakers yeast production process on 200 m3 scale [9]. Again, with an improved measurement system cell transition states could be visualized more accurately in a 500 L perfusion mammalian cell (CHO) cultivation process for production of recombinant human blood coagulation factor VIII [14]. It was possible to follow batch, fed-batch and perfusion stages on-line in the process. Also, states of high and low factor VIII productivity as well as lactate formation (high lactate concentrations are inhibiting to the metabolism of mammalian cells) could be visualized. The above principle was further extended by a quasi-quantification of the different metabolic states of cells during a bioprocess. In laboratory-scale Escherichia coli fed- 486 20 Process Monitoring

batch and Saccharomyces cerevisiae batch processes semi-quantitative estimation of the physiological and metabolic states of the cells was realized using simple partial least squares (PLS) models [15]. For a detailed description see Section 20.4, study 2. Another important area of application for electronic noses is the detection of contamination in bioprocess. Suitable on-line methods for identification of foreign growth in bioprocesses rarely exist. Instead, routine checks are made usually once a day by time-consuming incubations of media samples. In a production-scale CHO cell perfusion process it was shown that a bacterial infection with a Micrococcus sp. could be identified at least 1 day before the in-process analysis [14]. Intentional contaminations of laboratory-scale CHO cell perfusion processes with Bacillus cereus and Pseudomonas aeruginosa supported the above findings [16]. In shake-flask cultivations of Micromo- nospora carbonacea the pure culture could be discriminated from contaminated culture [17]. In order to optimize the productivity in recombinant protein fermentations, maximization of the replication and protein expression rate is desired in order to match the biosynthetic capacity of the cell. Such process optimization is much easier to achieve if a sensor technology is available that can identify metabolic burden on-line in the bioprocess. The applicability of an electronic nose to detect metabolic burden was assessed in a series of small-scale fed-batch fermentations using Escherichia coli producing human recombinant superoxide dismutase [18]. The quality of complex growth media is decisive for high growth rate and product yield. Successful discrimination of casein hydrolysates with high quality for growth in recombinant Escherichia coli from lots with low quality has been shown recently [17]. Also the quality of lignocellulose hydrolysates for production of ethanol with Sacchar- omyces cerevisiae has been predicted [19]. The outcome of the anaerobic yeast fermentation could be predicted concerning ethanol productivity by analyzing the hydrolysates before fermentation start. An application with great impact on the performance of final-stage production bioprocesses is the determination of preculture quality. A preculture is the preceding fermentation stage of the production scale fermentation and its quality is therefore of high importance for product quality and yield. The quality and state of inoculum for a 2.5 L recombinant Escherichia coli fed-batch fermentation was assessed successfully in a recent study [20].

20.2.3 At-line Food Process Monitoring

Only a few comprehensive studies exist on at-line food process monitoring applications. One promising application is the work done by Abass et al. who applied an electronic nose for at-line quality monitoring of cured hams [22]. They could demonstrate that the system successfully rejected all the hams that had been assessed as “bad” according to off-flavors by a trained panel. In another at-line food process monitoring application an electronic nose was applied for monitoring and controlling the aroma during the drying process of Iberian hams in chambers [23]. An application that has 20.4 Selected Process Monitoring Examples 487 been implemented recently in the food industry is the use of an ion-mobility based gas sensor system for at-line quality sorting of spoiled sugar beet [24]. Additional at-line food process monitoring examples are listed in Tab. 20.1.

20.3 Special Considerations

Reproducibility and repeatability is an issue in sensor technology due to sensor drift (see Chapter 12). Most chemical sensors do not remain stable over time due to loss in sensitivity and require a frequent recalibration to obtain stable pattern recognition and prediction models. In cases where the sensor drift exceeds the variation in the real measurement data a drift algorithm would be required. Different mathematical approaches have been used recently to handle this problem and they are based on the temporal variation in the sensor signal of repeated identical reference samples that are being measured together with the real samples [27–29]. In bioprocesses it has been shown that the background of the non-inoculated growth media can be used as a stable reference and sensor drift of up to 30 % over 1 year could be corrected for [12]. For the case studies investigated in this paper the drift of the sensors did not represent any major problem due to the fact that the real measurements by far exceeded the magnitude of the sensor drift and drift compensation was therefore not employed. Important considerations for instrument design are to include liquid protection and foam traps when measuring liquid samples on-line over a long period of time. Also heated gas transfer lines should be installed to avoid condensation. To reduce the influence of water and to minimize the difference in response intensity between sample and reference, the reference gas should be humidified (see Fig. 20.1).

20.4 Selected Process Monitoring Examples

20.4.1 On-line Monitoring of Bioprocesses

Conventional bioprocess monitoring still suffers from a lack of suitable on-line monitoring methods that can reveal process states, identify the concentrations of key metabolites or determine cell growth. The complexity of the metabolic network of cells results in a large amount of chemical compounds that could be analyzed in a bioprocess in order to obtain information about the cells’ metabolic or physiological state. However, measurement of such a vast number of analytes requires several different sensor systems to be connected to the bioprocess, many of which are difficult or impossible to operate in on-line mode. The lack of such on-line sensors that could capture comprehensive data about the metabolic state of the cell culture therefore impedes efficient process control. 488 20 Process Monitoring

One of the most important parameters to measure in bioprocesses is certainly the total cell mass. However, today’s existing cell mass monitoring methods can only partly cope with the requirements from modern bioprocesses, e.g., changing high aeration rates in the bioreactor, changing media composition or low cell mass. The first case study presented will outline the potential of the electronic nose for quantitative estimation of biomass in bioprocesses. Refined process control algorithms could be implemented in bioprocesses if it would be possible to on-line measure biomass, substrate, product and inhibitor concentrations. By calculating the uptake/production rates (physiological variables) there- of, the physiological state of the cell culture would be revealed and the culture could theoretically be controlled towards highest possible yield and product quality. The second case study will show the successful semi-quantitative estimation of the physiological state of a cell culture using an electronic nose.

20.4.2 At-line Monitoring of a Feed Raw Material Production Process

The third case study is a feasibility study that focuses on the use of an electronic nose for monitoring the quality of slaughter waste. Waste from slaughterhouse’s represent an important raw material that is being utilized for production of different animal feeds. Due to the possible link between animal cannibalism and BSE, quality control of the waste processing is of great importance in order to obtain products based on pure raw material from the same type of animal. At the waste processing plant the quality of the delivered waste will differ due to different extents of bacterial decay of the slaughter waste, and type of material depending on transport time and season. The off-odor perceived at delivery will be a combination of volatile compounds derived from body effluents (urine and feces), lipid oxidation and bacterial spoilage processes. Accordingly, the major components in the headspace will be volatile acids, aldehydes, ketones, sulfides and amines. With increasing onset of spoilage, the volatile secondary metabolic products (sulfides and amines) will be dominating the off-odor of the waste. Waste from animal slaughterhouses consists of blood and a slurry of particulate (matter) slaughter waste with a high water content, which have been separated before they enter the plant. The different process steps of slaughter waste processing are: (a) the blood is coagulated by water vapor and the dry matter is separated; (b) the water phase is recycled and the dry matter is mixed with the slaughter waste after the grinder; (c) the slurry with particulate matter is delivered by truckloads from different slaughterhouses to the processing plant and fed via a huge funnel into a grinder (particle size of 0.5 cm) representing the first processing step; (d) the waste is dried thereafter, by heating at 100 8C at atmospheric pressure for 20 minutes. The drying process decreases the water content to 42 %; (e) the material then goes into the dry smelter (autoclave) where it is heated to 136 8C with pressure up to 3.2 bar for another 20 minutes. The pressure is decreased and the mass is vaporized until the remaining humidity is 5–8 %. The mass from the dry smelter contains about 40 % fat and 60 % 20.4 Selected Process Monitoring Examples 489 dry matter, which are sent through a press for separation. The final products are a pure lipid phase and bone flour, which are used as raw material for animal feed production.

20.4.3 Monitoring Setup

The gas sensor arrays used in the bioprocess monitoring studies (study 1 and 2) were equipped with a set of 10 metal-oxide semiconducting field effect transistor sensors

(MOSFET), up to 19 metal oxide semiconductor sensors (MOS) and 1 CO2-monitor based on infrared adsorption. The MOSFET sensors were produced in-house at Lin- ko¨ping University (Linko¨ping, Sweden) with different catalytic metal gates of Pd, Pt and Ir at metal film depths between 70 and 400 A. The MOS sensors were commercially available sensors of Taguchi (TGS) or FIS type fabricate. The electronic nose used in case study 3 was a commercial on-line sensor array system (NST 3210, Nordic Sen- sor Technologies AB, Sweden) consisting of 10 MOSFET and 5 MOS sensors (Taguchi type). In the presented case studies the electronic noses had a built-in membrane pump and a mass flow controller to supply the sensor array with a constant flow of gas at all times. Repetitive cycles of, alternately, reference gas and sample gas were measured in order to be able to relate the sensor signal to a stable baseline value, and hence to obtain accurate and reproducible measurements. In case studies 1 and 2 a compensator vessel formed the interface to the bioreactor exhaust gas stream in order to compensate for minor variations in flow rate or gas concentration and to trap condensation (see Fig. 20.1). The reference gas used was the same as the process air to the bioreactor (compressed and filtered air). The humidity of the reference gas was adjusted to approximately the same value as the bioreactor exhaust gas by bubbling the reference gas through distilled and sterile water. In case study 3, dehumidified and active charcoal filtered ambient air was used as references gas.

20.4.4 Signal Processing

The definitions of the signal parameters that have been extracted from the gas sensors are illustrated in Fig. 20.2. The frequency of collecting the sensor signals was set to 1 Hz in all case studies. The total measurement cycle time was 10 and 15 minutes in case studies 1 and 2, respectively. The interval for measuring the sample from the bioreactor exhaust gas was between 20 and 30 seconds. The mean value of the last 20 seconds of the baseline measurement was taken as sensor baseline value for each cycle. The sensor response, on-derivative and on-integral values were all calculated relative to the baseline. The response is the average over the last 6-second sample measurement period. The on-derivative is the value of the fourth measurement point of the sample measurement, and the on-integral is the average of the first 21 seconds of the sample measurement. The off-derivative and off-integral values were calculated 490 20 Process Monitoring

Fig. 20.2 Signal parameter extraction for gas sensors. The sensors are continuously exposed to reference gas and interrupted by short sampling periods (Sample gas/Reference gas). Response and baseline values are calculated as the mean value over a defined time interval. On- and Off-integral values represent times over which the signal is integrated. The derivates on the rise and fall of the signal are the on- and off-derivative values

relative to the response, as the fourth measurement point after the sample measurement and the average over the first 21 seconds after the sample measurement, respectively. Measurement conditions used in case study 3 were as follows: storage experiment – 100 sec baseline, 10 sec sampling (12 ml sample volume) and 40 min recovery. In addition to the three batches of slaughter waste, also the background air was measured repeatedly. Each of the waste batches was measured every 2.8 hour over 5.5 days, i.e. 47 measurements of each sample batch; field experiment – 20 sec baseline, 40 sec sampling (50 ml sample volume) and 40 sec recovery. Between measurements of every new waste batch, both ambient air and water vapor were alternately pumped over the sensors in order to flush the sampling tube and inlet system. Time between each new sample (truck delivery) lasted from 5 minutes up to one hour. Accordingly, the minimum recovery time was about 5 minutes. Average sensor responses (signal height relative to baseline) of the last two replicate measurements were used for the data analysis. 20.4 Selected Process Monitoring Examples 491

20.4.5 Chemometrics

The structure of the ANN used in case study 1 was a one-hidden layer back-propagation network with a sigmoidal activation function and one output node. In the storage experiment of case study 3 the responses of the five MOS sensors were used as inputs to a back-propagation network. A four-hidden layer network with a sigmoidal transfer function and three output nodes was used. Minimization of the network output error was in all cases performed using the Levenberg-Marquardt algorithm. For efficient sensor variable selection a forward selection algorithm was used in case study 2 [30]. The objective of this algorithm is to find a subset of the original sensor signals that minimizes a selection criterion. The selection criterion is the prediction error from a multiple linear regression model towards the desired model output (the process variable). A forward selection adds one variable at a time to the model until the selection criterion reaches a minimum. The PLS models in case study 2 were built using the NIPALS algorithm. All calculations in case studies 1 and 2 were performed using MATLAB (The MathWorks Inc., MA, USA) and PLS-toolbox for MATLAB (Eigenvector Technologies, Manson, WA). The PLS calculation in study 3 was performed using The Unscrambler (v7.5, Camo, Trondheim).

20.4.5.1 Study 1: Estimation of Cell growth in Escherichia coli Fermentations This study was performed using a recombinant Escherichia coli strain producing human carbonic anhydrase. A total of five batch cultivations on a 2 L scale were carried out with a fermentation time of 22 h. Details on this study can be found in Bachinger et al. [10]. Investigation of the raw sensor signals from the gas sensor array reveals several interesting aspects. In Fig. 20.3a, selected sensor signals, biomass and dissolved oxygen level for one of the batch processes are shown. The response pattern from most of the sensors can be directly associated with the three phases of a typical batch process: the lag phase (0–2 h), a phase associated with growth (2–11 h) and the stationary phase (11–22 h). This characteristic sensor pattern can be related either to the cell metabolism or the physical parameters in the broth. For example, some sensors mirrored the dissolved oxygen level in the broth, while the increase in MF8resp after 6 hours occurred at the same time as the depletion of the carbon source in the medium. Since a nutrient rich 2 LB-medium was used in this process metabolic activity did not stop after carbon source depletion, which is reflected in the signal of, for example, MF6resp. The infrared sensor (IR) followed the cell mass evolution proportion- ally with time during the exponential phase of the fermentation. In order to estimate biomass ANN technology was used to relate the gas sensor signal pattern to the cell mass in this process. A trial-and-error procedure was performed to identify the best set of input variables and the structure for the ANN. The input pattern that resulted in the lowest training error for the biomass estimation was identified as MF1(resp), MF2(resp), MF3(resp), MF4(resp), MF7(resp), MF8(resp), MF9(resp), MOS3(resp) and IR(resp). 492 20 Process Monitoring

Fig. 20.3 (a) Selected gas sensor signals and the dissolved oxygen

concentration (pO2) from a 22 h Escherichia coli batch fermentation. (b) ANN validation and off-line values for biomass in a 22 h Escherichia coli batch fermentation 20.4 Selected Process Monitoring Examples 493

A 9-8-1 network with the biomass as network output was trained using the above sensor signal parameters from four Escherichia coli fermentations. The result of the model validation on the fifth fermentation is shown in Fig. 20.3b. In the figure, the off-line biomass values for the fermentation are compared with the estimated biomass values from the ANN model. The mean deviation between off-line and estimated biomass values was 0.043 gL1 and the accuracy reached 1.46 %.

20.4.5.2 Study 2: Physiologically Motivated Monitoring of Escherichia coli Fermentations An on-line approach of monitoring the physiological changes of the cells in a bioprocess is presented in this study. The basic idea was that for the task of a simple physiological state (PS) description it should be sufficient to focus on state identification instead of quantification. We are therefore not specifically interested in the exact values given in the physiological variables instead we would like to trace fast changes in metabolic state. A semi-quantitative method for PS identification is therefore proposed that can be performed without the need for sensor calibration. In this method the trajectory representation of the gas sensors is directly related to the physiological state of the cell culture. Thereby the precise response height or intensity values of the sensors are not critical. Details on this study can be found in Bachinger et al. [15]. The principle is explained on the basis of five 35 h fed-batch fermentations with a recombinant Escherichia coli strain producing b-galactosidase. Figure 20.4a shows selected gas sensor signals, dissolved oxygen level and biomass for one fed-batch process. Similar sensor response characteristics can be observed in this fermentation compared to study 1. Several of the sensors follow the dissolved oxygen concentration and the stages of the fed-batch process can be clearly associated with the sensor response pattern. The strategy developed for PS characterization is relying on PLS methods. For every physiological variable a specific PLS-model is calculated from selected sensor response signals in a standard fermentation. The latent variable with highest correlation towards the desired physiological variable is identified as the models output. The resulting PLS-models serve as the base models for respective physiological variable predictions in subsequent fermentations. More specific, sensor signals from a new fermentation are projected on-line onto a defined PLS-model resulting in new latent variable scores that represent the physiological variable of interest. As first example the physiological variable ‘growth rate’ was modeled accurately by this approach as can be seen in Fig. 20.4b. The sensor signals MF7(on der.), MOS3(off int.), MF7(abs resp.), MOS12(abs resp.), MOS16(abs resp.), MOS5(on int.), IR(resp.), MF5(abs resp.), MOS7(off der.), and MOS19(abs resp.) were selected by a forward selection method for PLS-model building [30]. The PLS-model was built with the data from the first fermentation and a latent variable was selected from the model by visual evaluation to represent the growth rate. Both actual growth rate and the LV score from a new fed-batch process are shown in the plot. The arrows in the figure indicate the coincidence of changes in direction in time in both actual and modeled growth rate. A second model was calculated for the physiological variable ‘glucose uptake rate’. Figure 4d shows LV scores for modeled physiological variables of growth and glucose 494 20 Process Monitoring

Fig. 20.4 (a) Selected gas sensor signals, dissolved oxygen concen-

tration (pO2) and biomass from a 35 h Escherichia coli fed-batch fermentation. (b) Actual growth rate (GR) and latent variable (LV) representing growth rate in the E. coli process. (c) Actual growth rate and glucose update rate values for the E. coli process. (d) Latent variables representing growth rate and glucose update rate in the E. coli process 20.4 Selected Process Monitoring Examples 495

Fig. 20.4c, d uptake rates, plotted next to each other. The selected sensor signals for glucose uptake rate were IR(resp.), MOS18(off der.), MOS12(on int.), MOS5(off der.), MF5(on int.), and MOS2(off der.). It can be seen that the changes in both trajectories occur at the same instances in time as the original physiological variables seen in Fig. 20.4c. 496 20 Process Monitoring

20.4.5.3 Study 3: Quality Control of a Slaughter Waste Process The traditional quality criteria for animal feed products is based on the content of free fatty acids (FFA) in the lipid product. A high FFA value corresponds to a poor product quality. The objective of the study was to investigate whether the electronic nose technology could be used to determine the quality of the waste raw material before it enters the processing plant. Mixing of raw material of different spoilage quality would result in poor quality of the final product. One of the objectives was therefore also to use the technique to sort out waste of similar quality in terms of spoilage status and final FFA value. Two different experiments have been carried out. One was a small-scale experiment for simulating the spoilage processes taking place in the waste during storage or transport on trucks to the waste processing plant. The second was a field experiment, measuring the truckloads of slaughter waste directly at delivery by the processing plant. For the storage experiment batches of fresh slaughter waste from pure pork, pure cattle and a mixture of both were investigated. The off-gas production from the bacterial decay of the waste was monitored continuously. In the second experiment a quality monitoring was performed on the waste directly on the truckloads before they were fed into the processing plant. The waste consisted of pure pork, pure cattle, mixture of cattle and pork and pure poultry.

Storage experiment Three batches of 30 L fresh slaughter waste were stored indoors in open tanks of 80 cm in diameter at 8 8C over 5.5 days. Each of the batches was covered by odorless plastic lids that were connected to tubing under the room ceiling, where the off-odors were passively drained through an outside ventilation system into the open air. This was done to prevent contamination of off-odors between the batches. Off-gases were monitored continuously in the process with the electronic nose. The sampling tubes to the sensor array were located about 30 cm above the waste at the center of each tank. Typical sensor responses are shown for selected sensors in Fig. 20.5 for the pork waste. The other sensors of the array showed a similar distribution over time for the different waste types. There is a period of 62 hours (the bacterial lag phase) before the bacteria enter the exponential growth phase, which is reflected in a simultaneous increase in the gas production. After 5 days the stationary phase was still not reached when the experiment had to be terminated due to a high concentration of sulfides that caused poisoning of some of the sensors so they did not recover back to baseline. A principal component analysis (PCA) plot based on the MOS sensor responses (Fig. 20.6) shows that the different categories of slaughter waste at the given conditions show different off-gas profiles throughout the whole measurement period from the start and far into the exponential phase. At the end of the experiment the pork and cattle waste seem to become similar indicating a production of similar off-gases at this stage of the bacterial decay process. In Fig. 20.7 the PCA plot is based entirely on the MOSFET-sensor responses that show a slightly different distribution. At the start of the experiment the wastes are very similar. After some time, however, the gas sensor profiles become separated and proceed in different directions, which 20.4 Selected Process Monitoring Examples 497 corresponds to the onset of the exponential bacterial growth phase. The onset of the growth phase could also easily be perceived by the change in odor of the samples. During the lag phase the odor had a fresh note and at the onset of the growth phase characteristic strong off-odors were perceived related to increased production of sulfides. Later on, the wastes are clearly separated and only the pork and the mixture wastes are becoming similar at the end of the measurement period, whereas the cattle waste stays significantly different in the off-gas profile. PLS modeling was used to obtain freshness prediction models based on the sensor responses and storage time. The results are listed in Table 20.2. The storage time could be predicted with an error of about 5 hours, which represents an error of 3.7–4.2 %. A back-propagation neural network model has been applied to the sensor responses in order to obtain a prediction model for classification of the different waste types. The responses of the five MOS sensors were used as inputs to the network. A four-hidden layer network with a sigmoidal transfer function and three output nodes was used. The outputs represented the three waste classes (C, P and P þ C). 30 % of the measurement data were used as the training set and the rest was used for validation. The results are shown in Table 20.3. A high classification rate (96–98 %) was obtained for all three waste types. For the pork and cattle waste there was only one measurement that was wrongly classified as belonging to the mixed pork and cattle waste, whereas two measurements of the mixture (P þ C) was undefined in that they could not be fitted to any class.

Field experiment The measurement device was set up in a room next to the feeding funnel. A 7 m stainless tube (1.5 mm inner diameter) that was connected to the sampling inlet of

Fig. 20.5 Selected sensor responses from pork measurements in the storage experiment 498 20 Process Monitoring

Fig. 20.6 PCA plot based on the MOS sensor responses from the storage experiment (C ¼ cattle, P ¼ pork, P þ C ¼ pork and cattle mixture)

the instrument led outdoors and positioned at about 3 m height above ground, was used for sampling the off-gases directly on the truckload. The inlet of the sampling tube was positioned inside, below the cover of the truckload into the headspace above the slaughter waste where measurement took place. The indoor temperature around the measurement device was consequently higher than outside, where the sampling

Fig. 20.7 PCA plot based on the MOSFET sensor responses from the storage experiment (C ¼ cattle, P ¼ pork, P þ C ¼ pork and cattle mixture) 20.4 Selected Process Monitoring Examples 499

Tab. 20.2 Results from PLS regression between the MOS sensor responses and storage time from storage experiment. Number of samples used for the model (n), correlation coefficient (r), and root mean square error of prediction (RMSEP) in hours

Waste nr RMSEP

Cattle 47 0.98 5.6 Pork 47 0.99 5.1 Pork/cattle mix 47 0.99 5.0

Tab. 20.3 Results from ANN classification of different types of slaughter waste. Number of samples is given in brackets

Predicted class Pork þ Cattle Pork Cattle Undefined Total

Pork þ Cattle 96 % (46) 0 % (0) 0 % (0) 4 % (2) 48 Pork 2 % (1) 98 % (47) 0 % (0) 0 % (0) 48 Cattle 2 % (1) 0 % (0) 98 % (47) 0 % (0) 48

took place, hence preventing condensation of gases in the sampling tube. Ambient air was dehumidified, coal filtered, and used as reference air in order to keep a stable sensor baseline. Three replicate measurements were performed on each sample (on truckload delivery). The samples consisted of pure pork, cattle and poultry waste, and mixtures of cattle and pork waste. Due to different transport times the waste was deteriorated to a different extent representing different states of bacterial spoilage. In parallel to the electronic nose analysis at delivery of the waste raw material, the quality of the material was also assessed by giving it a sensory score according to color, odor quality (bad, good), and intensity. The waste material differed in quality from fresh “pleasant” smelling to different extents of spoiled and unpleasant (sulfide, ammonia like) smelling samples due to bacterial spoilage processes. Figure 20.8 shows the output signals of selected sensors for the different waste types measured. It is seen that the response is increasing with increasing FFA values of final product. Increasing sensor signals were also in accordance with the sensory assessment of the raw material. Fresh material having good odor and low odor intensity showed low sensor responses in comparison to samples with unpleasant off-odors and discolor that showed increasing sensor responses. The off-odors and discoloring of the pork and cattle waste was similar to what had previously been observed during the storage experiments. The pork samples showed lower responses at low FFA values compared to the cattle samples. PLS regression between sensor responses and FFA values was used to obtain prediction models for the FFA value for the different types of waste. The results are summarized in Table 20.4. The results show that the quality of the waste raw material is correlated with the quality of final product in terms of the FFA values of the lipid product. 500 20 Process Monitoring

Fig. 20.8 Sensor responses of selected sensors for cattle, pork, cattle and pork mixture and poultry slaughter waste samples with increasing content of free fatty acids (FFA)

20.4.5.4 Discussion Several conclusions can be drawn from the presented bioprocess monitoring studies. Specific compounds (cell metabolites) that are non-volatile or of a very low concentration below the detection limit of the gas sensor array can be monitored indirectly by measuring the vapor phase. The results suggest that it may be possible to predict metabolites of a biological system because they are correlated with volatile compounds via stoichiometric or other complex correlations. In both case studies all the cell transition states could be predicted with a constant and high accuracy including the very small biomass values at the beginning of the cultivation in case study 1 Even though the result was obtained at low biomass concentration the derived neural network model gave an accuracy similar to that for conventional wet chemical techniques. Physio- logical state changes could be tracked in case study 2. It was not necessary to achieve quantitative resolution of the PS, instead fast cell transition states were monitored in a

Tab. 20.4 Results from PLS regression between sensor responses and FFA values from field experiment. Number of samples used for the model (n), correlation coefficient (r), and RMSEP as a percentage of the measurement range

Waste nr RMSEP

Cattle 30 0.95 12.5 Pork 20 0.83 12.5 Cattle/pork mix 20 0.92 6.2 Poultry 16 0.93 5.0 20.5 Future Prospects 501 semi-quantitative approach. The method has the advantage that a representative model can already be built on the basis of a single fermentation. Since quantitative information is not acquired, drift counteraction and calibration problems are not a major com- plication. The possibilities of on-line and non-invasive operation of the measurement make it a simple and fast method for the monitoring of industrial bioprocesses. The results from the storage experiment in case study 3 suggest that the electronic nose has a potential for sorting out slaughter waste based on the extent of spoilage. The bacterial transient responses expressed in the sensor signals were strongly correlated to characteristic odor differences varying between fresh and spoiled odor. In addition, the results indicate the possibility to determine the purity of the slaughter waste in terms of animal content. Results obtained in the field experiment indicate that electronic nose technology can have a potential in quality control of slaughter waste in terms of spoilage status of the waste before it is fed as raw material into the waste processing plant. In addition, it could be demonstrated that the quality in terms of FFA values of the final product can be predicted by analyzing the odor of the raw material before it enters the process. This demonstrates the importance the quality of raw material may have for the quality of final product.

20.5 Future Prospects

Even though many attempts are made to employ electronic noses for quantitative monitoring, direction of application focuses mainly onto the more successful qualitative monitoring approaches. This favors biological process monitoring, since detection of process abnormalities or cell/process states does not rely purely on quantitative information. Further development and success of the electronic nose technology in process monitoring applications would profit greatly from sensors with improved stability, selectivity, less signal drift, and faster update speeds. There is a rapidly advancing research and development going on both on sensors and instrument hardware and software in order to enhance selectivity, sensitivity and reproducibility of the gas sensors. Application-specific sensor selection, improved calibration modeling and adapted pattern recognition analysis will enable us to expand the area of applicability even further. It becomes clear from the presented material that this technology has a potential for process control by monitoring the volatile compounds produced throughout a process that will allow fast/rapid detection of process abnormalities/deviations in order to ensure the final product quality. However, since this technology does not provide specific chemical information due to the limited selectivity of chemical sensors, it mostly provides little insight into the causes when deviations occur. For some applications the monitoring of the vapor phase may therefore not be sufficient to obtain the essential process information and additional sensors would be required. Sensor fusion with other on-line/at-line measured process parameters could, especially in bioprocesses, lead to a better understanding of the signal responses. A fully automated multi-sensor system methodology consisting of different sensor technologies to monitor 502 20 Process Monitoring

the essential process parameters required for assuring the quality of both raw material, process and final product may therefore be the future solution for some applications [31]. Gas-sensors would make up a vital part of such a multi-sensor system. This thought leads to the integration of the electronic nose into knowledge-based systems supporting process control [32]. Process reproducibility and in turn product quality and safety could be improved in the first place and the technology could even be useful in supporting process development. This may be realized in industry in the not so distant future.

Acknowledgments Drs Carl-Fredrik Mandenius, Per Martensson, Tomas Eklo¨v, Helena Lide´n and Martin Holmberg are acknowledged for their valuable contributions to the development of the electronic nose technology for bioprocess monitoring. Process engineer Oliver Tomic is acknowledged for his contribution in the slaughter waste field study.

References

1 T. D. Gibson, O. Prosser, J. N. Hulbert, R. W. 13 C. F. Mandenius, A. Hagman, F. Dunas, H. Marshall, P. Corcoran, P. Lowery, E. A. Sundgren, I. Lundstro¨m.Biosens. Bioel. 1998, Ruck-Keene, S. Heron. Sensors Actuators B 13, 193–199 1997, 44, 413–422 14 T. Bachinger, U. Riese, R. K. Eriksson, C. F. 2 M. Holmberg, E. G. Ho¨rnsten, F. Winquist, Mandenius. J. Biotechnol. 2000, 76, 61–71 I. Lundstro¨m, L. E. Nilsson, F. Gustafsson, L. 15 T. Bachinger, C. F. Mandenius. Eng. in Life Ljung. Biotechnol. Techn. 1998, 12(4), 319– Sciences 2001,1,33–42 324 16 T. Bachinger, U. Riese, R. K. Eriksson, C. F. 3 T. Bo¨rjesson, T. Eklo¨v, A. Jonsson, H. Mandenius. Biosens. Bioel. 2002, 17, 395– Sundgren, J. Schnurer. Cereal Chem. 1996, 403 73, 457–461 17 P. K. Namdev, Y. Alroy, V. Singh. Biotechnol. 4 F. Winquist, E. G. Ho¨rnsten, H. Sundgren, Prog. 1998, 14, 75–78 I. Lundstro¨m. Meas. Sci. Technol. 1993,4, 18 T. Bachinger, C. F. Mandenius, G. Striedner, 1493–1500 F. Clementschitsch, E. Du¨rrschmid, M. 5 W. Ping, T. Yi, X. Haibao, S. Farong. Biosens. Cserjan-Puschmann, O. Doblhoff-Dier, K. Bioel. 1997, 12, 1031–1036 Bayer. Chem. Technol. Biotechnol. 2001, 76, 6 C. F. Mandenius, I. Lundstro¨m, T. Bachin- 885–89 ger. 1st Eur. Symp. Biochem. Eng. Sci. 1996, 19 C. F. Mandenius, H. Lideń, T. Eklo¨v, M. 104 Taherzadeh, G. Lideń. Biotechnol. Prog. 1999, 7 C. F. Mandenius. Adv. Biochem. Eng. Bio- 15, 617–621 technol. 1999, 66, 65–83 20 C. Cimander, T. Bachinger, C. F. Mande- 8 T. Bachinger, C. F. Mandenius. Trends Bio- nius. Biotechnol. Prog. 2002, 18, 380–386 technol. 2000, 18, 494–500 21 T. Bachinger, H. Lideń, P. Martensson, C. F. 9 C. F. Mandenius, T. Eklo¨v, I. Lundstro¨m. Mandenius. Seminars Food Anal. 1998,3, Biotechnol. Bioeng. 1997, 55, 427–438 85–91 10 T. Bachinger, P. Martensson, C. F. Mande- 22 A. K. Abass, L. D. Coper et al.. Electronic nius. J. Biotechnol. 1998, 60, 55–66 Noses & Sensor Array Based Systems, Design 11 H. Lideń, T. Bachinger, L. Gorton, C. F. and Applications. W. J. Hurst. (Ed.) 1999, Mandenius. Analyst 2000, 125, 1123–1128 Pennsylvania, USA, Technomic Publishing 12 T. Bachinger, U. Riese, R. K. Eriksson, C. F. Company Inc Mandenius. Bioproc. Eng. 2000, 23 (6), 637– 23 M. C. Horillo, I. Sayago et al.. ISOEN 2000, 642 Brighton, UK 20.5 Future Prospects 503

24 A. Kaipanen. Electronic Noses in the Food 28 J. E. Haugen, O. Tomic, K. Kvaal. Anal. Industry, A state of the art symposium 1998, Chim. Acta 2000, 407, 23–39 49–52, Stockholm, Sweden 29 T. Artursson, T. Eklo¨v, I. Lundstro¨m, P. 25 P. Mielle, F. Marquis. Sensors Actuators B Martensson, M. Sjo¨stro¨m, M. Holmberg. J. 2001, 3795, 1–7 Chemometrics 2000, 14, 711–723 26 J. E. Haugen, O. Tomic, F. Lundby, K. Kvaal, 30 T. Eklo¨v, P. Martensson, I. Lundstro¨m. Anal. E. Strand, L. Svela, K. Jørgensen. In: Elec- Chim. Acta 1999, 381, 221–232 tronic Noses and Olfaction 2000, ISBN 31 V. Steinmetz, F. Se´vila, V. Bellon-Maurel. J. 0750307641, pp. 265–271 Agric. Engng. Res. 1999, 74, 21–312 27 M. Fryder, M. Holmberg, F. Winquist, I. 32 M. D. Naish, E. A. Croft. Mechatronics 2000, Lundstro¨m, Proc. Transducers ’95 and Euro- 10, 19–51 sensors IX 1995, Stockholm, 683–686 505

21 Food and Beverage Quality Assurance

Corrado Di Natale, Roberto Paolesse, Arnaldo D’Amico

Abstract Among the numerous applications of electronic nose technology, the analysis of foodstuff is one of the most promising, and also the most traveled road towards industrial applications for this technology. Because human senses are strongly involved in an individual’s interaction with foods, the analysis of food provides an excellent field to compare the performances of natural and artificial olfaction systems. Because the electronic nose is non-destructive and directly correlates, in principle, to the way the consumer perceives food products, it is a good candidate for use as an evaluation tool for quality assessment. In this chapter, a review of the applications of the electronic nose (and its liquid counterpart the electronic tongue) to the evaluation of quality in foods and beverages is given. Also included is an example case study: the measure of the quality of fish. The experiment described was performed with an electronic nose developed by the authors, a description of which is also provided in the text.

21.1 Introduction

Food analysis is a complex discipline involving many different basic sciences. A multitude of different principles of instrumental analysis is currently being investigated and used for the analysis of foods and beverages. At the industrial level, the objectives of these measurements are directed towards safety (e.g. the search for contaminants), biochemical composition (to identify the basic constituents), and the effects of food treatment and processing. For each of these concerns, a number of techniques are currently being studied and developed. They span from the classical analytical chemistry to the more advanced diagnostic imaging techniques such as nuclear magnetic resonance (NMR) [1]. In order to optimize the evaluation of quality and to enhance the marketability of the products, there is an increasing interest for non-destructive methods to assist in the complex classification of fresh products.

Besides the classical objectives mentioned above, recently there has been increased emphasis on the certification of quality. In a time of increasing globalization of manufacturing and markets, quality improvement is becoming one of the latest trends in food marketing. For instance, the consuming public wants to recognize those protection classified products that may be identified with certain animal or vegetable species or with particular production methods. In this regard, we are witnessing the extension of the same classification criteria traditionally adopted for wines, to foodstuffs like meats and fruit. Quality is a global character of a food – it is concerned with all aspects of the interaction between food and consumers. Hence, the perfect instruments to determine food quality are the human senses. Actually, trained panels of tasters are used to set the criteria of quality, to assess the quality of food, and to help in the development of new products. Although the science of food assessment by a human panel is well understood and accepted as the ‘gold’ standard of sensory analysis, the actions of panels are affected by imprecision, are scarcely repeatable, and should not be used for routine operations. One of the major difficulties with panels is the comparison of analysis done at different times. For instance, the evaluation of wines performed in two different years may give inconsistent results. Although these limitations are widely known, the importance of panels is growing. As an example, the European Union employs human panels to assign market values for olive oils. For these reasons, it is important to replace evaluation by panels with an accurate instrumental technique that can perform measurements in real-time and generate the same information as a panel, but in a reproducible and stable way. An example of the complexity of the measure of food quality is demonstrated by the case of fruit. Currently, fruit quality is assessed by measuring mechanical properties (texture, firmness, and acoustic properties) [2, 3], external images (visible and infrared (IR)) [4], internal images (NMR) [5], electrical properties (complex impedance) [6], sound-wave propagation [7], reflectance spectroscopy (visible and infrared) [8], and, of course, sensory analysis as the reference method to which the instrumental measurements are correlated [9]. Each of these techniques provides partial information, and only through the integration (fusion) of all of them is it possible to achieve quality. Many measurements are made by creating a headspace above a fruit sample. This headspace is studied with conventional analytical chemistry equipment (such as gas chromatography and mass spectroscopy). Correlation between the quality aspects of foods and beverages and the composition of their headspaces (both in quantitative and qualitative terms) has been found in many different cases. Specific biochemical modeling of the production of volatile compounds is also available in many cases (e.g. for fish [10]). Despite these encouraging findings, the measure of the composition of headspaces has not resulted in any practical industrial instrumentation to measure food quality. Currently, the information from the headspace is mostly exploited by the senses of human panels, who provide their judgments about the quality of products. The development of artificial olfaction machines (electronic noses) that are easy to use, portable, and provide a simplified sampling method, appears extremely appealing 21.2 Literature Survey 507 in this field. It could make possible the practical exploitation of a fundamental source of information to determine food and beverage quality. Recently entering the scene is the liquid counterpart of electronic noses, namely arrays of sensors working in solution: the so-called electronic tongue. Such devices are of extreme interest to researchers who want to characterize beverages and, in some cases, foodstuff. For all these reasons, food and beverage quality is the most practiced application of the electronic noses. Research results reported in the literature are finding their way into industrial practice. In the next section, a survey of food application studies available in literature is given with a discussion of general arguments about using electronic noses in this field. Following that, a selected case study will be presented with a description of an electronic nose developed by the authors.

21.2 Literature Survey

Table 21.1 lists a number of foods and beverages that have been the subject of electronic nose analysis. Many of the papers appearing in this area originated from the desire of electronic nose researchers to test the recognition capabilities of their sensor arrays, so that many of these papers deal with questions of little interest to food industries. This is the case, for instance, in the classification of wines of different varieties. Although researchers have focused on this question, the wine industry would rather study differences occurring among wines of the same variety. For each different foodstuff, attention has been devoted to particular aspects. In the case of meat, the effects of processing and the microbial quality have been investigated [11, 12]. As a specific example, boar taint detection is an important factor in the quality of pork meat [13]. It is interesting to note that the boar taint is due to the presence of androstenone, this is a typical compound present in human male sweat. Electronic nose sensitivity to androstenone helped in the analysis of human skin for medical diagnosis purposes [14]. Because the presence of this compound is related to the sexual status of the animal, the counteraction of the boar taint is achieved by the castration of the animal. The relation between castration and meat quality has been found in other animals; for instance, it has been used in evidence by an electronic nose study of South American camelids meat [15]. In the case of fruit and vegetables, attention has mostly been given to measuring the headspace composition variations due to post-harvest processes [16–20] and their correlation with the presence of defects, such as mealiness in apples [21]. From an industrial point of view there is also a strong requirement for the identification and selection of cultivars. Recently, the problem of the identification of the optimal harvest time has been addressed in the case of apples, with an electronic nose obtaining results comparable with the most widely used destructive methods [22]. Olive oil is another special case where electronic noses are requested to be competitive with sensory analysis panels. The European Community requires that human panels assign each olive oil to a market-value category [23]. The use of an economical instrument able to overcome the scarcity of trained human panels has become an 508 21 Food and Beverage Quality Assurance

Tab. 21.1 List of some applications of electronic noses to food and beverage quality. Some electronic tongue applications are also listed

Food Description Reference

Meat Fermentation of sausages De Meyer et al, 2000 [61] Eklov et al., 1998 [62] Processed chicken meat Pfannahuser, 1999 [11] Packaged beef meat Blixt et al., 1999 [12] Ground meat Di Natale et al., 1997 [33] Ground meat Winquist et al., 1193 [34] Boar taint Annor-Frempong at al., 1998 [13] Alpaca and llama meat quality Neely et al., 2001 [15]. Fruit and vegetables Aroma of pears Oshita et al., 2000 [63] Quality of tomatoes Sinesio et al., 2000 [39] Maul et al., 1998 [16] Bacteria infection in potatoes De Lacy Costello et al., 2000 [27] Quality of straweberries Hirschfelder et al., 1998 [64] Ripeness detection Benady et al, 1995 [17] Peaches: correlation with sensory analysis Di Natale et al., 2000 [40] Apple ripeness Hines et al., 1999 [18] Apple picking time Saeveles et al., 2001 [22] Quality of apples and citruses Di Natale et al., 2000 [21] Banana ripeness Llobet et al., 1999 [19] Blueberries – quality sorting Simon et al., 1996 [20] Vegetable oils Defects and rancidity of olive oil Aparicio et al., 2000 [24] Classification of vegetable oils Martin et al., 1999 [65] Classification of olive oil Stella et al., 2000 [66] Di Natale et al. 2000 [68] Cereals Mite infestation Ridgway et al, 1999 [25] Microbial quality Jonsson et al, 1997 [26] Odor classification Borjesson et al, 1996 [67] Wine Toasting of barrels Chatonnet, 1999 [69] Vintage years Di Natale et al., 1995 [70] Vineyards of production Di Natale et al., 1996 [71] Denomination (electronic tongue) Legin et al. 1997 [72] Correlation with sensory analysis Legin et al., 1999 [37] (electronic tongue) Correlation with chemical analysis Di Natale et al., 2000 [38] (electronic nose and tongue) Vinegar Anklam et al., 1998 [73] Dairy products Cheese ripening Schaller et al., 1999 [74] Off flavors in milk Marsili, 1999 [75] Cheddar cheese aroma Muir et al., 1997 [76] Aroma of UHT milk Di Natale et al., 1998 [35] Milk freshness (electronic nose and tongue) Di Natale et al., 2000 [77] Milk freshness (electronic tongue) Winquist et al., 1999 [36] Coffee Aroma discrimination Gretsch et al., 1998 [78] Discrimination of blends Gardner et al., 1992 [79] Discrimination of blends (electronic tongue) Fukunaga et al., 1996 [80] Discrimination of blends (electronic tongue) Legin et al. 1997 [81] Brewery Aroma detection in brewery Tomlinson et al, 1995 [82] Flavor detection Pearce et al., 1993 [83] 21.2 Literature Survey 509

Tab. 21.1 Continued

Food Description Reference

Fishes Trout freshness Schweizer-Berberich et al., 1994 [29] Freshness of cod Di Natale et al., 2000 [30] Cod-fillet storage time Di Natale et al., 1996 [31] Freshness of capelin Olafsdottir et al., 1997 [32] Spirits Sake (electronic tongue) Arikawa et al., 1996 [84] urgent issue. In this direction, the detection of defects and rancidity (the two main descriptors of human panel scores) by means of an electronic nose represents a positive result [24]. Other interesting applications of great social and industrial relevance are those related to the safety of food. As an example, the infestation of mites in cereals [25], the microbial quality of grain [26], and potatoes [27]. The detection of spoilage processes in fish [28–32], meat [33, 34], and milk [35, 36] are also of great importance most of all for processing industry. Food analysis also offers the possibility to compare the electronic nose evaluation with those of expert panels, namely with the human senses at their best. This particular field is fully detailed elsewhere this book. Here it is interesting to note that in those applications where sensory analysis has a long and established tradition, the descriptor used by panelists are so specialized that poor correlation with electronic nose data is found. Typical examples are found in wine [37, 38]. In other cases, when more simple descriptors are used that are less involved with fine human perceptions but rather linked to general quality, the correlation is found to be much better [39, 40]. This suggests that to pursue the utilization of electronic noses a reformulation of sensory profiles is, in some cases, perhaps necessary. From a methodological point of view, all these applications can be classified into two main categories: static classifications and dynamic classifications. Static classification is related to those applications where the electronic nose is expected to recognize samples of foods and beverages as belonging to definite classes. Dynamic classification considers the capability of electronic noses to monitor the evolution of foods from the fresh product. Often in this case, samples are represented along a ‘freshness lad- der’, going from perfectly fresh up to the state of non-edibility. Many of the applications listed in Table 21.1 belong to two opposite classes. The first is the class of studies done by electronic noses researchers. In these papers, the choice of the application and the sample treatment are often na¨ve.Also, in some cases, great attention is devoted to sensor development and sometimes to data analysis, so the results are of little interest to food scientists. However, there are studies by food scientists using commercially available electronic noses. In these cases, the major focus is devoted to the sample, with insufficient attention being paid to the sensors and data analysis, resulting also in reports with little practical use. In those cases where both the researchers and industrial scientists co-operate, the most promising results are achieved. 510 21 Food and Beverage Quality Assurance

21.3 Methodological Issues in Food Measurement with Electronic Nose

From the point of view of measurement methodology, electronic nose measurements have some peculiar issues to be considered. In particular, it should be clear to the electronic nose user that the sensor’s signal is a combination of the sensor sensitivity and the concentration of volatile compounds. Preliminary knowledge of these two quantities is a fundamental pre-requisite to foresee the meaning of the data obtained from the electronic nose. The sensors used should not be very sensitive to volatiles carrying low information about the sample under measurement. A typical example of this is found in olive oil, which is characterized by a large difference between the compositions of the oil and its headspace. Dominant compounds in the headspace are methanol and ethanol, whose presence in liquid is scarce and of no importance to defining the oil characteristics. On the other hand, those substances responsible for the sensory properties (e.g. hexanal, trans-2-hexanal, and ethylacetate among the others), and which are abundant in the oil due to their high boiling points, are found at low concentrations in the vapor phase [41]. In this situation, sensors with high sensitivity to alcohols (e.g. metal-oxide semiconductors and conducting polymers) may give rise to signals that are poorly correlated with the relevant properties of the samples. On the other hand, the sensor nose should be sufficiently sensitive to be able to capture the variations of relevant compounds in the different classes of the inspected samples. Environmental parameters, such as temperature, may greatly affect, directly or indirectly, the sensor responses. We can call direct disturbances those related to the sensitivity of the sensors to the environmental parameters, whereas indirect disturbances are those concerned with the effects of the environment on the samples under test. This last aspect is associated with the performances of the sampling methodology. Generally, attention is paid to insulating the sensors from the actions of the environment, e.g. with proper temperature conditioning, making the direct disturbances almost negligible. On the other hand, it has to be clear that what the electronic nose really measures is an image of a solid or liquid foodstuff. The image (i.e. the composition of the headspace) may be, in some cases, very different from the sample itself; furthermore, it is strongly dependent on the environmental parameters. The concentration in the headspace of a compound present, for instance, in a liquid phase, is related to the vapor pressure and to the liquid phase concentration of the compound, and is a function of the temperature. This means that more volatile compounds tend to be more abundant in the headspace than their relative abundance in the sample. Furthermore, the headspace changes dynamically with the variation of temperature. It is well known that for each foodstuff an optimal temperature exists at which the richest expression of the aroma is achieved. A classic example of this is found in red wine and spirits such as cognac. The framework outlined above holds for ideal solutions, namely those for which the mixing enthalpy is zero. In this case Raoult’s law applies and changes of temperature produce a scaling of all the headspace concentrations [42]. For non-ideal liquids (e.g. water-ethanol mixtures) significant deviations from Raoult’s law can occur, and tem- 21.4 Selected Case 511 perature variations result in a progressive distortion of the headspace composition. Foods are, almost always, complex and non-ideal mixtures. Therefore, sampling has to be carefully designed, and when sensory analysis is involved as a reference method, the use of the same temperature range is a constraint to achieving significant results. Optimal sampling systems should completely isolate the sample from the outside environment. In practice this condition is not completely fulfilled, and changes of the environmental parameters results in variations of both quantity and quality of the headspace. These give rise to an additional signal source that can sometimes completely hide the resolution of the electronic nose. The straightforward way to counteract the problem is to increase the performance of the sampling system, improving the sample temperature conditioning and using synthetic carriers. On the other hand, portability and economic requirements are in contrast with a sampling system that is too sophisticated. It is worth noting that, except for a few exceptions, previous work did not pay sufficient attention to the difference between the intrinsic sensor drift and the disturbances induced by the experimental set-up [43]. For some food, the interaction with the environment can also induce irreversible modification of the sample itself. A typical example of this effect is found in wine that is gradually oxidized when exposed to air. As a result of the effect, successive measurements of the same sample are not reproducible. A way to avoid this problem in wine consists of introducing two needles in the cork and using nitrogen as a carrier to sample the bottle headspace. The use of nitrogen as the carrier does not modify the chemical state of the wine [37]. With this arrangement, the wine is measured without opening the bottle.

21.4 Selected Case

As an example of applications in food quality analysis, the freshness of fish will be described in some detail. This example is concerned with measurements performed with an electronic nose conceived, designed, and fabricated by the authors at the Uni- versity of Rome ‘Tor Vergata’. These activities, started in 1995, resulted in a full op- erative instrument named LibraNose in 1999. In the following, a detailed description of the instrument is given, followed by the selected case study.

21.4.1 LibraNose

LibraNose is based on an array of thickness shear mode resonators (TSMR) also known in literature as quartz microbalance (QMB) sensors. The chemical sensitivity is given by a molecular film of pyrrolic macrocycles (mostly metalloporphyrins and similar compounds). In the current configuration eight sensors are used [44]. 512 21 Food and Beverage Quality Assurance

The most well known pyrrolic macrocycles are porphyrins and pthalocyanines. The sensing properties of phtalocyanines have been studied extensively in the past [45]. Porphyrins have only rarely been used, however, and even then only their optical properties were exploited to fabricate sensors for working in solutions. In spite of this, porphyrins are among the most important molecules in nature, their functions as complexing ligands or redox catalysts are essential for all organisms [46]. The mimicking of their biological functions in an electronic nose sensor array has been so attractive that these molecules have become the fundamental component of sensor system. Figure 21.1 shows the basic porphyrin molecule. A number of features make porphyrins eligible as a ‘sensing material’ able to detect the volatile organic compounds. Porphyrins are rather stable and their properties can be finely tuned by simple modifications of their basic molecular structure. The coordinated metal, the peripheral substituents, and the structure of the macrocyclic ske- leton influence the coordination and the related sensing properties [47, 48]. The adsorption properties of solid-state porphyrins are characterized by large sensitivities and wide selectivities: both of these features are particularly appealing for electronic nose applications. While the wide selectivity is generally related to weak interactions (such as Van der Waals force and hydrogen bonding), an additional term, due to the coordination of analytes, has be taken into account. Both the interactions are expected to co-operate. The double interaction is expected to give rise to a non-linear adsorption isotherm resulting from specific p-p-interaction between the aromatic systems of porphyrin and an aromatic analyte (such as benzene). This double interaction has been recently introduced to model interactions in analog molecules [49]. This interaction takes place at low concentrations of benzene and is ruled by a Langmuir isotherm. At higher concentration, after the saturation of the specific sites, only the non-specific adsorption occurs and the shape of the isotherm becomes linear (Henry-type behavior). In general, the selectivity frame of metalloporphyrins towards different analytes depends on several factors, such as peripheral substituents, solid-phase packing, deposition techniques and so on. Among them, a key factor is the metal coordinated to

Fig. 21.1 The basic porphyrin molecule. The molecule can be functionalized by adding lateral substituents at the R R 0 positions, and a metal ion at the core of the ring 21.4 Selected Case 513 the porphyrin ligand; coordination of the analyte to the central metal in this case contributes significantly to sensing material-volatile compound interactions. The strength of these interactions can be broadly predicted by the hard-soft acid-base theory (HSAB): hard acid (metals) prefers to interact with hard base (ligands) and vice versa. In our case, for example, Cr, Mo and V porphyrins (containing hard metals) strongly bind hard ligands, such as alcohols or organic acids, while soft metals (Cu and Ni for example), prefer to bind soft ligands, such as sulfur compounds. In order to be exploited as a sensor, the porphyrins need to be deposited as a solid film onto a substrate. Different techniques are available for this purpose and, among them, the following have been used: solvent casting, Langmuir-Blodgett [50, 51], self-

Fig. 21.2 The figure shows a PCA biplot of scores and loadings of an experiment aimed at evaluating the volatile organic compounds (VOCs) discrimination of arrays of porphyrin-based QMB sensors. Scores are indicated by circles and loadings by crosses. Loadings from 1 to 7 are related to a tetraphenylporphyrin functionalized with different metals (in order: cobalt, molybdenum, copper, iron, vanadium, nickel and chromium) while from 8 to 14 the same cobalt-tetraphe- nylporphrin with different functional groups at lateral positions. The figure shows that metal ions are responsible of a different behavior of the sensitivity. It is also worth noting that compounds are separated in four main groups: amine, aromatic, alcohol and acid, and aldehyde and alkane. The separation indicates the way VOC interact with porphyrin film [76] 514 21 Food and Beverage Quality Assurance

Fig. 21.3 The currently available versions of LibraNose

assembled monolayers [52, 53], and electropolymerization. The adsorption of analytes into solid-state porphyrin layers induces the variation of physical quantities. Each of these quantities can be transformed in an electrical signal matching the porphyrin layer with a proper transducer. Porphyrin-based sensors have been demonstrated with different basic transducers such as TSMR [47], surface acoustic wave [54], conductivity [48], work function [55], and optical [56]. TSMR sensors have been chosen for the implementation of a porphyrin-based electronic nose. These sensors consist of a thin slab of crystalline quartz, cut along a certain symmetrical axis (usually the crystallographic AT direction is used) to obtain material able to sustain bulk electroacoustical oscillation at frequencies from 5 to 30 MHz [57]. The quartz property that makes it interesting as a sensor is that the resonance frequency is, in a limited linear range, inversely proportional to the mass gravitating onto the surface of the quartz. This behavior is exploited to turn the quartz into a chemical sensor when some chemically interactive material, able to capture molecules from the environment, is used as a coating.

Tab. 21.2 List of the main features of the LibraNose instrument

Sensors Eight thickness shear mode resonators, fundamental frequency: 20 MHz Sensor chamber Stainless steel, volume: 10 cm3 Internal tubing Stainless steel Pneumatic components Peristaltic pump, flux:0–0.2 sccm Three two-ways electrostatic valves Sample channels 2 input (sample and cleaning) 1 output Electronics Eight ‘Pierce Oscillators’ at large dynamics Motherboard: microcontroller (Motorola HC05) and programmable logics (Xylinx) Surface-mounted components RS232 serial connection to an external computer Software Cþþ builder for MS/Windows 21.4 Selected Case 515

First studies on a porphyrin-coated QMB showed the fundamental properties of these materials, namely the role played in defining the performances of the sensor by the metal, both in metalloporphyrin complexes and the lateral substituents [47]. Results always confirmed a wide selectivity range that can be adjusted by changing the metal and the peripheral substituents. This property of these sensors satisfies the general requirement of sensors to be employed in electronic noses. Figure 21.2 shows, as a PCA score plot, the ability of the LibraNose to distinguish among different volatile compounds. Figure 21.3 shows the LibraNose. The instrument is linked to an external computer that manages measurements, data collection, and analysis. Pneumatic components (pump and valves) are installed on-board to provide the necessary sample delivery to the sensor chamber. Table 21.2 gives the technical specification of the electronic nose.

21.4.2 Case Study: Fish Quality

For fish it is important to determine the freshness degree, defined as the number of storage days at a certain temperature. For this kind of product, issues such as the distinction between fresh and thawed samples and the maintenance of a constant temperature during storage are of great importance. Currently, many methods based on different measurement principles are available to give a measure of fish freshness [58]. The physical properties of the fish such as the rheological characteristics (firmness and texture) and the electrical properties (impedance) can sometimes be correlated with storage days. For instance, the impedance of fish is, for many species such as cod and salmon, a good indicator of the time after catch. Nonetheless, this method is not effective in case of frozen and thawed fish. The composition of the fish headspace is a source of information about the freshness degree of a sample. Spoilage in fish can be detected through the measure of the amount of amines, such as trimethylamine. Some methods, based on analytical chemistry procedures, are currently available to get information about the content of volatile trimethylamine in the headspace. Nevertheless, the formation of amines due to decomposition starts some days after the catch. Chemical investigations using gas chromatographic techniques have shown that there are five sources of odors, which when combined, give rise to the overall odor of fish [59]. Fresh fish odor is a characteristic related to the individual species. Long-chain alcohol and carbonyls, bromophenols, and N-cyclic compounds are the basic contributors. Opposite to the fresh fish odor is the microbial spoilage odor – caused by compounds that are microbially formed during the spoilage processes. These compounds are short-chain alcohol and carbonyls, amines, sulfur compounds, aromatics, N-cyclic compounds, and some acids. The concentration of these volatiles increases with time as the fish spoils; in fact, some of these are often used as indicators of spoilage [59]. Other sources of odors can be environmental (such as petroleum odors), or due to the processing of fish, and from products of lipid oxidation. 516 21 Food and Beverage Quality Assurance

Due to the high number of volatile compounds involved in the process, and to the fact that they also dynamically change, the measure of fish freshness over a long period of storage can be achieved with a multicomponent approach. This is a typical electronic nose application where a number of non-selective and partially cross-correlated sensors are used to get a qualitative analysis of samples. Different electronic noses have been applied in the past to the detection of fish freshness. Interesting results have been obtained with different sensor technologies such as metal-oxide semiconductor gas sensors [28, 29], electrochemical sensors [32] and TSMRs [31]. In the following, the application of the LibraNose to the measurement of freshness, expressed as storage days, of a number of samples of cod fillets is described. It is useful to discuss some properties of metalloporphyrin-based sensors with regard to the fish-freshness application. As stated earlier, some of the selectivity properties of metalloporphyrins can be derived from the HSAB principle. In this context the use of Mn(III) ion, a hard acid, is expected to provide greater sensitivity to oxygen- based ligands, while a metal ion like Co(II) is expected to give higher sensitivity towards amines or sulfur-containing metals. This scheme is simplified because it does not consider the role of the porphyrin ligand, but experiments have shown that it offer a good explanation just for the selectivity towards amines, alcohols, and sulfur [47]. Therefore, metalloporphyrins offer a way to design sensors optimized to catch fish odor at earlier and late stages of storage. As a reference method, trimethylamine (TMA) and total volatile bases nitrogen (TVB-N) have been measured in the same samples. The data discussed here are related to an experiment performed at the Icelandic Fisheries Laboratory in Reykjavik from 15–20 Nov 1999. Three batches of Atlantic Cod were collected for the experiment. Fish was caught with long line, gutted, and iced immediately after catch and brought to the Icelandic Fisheries Laboratories the following day. Fish was kept at 0 8C before being analyzed. Samples were filleted and de-skinned prior to measuring on the storage days: 1, 2, 3, 4, 7, 9, 11, 15, and 17. Eight samples per storage day were measured; a total of 72 fish. The measurements were performed on fillets. For each fish the right side fillet was measured, and the other side reserved for experiments not described here. Fillets were prepared about one hour before the analysis and were held constantly on an ice-bed until measured.

Fig. 21.4 The fish odor sampler. The probe has a diameter of 5 cm. Air refill is provided by a series of small holes immediately over the fish surface, so that the odor concentration in the supplied air is very close to the equilibrium value. Measurement of a salmon is shown on the right 21.4 Selected Case 517

Fish odor measurements were done using a suitably designed fish odor sampler (see Fig. 21.4), which is a metallic capsule with an internal volume of 10 mL, approximately equal to the volume of the sensor chamber. The capsule is endowed with a series of small orifices for air refilling. The sampler works in contact with the fish fillet, and a stable and reproducible (from the point of view of sensor response) headspace is established in five minutes. During the experiment, the bone side of the right fillets was measured for each fish, and each sample measured twice. The variation in resonant-frequency of QMB, considered in steady state, was used as the sensor feature. Filtered ambient air was used as a reference. Ethanol, at its saturated pressure, was measured before and after each measurement session, in order to control the stability of the sensors. The temperature of the fillet surface, monitored during the measurement, varied from 7 8Cto108C, and no correlation of sensor responses with the fillet temperature was observed. TMA and TVB-N, extracted from fish muscles, were measured using a conventional flow injection analysis-gas diffusion method [60]. Electronic nose data were analyzed by partial least square discriminant analysis (PLS-DA). All calculations were carried out in Matlab 5.0. PLS-DA is a supervised classification method where the search for optimal discriminant directions is performed using PLS. Class membership is numerically represented with a so-called one-of-many encoding. Namely, the y-block in PLS contains a number of variables equal to the number of classes, and the membership of a single data point is expressed by putting the corresponding variable to one and all the others to zero. An unknown sample is then assigned to the class whose output is higher than the others. This procedure is standard when quantitative oriented classifiers are used, such as neural networks. PLS-DA provides both a quantitative estimation of class discrimination, and score and loading plots for a visual inspection of data separation, and the contribution of single sensors to the array. The meaning of these plots is different from those obtained by principal component analysis. In this case, the latent variables are determined in a supervised procedure aimed at fitting the declared class membership, so that, even if the score plot of the first two latent variables may show class overlapping, the globality of all the latent variables can achieve a class separation. Nonetheless, these score plots, being linear projection over some basis, are indicative of the distribution of data in the sensor space. An evaluation of the classification properties can be obtained through a training and validation procedure using the one-leave-out validation technique. Figure 21.5 shows the LibraNose data plotted on a basis identified by the first two latent variables. Samples stored up to three days are clearly gathered in close clusters, the fourth day is overlapped with days 11 and 17, while the days 7,9, and 15 are also overlapping. This tendency to overlap the last days of storage with the first days, namely the inability to distinguish fish at two very different stages of storage, will be shown to be consistent in this experiment. Here we have to keep in mind that there were three batches of fish and evidently there was a slight variation in the spoilage rate of the different batches. There may have been slight variations in handling during the first 24 hours after catch resulting in different spoilage rates of the batches. 518 21 Food and Beverage Quality Assurance

Fig. 21.5 Plot of the first two latent variables of the PLS-DA for the LibraNose data. Days 1-3 are separated while days 4-11-17 and 7-9-15 form grouped clusters

The result of sensory analysis (data not shown here) confirms this effect and in fact the spoilage rate of the second batch appears to be slower than the first one. To clarify, days 1, 2, 3 and 4 are from the first batch, days 7, 9 and 11 are from the second batch and finally days 15 and 17 are from the oldest batch. Class identification is shown in Table 21.3 as a confusion matrix. The validation has been performed on the whole data set because the one-leave-out validation technique has been used. Almost 90% of the samples were correctly identified. Nevertheless, errors, although numerically few, are qualitatively not negligible. Indeed, some samples belonging to storage days from 7 to 15 are classified as belonging to the first day. An interpretation of the errors can be obtained by considering the values of TMA and TVB-N. Figure 21.6 shows the measured values of these two important indicators. As reported in the literature, TMA values become considerably different from zero only after 9 days of storage, whereas TVB-N shows a non-linear and a non-monotonic behavior with time. At the beginning of storage, TVB-N increases to reach a maximum after approximately 4 days, then reaches the same levels as the very fresh fish after 7 days, and then increases following the behavior of TMA. Figure 21.7 shows the plot of TVB-N versus TMA, a log-log scale has been chosen in order to avoid the different evolution of the two indicators. The plot shows basically the same distribution exhib- 21.4 Selected Case 519

Tab. 21.3 Confusion matrix, estimated versus true, storage days in fish freshness experiment

123479111517

18 27 1 38 48 71 7 92 5 1 11 8 15 1 7 1 17 8 ited by the electronic nose systems, namely a straight evolution from days 1 to 4 and a folding back from 7 to 11 and then a net separation of the last days. It is worth mentioning that the similarity of the log-log plot with the electronic nose score plot suggests that a logarithmic-like relationship between sensor response and volatile concentration should exists for the sensors considered here.

Fig. 21.6 Measure of TMA and TVB-N on the samples. TMA becomes important after day 11, whereas TVB-N shows a non-monotonic behavior during the first part of the evolution. In both the plots, the inter-class dispersion grows with the number of storage days 520 21 Food and Beverage Quality Assurance

Fig. 21.7 The log-log plot of TVB-N versus TMA reveals a class distribution very similar to that achieved by the electronic noses. This result confirms that the class overlapping (a sort of folding back effect) may be considered as intrinsic to the examined samples

The results of TMA and TVB-N show that the sensors are mostly correlated with these two parameters, and most of all that the evolution of the chemical composition (qualitative and quantitative) does not provide a straightforward indication of the freshness represented as storage days. This may be explained by slightly different spoilage rates of the three batches used, indicating that days of storage may not give the best information about the freshness status of the fish when different batches of fish are considered.

21.5 Conclusions

The quality of foods and beverages is certainly among the most explored area of applications of electronic noses. Nonetheless, the reported studies have been mostly performed at academic institutions. In many cases the results are certainly interesting for the improvement of the field, but only rarely do they constitute a basis for immediate industrial exploitation. The field still requires more basic research. Most of the 21.6 Future Outlook 521 research reports have concentrated on the improvement of sensors, while other important areas, like the reliability of the sampling systems, have been neglected. However, a couple of conclusions can be made. The first is that the results achieved so far are a sound basis for continuing towards reliable and industrially applicable quality measurement systems. To make rapid progress, the co-operation of electronic nose researchers and food scientists is necessary in order to customize a general-purpose technology like the electronic nose to the specific requirements of food and beverage industries. The second more general conclusion is that the electronic nose is not an analytical instrument, because it does not provide separation of volatile organic components. The future is bright. For the first time, the principles of natural olfaction are being exploited to obtain a chemical measurement. A cultural revolution is emerging that has still to permeate the academic and industrial organizations, as well as the mentality of end users.

21.6 Future Outlook

All the participants in the food chain (producers, processors, and consumers) are potential users of electronic nose technology. Each step of the food chain has peculiar needs that an electronic nose approach can satisfy in principle. As an example, at producer level the increment of quality and yield, at processor level the screening of quality of incoming products to optimize the processing and to sort processed food, and finally at consumer level the control of quality and safety both on the market and at home. All these applications require instruments that work on-site. Food-related sites are usually highly contaminated from the point of view of odor. At the current state of the art, sensors are not able to distinguish between background and relevant odor. From this perspective, portable systems without any conditioning of the samples are of limited use in food analysis. For example, measuring the peculiar odor of a fish in a typical storage room among dozens of stacks of fish crates would be difficult. However, there are certainly applications, interesting at industrial level, where existing electronic noses can be specialized, in terms of sampling and data presentations, in order to fulfill user requirement. For this it is necessary to have strong co-operation between electronic nose producers and end users in order to optimize practical solutions. At this level it is important to have a correct and careful analysis of user needs and expectations, and an educational effort towards the users in order to disseminate the intrinsic novelty carried by the artificial olfaction machines. It is also important that developers and users are aware of the intrinsic limit of information that is carried by the volatile part of a food. For instance, it is important to consider that sensory analysis is almost never just confined to olfactory perception. Actually, synergetic action among the senses is required to form a full judgment over a particular food sample. As an example, in fish analysis, a quality index, linearly correlated with the days in ice, is calculated considering visual, tactile, and olfactory perceptions [60]. This suggests that, to fully reproduce the perceptions of humans with 522 21 Food and Beverage Quality Assurance

artificial sensors, the electronic nose has to be compared and integrated with instruments providing information about visual aspects, texture, and firmness. This opens a further novel investigation direction involving researchers from different areas, con- firming that the interdisciplinary nature is the most strong added value for food analysis.

References

1 J. R. J. Pare, J. M. R. Belenger (Eds.). ‘In- 18 E. Hines, E. Llobet, J. W. Gardner. Electronic strumental techniques in food analysis’, Else- Letters, 1999 35, 821–823. vier, Amsterdam (The Netherlands), 1997. 19 E. Llobet, E. L Hines, J. W Gardner, S. 2 F. Paoletti, E. Moneta, F. Sinesio. Food Franco. Measurement Science and Technology, Science and Technology, 1993, 26, 264–270. 1999 6, 538–548. 3 N. Galili, I. Shmulevich, N. Benichou. 20 J. E. Simon, A. Hertzoni, B. Bordelon, G. E. Transactions of ASAE, 1998, 41, 399–407. Miles, D. J. Charles. Journal of Food Science, 4 Y. Tao, C. T. Morrow, P. H. Heinemann, J H. 1996 61, 967–969. Sommer. Transactions of ASAE 1990 90, 21 C. Di Natale, A. Macagnano, E. Martinelli, R. 3531–3554. Paolesse, E. Proietti, A. D’Amico. Sensors 5 S. I. Cho, G. W. Krutz, H. G. Gibson, K. and Actuators B, 2001 78, 26–31. Haghighi. Transactions of ASAE, 1990,3, 22 S. Saevels, C. Di Natale, B. Nicolai. Procee- 1043–1050. dings of the 6th International Symposium on 6 C. Curt. Science des aliments, 1997, 17, 435– Fruit, Nut, and Vegetable Production Engin- 456. eering, Potsdam (Germany) 11–14 Sept. 7 N. Sarkar, R. R. Wolfe. Transactions of ASAE, 2001. 1983, 26, 624–629. 23 European Communities Regulation 2598/ 8 V. Bellon, J. L. Vigneau, M. Leclerq. Applied 91, Off;. J. Eur. Communities Legis. 248 (1991) Spectroscopy, 1992, 47, 1079–1083. 1–33. 9 G. Jellinek. Sensory evaluation of food theory 24 R Aparicio, S. M. Rocha, I. Delgadillo, M. T. and practice, Ellis Horwood Ltd. Publ. Chi- Morales. Journal of Agricultural and Food chester (UK), 1985. Chemistry 2000 48, 853–860. 10 T. Kawai. Critical Reviews in Food Science and 25 C. Ridgway, J. Chambers, E. Portero-Larra- Nutrition, 1996 36, 257–298. gueta, O. Prosser. Journal of the Science of 11 W. Pfannhauser. European Food Research and Food and Agriculture 1999 79, 2067–2074. Technology 1999 208, 336–341. 26 A. Jonsson, F. Winquist, J. Schnuerer, H. 12 Y. Blixt, E. Borch. International Journal of Sundgren, I. Lundstro¨m.International Journal Food Microbiology 1999 46, 123–134. of Food Microbiology 1997 35, 187–193. 13 I. E. Annor-Frempong, G. R. Nute, J. D. 27 B. P. DeLacy Costello, P. J. Ewan, H. E Wood, F. W. Whittington, A. West. Meat Gunsam, W. M. Ratcliffe, P. T. N. Spencer Science 1998 50, 139–151. Philips. Measurement Science Technology 14 C. Di Natale, A. Macagnano, R. Paolesse, E. 2000 11, 1685–1691. Tarizzo, A. Mantini, A. D’Amico. Sensors and 28 R. Olafsson, E. Martinsdottir, G. Olafsdottir, Actuators B, 2000 65, 216–219. T. I. Sigfusson, J. W. Gardner. in Sensors and 15 K. Neely, O. Prosser, P. F. Hamlyn. Meat sensory systems for an electronic nose,J.W. Science, 2000 58, 53–58. Gardner and P. Bartlett (eds.). Kluwer Aca- 16 F. Maul, S. A. Sargent, D. J. Huber, M. O. demic Publishers, Dordrecht (Netherlands), Balaban, D. A. Luzuriaga, E. A. Baldwin. 1992. Proceedings of the Florida State Horticultural 29 M. Schweizer-Berberich, S. Vahinger, W. Society 1997 110, 188–194. Go¨pel. Sensors and Actuators B 1994 18, 282– 17 M. Benady, J. E.Simon, D. J. Charles, G. E. 290. Miles. Transactions of the ASAE, 1995 38, 30 C. Di Natale, G. Olafsdottir, S. Einarsson, E. 251–257. Martinelli, R. Paolesse, A. D’Amico. Sensors and Actuators B 2001 77, 572–578. 21.6 Future Outlook 523

31 C. Di Natale, J. A. J. Brunink, F. Bungaro, F. 47 J. A. J. Brunink, C. Di Natale, F. Bungaro, F. Davide, A. D’Amico, R. Paolesse, T. Boschi, A. M. Davide, A. D’Amico, R. Paolesse, T. M. Faccio, G. Ferri. Measurement Science and Boschi, M. Faccio, G. Ferri. Analytica Chi- Technology 1996 7, 1103–1114. mica Acta 1996 325, 53–60. 32 G. O´ lafsdo´ttir, E. Martinsdo´ttir, E. H. 48 C. Di Natale, A. Macagnano, G. Repole, G. Jo´nsson. Journal of Agriculture and Food Saggio, A. D’Amico, R. Paolesse, T. Boschi. Chemistry 1997 45, 2654–2659. Material Science and Engineering C, 1998 5, 33 C. Di Natale, A. Macagnano, F. Davide, A. 209–214. D’Amico, R. Paolesse, T. Boschi, M. Faccio, 49 K. Bo¨denhofer, A. Hierleman, M. Juza, V. G. Ferri. Sensors and Actuators B 1997 44, Schurig, W. Go¨pel. Analytical Chemistry 521–526. 1997 69, 4017–4031. 34 F. Winquist, E. G. Hornsten, H. Sundgren, 50 G. Roberts. Langmui-Blodgett films , Plenum I. Lundstro¨m. Measurement Science and Press (New York, USA) 1990. Technology 1993 4, 1943–1950. 51 C. Di Natale, R. Paolesse, A. Macagnano, V.I. 35 C. Di Natale, A. Macagnano, A. Mantini, E. Troitsky, T. S. Berzina, A. D’Amico. Analy- Tarizzo, A. D’Amico, R. Paolesse, T. Boschi, tica Chimica Acta, 1999 384, 249–259. F. Sinesio, F. M. Bucarelli, E. Moneta, G. B. 52 C. D Bain, G. M. Whiteside. Angewandte Quaglia. Sensors and Actuators B 1998 50, Chemistry International Edition, English 1989 246–252. 101, 522–525. 36 F. Winquist, C. Krantz-Ru¨lcker, P. Wide, I. 53 C. Di Natale, R. Paolesse, A. Mantini, A. Lundstro¨m. Measurement Science and Tech- Macagnano, T. Boschi, A. D’Amico. Sensors nology, 1998 9, 1937–1946. and Actuators B 1998 48, 369–373. 37 A. Legin, A. Rudnitskaya, Yu. Vlasov, C. di 54 C. Caliendo, P. Verardi, E. Verona, A. D’A- Natale, E. Mazzone, A. D’Amico. Sensors and mico, C. Di Natale, G. Saggio, M. Serafini, R. Actuators B 2000 65, 232–234. Paolesse, S. E. Huq. Smart Materials and 38 C. Di Natale, R. Paolesse, A. Macagnano, A. Structures 1997 6, 689–698. Mantini, A. D’Amico, M. Ubigli, A. Legin, L. 55 C. Di Natale, D. Salimbeni, R. Paolesse, A. Lvova, A. Rudnitskaya, Yu. Vlasov. Sensors Macagnano, A. D’Amico. Sensors and Ac- and Actuators B 2000 69. tuators B 2000 65, 220–226. 39 F. Sinesio, C. Di Natale, G. Quaglia, F. Bu- 56 D. S. Ballantine (ed.). Acoustic Wave Sensors: carelli, E. Moneta, A. Macagnano, R. Pao- Theory, Design, and Physico-Chemical Appli- lesse, A. D’Amico. Journal of the Science of cation, Academic Press (New York, USA) Food and Agriculture, 2000 80, 63–61. 1996. 40 C. Di Natale, A. Macagnano, E. Martinelli, E. 57. G. O´ lafsdo´ttir, E. Martinsdottir, J. Oehlen- Proietti, R. Paolesse, L. Castellari, S. Cam- schla¨ger, P. Dalgaard, B. Jensen, I. Unde- pani, A. D’Amico. Sensors and Actuators B land, I. Mackie, G. Henehan, J. Nielsen, H. 2000 77, 561–566. Nilsen. Trends in Food Science Technology 41 M. T. Morales, A. J. Berry, P. S. McIntyre, R. 1997 8, 258–265. Aparicio. Journal of Chromatography A, 1998 58 D. B. Josephson, R. C. Lindsay and G. 819, 267–275. O´ lafsdo´ttir. in D. F. Kramer, L. Liston. (eds); 42 R.A.Alberty.Physical Chemistry(sixth edition), Seafood quality determination Symposium, J. Wiley and sons (New York, USA) 1983. Nov 10–14, 1986, Elsevier, Amsterdam, 43 P. Mielle, F. Marquis. Sensors and Actuators 1986. B, 1999 58, 526–535. 59 S. Sadok, R. Uglow, S. Haswell. Analytica 44 A. D’Amico, C. Di Natale, A. Macagnano, F. Chimica Acta 1996 334, 279–285. 60 J. B. Davide, A. Mantini, E. Tarizzo, R. Paolesse, Luten, E. Martinsdottir. in Methods to deter- T. Boschi. Biosensors and bioelectronics 1998 mine the freshness of fish in research and in- 13, 711–721. dustries, Institut International du Froide, 45 C. C. Lezenoff, A. B. P. Lever (eds.). Phta- Paris 1997. locyanines: Properties and Applications, VCH 61 D. Demeyer, M. Raemaekers, A. Rizzo, Publ. (Weinheim, Germany);, 1989. A. Holck, A. Smedt, B. de Brink, B. ten 46 D. Dolphine (ed.). The Porphyrins, Vol. VI Hagen, C. Montel, E. Zanardi, E. Murbrekk, part A and Vol. VII part B, Academic Press F. Leroy, F. Vandendriessche, K. Lorentsen, (New York, USA) 1978. K. Venema, L. Sunesen, L. Stahnke, L. Vuyst, 524 21 Food and Beverage Quality Assurance

R. de Talon, R. Chizzolini, S. Eerola. Food 72 A. Legin, A. Rudnitskaya, Y. Vlasov, C. Di Research International 2000 33, 171–180. Natale, E. Mazzone, A. D’Amico. Electroa- 62 T. Eklov, G. Johansson, F. Winquist, I. nalysis 1999 11, 1–7. Lundstrom. Journal of the Science of Food and 73 E. Anklam, M. Lipp, B. Radovic, E. Chiavaro, Agriculture 1998 76, 525–532. G. Palla. Food Chemistry 1998 61, 243–248. 63 S. Oshita, K. Shima, T. Haruta, Y. Seo, Y. 74 E. Schaller, J. O. Bosset, F. Escher. Chimia Kawagoe, S. Nakayama, H. Takahara. Com- 1999 53, 98–102. puters and Electronics in Agriculture 2000 26, 75 R. T. Marsili. Journal of Agricultural and Food 209–216. Chemistry 1999 47, 648–654. 64 M. Hirschfelder, D. Ulrich, E. Hoberg, D. 76 D. D. Muir, E. A. Hunter, J. M. Banks. Hanrieder. Gartenbauwissenschaft (in eng- Milchwissenschaft (in English) 1997 52, 85– lish) 1998 63, 185–190. 88. 65 Y. G. Martin, J. L. Perez-Pavon, B.M. Cor- 77 C. Di Natale, R. Paolesse, A. Macagnano, A. dero, C. G. Pinto. Analytica Chimica Acta, Mantini, A. D’Amico, A. Legin, L. Lvova, A. 1999 384, 83–94. Rudnitskaya, Y. Vlasov. Sensors and Actuators 66 R. Stella, J. Barisci, G. Serra, G. G. Wallace, B 2000 64, 15–21. D. De Rossi. Sensors and Actuators B, 2000 78 C. Gretsch, A. Toury, R. Estebaranz, R. 63, 1–9. Liardon. Seminars in Food Analysis 1998 3, 67 T. Borjesson, T. Eklov, A. Jonsson, H. 37–42. Sundgren, J. Schnurer. Cereal Chemistry 79 J. W. Gardner, H. V. Shurmer, T. T. Tan. 1996 73, 457–461. Sensors and Actuators B, 1992 6, 71–75. 68 C. Di Natale, A. Macagnano, S. Nardis, R. 80 T. Fukunaga, K. Toko, S. Mori, Y. Naka- Paolesse, C. Falconi, E. Proietti, P. Siciliano, bayashi, M. Kanda. Sensors and Materials, R. Rella, A. Taurino, A. D’Amico. Sensors 1996 8, 47–56. and Actuators B 2001 78, 303–309. 81 A. Legin, A. Rudnitskaya, Y. Vlasov, C. Di 69 P. Chatonnet, D. Dubordieu. Journal of Natale, F. Davide, A. D’Amico. Sensors and Agriculture and Food Chemistry 1999 47, Actuators B 1997 44, 291–296. 4319–4322. 82 J. B. Tomlinson. Ferment 1996 9, 85–89. 70 C. Di Natale, F. Davide, A. D’Amico, G. 83 T.C. Pearce, J.W. Gardner, S. Friel, P.N. Sberveglieri, P. Nelli. Sensors and Actuators B Bartlett, N. Blair. Analyst, 1993 118, 371– 1995 25, 801–804. 377. 71 C. Di Natale, F. Davide, A. D’Amico, P. Nelli, 84 Y. Arikawa, K. Toko, H. Ikezaki, Y. Shinha, S. Groppelli, G. Sberveglieri. Sensors and T. Ito, I. Oguri. Journal of Fermentation Actuators B 1996 33, 83–88. Bioengineering 1996 82, 371–376. 525

22 Automotive and Aerospace Applications

M. A. Ryan, H. Zhou

22.1 Introduction

The trainability of an electronic nose, along with the ability to select sensors for response to a suite of compounds, has made this type of device useful in several applications for monitoring air quality in an environment where the possible contaminants are known. In this chapter we will discuss its application to monitoring the presence of hazardous compounds for breathing air in an enclosed space. The application of an electronic nose as an air quality monitor is as an event monitor, where events of low concentration that do not present a hazard are not reported, but events of concentrations approaching a hazardous level are reported so remedial action can be taken. The electronic nose used in these applications is not an analytical device that analyzes the air for all compounds present, but neither is it an alarm that sounds at the presence of any change in the atmosphere. The device described here was used as an air-quality monitor in an experiment aboard NASA’s space shuttle Flight STS-95, and was designed to fill the gap between an alarm with no ability to distinguish between compounds and an analytical instrument.

22.2 Automotive Applications

Use of an electronic nose in the automotive industry is primarily conceptual today, but there are several areas in which such a device can be used. These include monitoring the exhaust for combustion efficiency, monitoring the cabin air for passenger safety, and monitoring the engine compartment for other conditions such as leaking oil or other fluids. Owing to offgassing of fabrics and materials (‘new car smell’), to leaks of coolant from the air-conditioning system, and intake of air from the roadway and the engine compartment, the passenger cabin of an automobile can be significantly more hazardous to human health than the outside air [1, 2]. Improvement of the air quality

in an automobile cabin can be accomplished rather simply, but as cabins will remain well sealed for climate control and energy conservation, a need to monitor the interior will remain. As environmental concerns spur development of more efficient combustion, it will be useful to monitor the exhaust for combustion products as well. Several automobile manufacturers have discussed the possibility of using an electronic nose in a system in which the exhaust is monitored for the presence of compounds indicative of incomplete combustion, and feedback to the engine will adjust engine settings to improve combustion efficiency.

22.3 Aerospace Applications

Electronic noses have been proposed for many applications in aerospace; some of those applications are realistic within the limits of today’s technology, and some will require more development. In the area of space exploration, electronic noses have been proposed for planetary atmospheric studies on landers. This application varies from addition of an electronic nose to a rover to study the atmosphere as the rover moves, to stationary devices, which will study the variations in atmosphere over days or seasons. In the search for evidence of life on other planets, electronic noses have been proposed as desirable sensors because the sensing media in the array can be selected to make it possible to distinguish between isomers and enantiomers [3], and because the sensor array can be configured to span a broad range of compounds. These applications require development of methods that will allow the electronic nose to deconvolute target vapors from an unknown background; work to develop devices with these capabilities is underway at the Jet Propulsion Laboratory (JPL). An immediate, and perhaps the most important, application is monitoring air quality in human habitats. The ability to monitor the recycled breathing air in a closed chamber is important to NASA for use in enclosed environments such as the crew quarters in the space shuttle and the International Space Station (ISS). Today, air quality in the space shuttle is generally determined anecdotally by crewmembers’ reports, and is determined after flight by collecting an end-of-mission sample and analyzing it in an analytical laboratory using gas chromatography-mass spectrometry (GC- MS). The availability of a miniature, low-power instrument capable of identifying contaminants in the breathing environment at part-per-million (ppm) and sub-ppm levels would enhance the capability to monitor the quality of recycled air and thus to protect crew health. Such an instrument is envisioned for use as an incident monitor, to notify the crew of the presence of potentially dangerous substances from spills and leaks, and to provide early warning of heating in electrical components that could lead to a fire. In addition to notification of events, it is necessary to have a reliable method by which judgments on the use of breathing apparatus can be made; if the crew has put on breathing apparatus while repairing a leak or cleaning a spill, it is necessary to know whether it is safe to remove the apparatus. These needs have led to the development of an electronic nose at JPL [4–6], with ultimate application to ISS intended and experiments on the space shuttle in the near-term. 22.3 Aerospace Applications 527

The qualities required for an incident monitor to be used in spacecraft are that it should be capable of identifying and quantifying target compounds at determined levels in a fairly wide range (see Table 22.1), that it be a low mass and volume device which uses low power, and that it require little crew time for maintenance, calibration, and air analysis. There are several possible sensing devices that could be used in the space shuttle or ISS, but all have limitations in terms of their requirements. These devices include GC-MS, volatile organic carbon analyzers, flame ionization detectors, and smoke alarms. Of these, only GC-MS discriminates among compounds; it also has the greatest sensitivity. However, it generally requires crew time in sample preparation, maintenance and calibration. An electronic nose does not, in general, have the sensitivity of GC-MS, however, for most target compounds ppm and sub- ppm sensitivity is required, but not the parts per trillion level found with GC-MS. An electronic nose meets the requirements for an incident monitor. It can identify and quantify compounds in its target set with a dynamic range of about 0.01 to 10 000 ppm, depending on the compound, it lends itself to miniaturization, and because it measures deviation from a background it does not require frequent calibration and maintenance. The electronic nose developed at JPL was designed to detect a suite of compounds and is suitable for use in the crew habitat of a spacecraft. The habitat is an enclosed space where air is recycled and where it is unlikely that unknown and unexpected vapors will be released into the air. It can be assumed that the air is clean at the beginning of a period of enclosure, and it is deviations from that state that the electronic nose will monitor, thus it is not necessary to have detailed knowledge of the constituents of the air initially. In addition, the contaminants which are likely to be present, and for which it is important to monitor, are well known, the number of compounds is not large (50 or so), and the probability of mixtures of 5 or more such compounds appearing at one time is small. It is possible, then, to design and train a device to monitor the air for deviation from a clean baseline and to analyze those deviations for the appearance of a set of target compounds. The air quality conditions in the crew quarters of a spacecraft are not radically different from the conditions in an aircraft cabin, or in the passenger cabin of a bus or automobile. In all those cases, it is reasonable to assume the air is clean at the beginning of a monitoring period, and there is a set of contaminants of concern to be monitored. With such conditions in mind, the JPL electronic nose was designed for a flight experiment where the crew habitat in the space shuttle was monitored continuously for six days. The JPL electronic nose is a low power, miniature device which, in its current experimental design, has the capability to distinguish among, identify and quantify 10 common contaminants which may be present as a spill or leak in the recirculated breathing air of the space shuttle or space station. It has as its basis an array of conductometric chemical sensors made from polymer/carbon composite sensing films developed at Caltech [7, 8]. It is an array of 32 sensors, coated with 16 polymers/carbon composites. The polymers were selected by analyzing polymer responses to the target compounds and selecting those that gave the most distinct fingerprints for the target analytes. The JPL development model was used in a flight experiment on the space 528 22 Automotive and Aerospace Applications

Fig. 22.1 The JPL electronic nose used in the flight experiment on STS-95 is shown as a block diagram and as a photo. The developmental device occupies a volume of 2000 mL and has a mass of 1.4 kg, including the HP 200 LX computer

shuttle flight STS-95 (October-November 1998) to determine whether it could be used as a continuous air quality monitor. A block diagram and photo of the JPL electronic nose are shown in Fig. 22.1. The device used in the flight experiment has a volume of 2000 mL and a mass of 1.4 kg including the HP200 LX computer used for control and data acquisition, and uses 1.5 W average power. The mass and volume were deter- 22.4 Polymer Composite Films 529 mined primarily by the spaceflight-qualified container required for the device to be used in an experiment; the volume and mass can be reduced by a factor of 4 with no modifications to the sensor head or the electronics and minor modifications to the pneumatic system.

22.4 Polymer Composite Films

The polymer/carbon composite films developed at Caltech are the sensing media used in the JPL electronic nose [7–10]. These films are made from insulating polymers loaded with a conductive medium such as carbon to make resistive films. When a polymer film is exposed to a vapor, some of the vapor partitions into the film and causes the film to swell; the degree of swelling is proportional to the change in resistance in the film because the swelling decreases the number of connected pathways of the conducting component of the composite material [7]. The electrical resistance of each sensor is then measured and the response of each sensor in the array is expressed as the change in resistance, dR. Using commercially available organic insulating polymers as the basis for conductometric sensing films allows ready incorporation of broad chemical diversity into the sensing array. The sensors respond differently to different vapors, based on the differences in such properties as polarizability, dipolarity, basicity or acidity, and molecular size of the polymer and the vapor. The polymer/carbon composite sensing films are sensitive to temperature and pressure change as well as to changes in the composition of the atmosphere. In a measuring mode where the device is sniffing the atmosphere and comparing it to a clean background with measurements of each a few minutes apart, temperature changes are generally not significant. However, in the case of continuous monitoring over several hours or days, both temperature and pressure changes will influence the location of the baseline, and it is necessary to distinguish among temperature and/or pressure change, slow buildup of compounds, and baseline drift. All of these issues were addressed in the device developed at JPL. Neither changes in pressure nor humidity which might be found in normal habitat have a significant effect on the differential sensor response, but temperature changes greater than 4–8 8C influence the magnitude of response across the sensing array as well as the fingerprint of individual analytes. While it is possible to measure temperature, pressure, and humidity and to subtract any effect of changes in these conditions from the sensor response data, the JPL electronic nose was built with the capability to control temperature, and pressure and humidity were measured separately. Temperature was controlled on the sensor substrates to stay constant at 28, 32, or 36 8C, both to eliminate apparent baseline drift (film resistance changes) caused by temperature change, and to aid the sensing process. Temperatures around 30 8C will assist the process of desorption of analytes from the films and will prevent hydrogen bonds from forming between analytes and the polymers. 530 22 Automotive and Aerospace Applications

22.5 Electronic Nose Operation in Spacecraft

While it is reasonable to assume clean air at the beginning of an enclosed period in the space shuttle, there are two scenarios in which a clean air baseline must be established. In one scenario, the electronic nose might be used to determine whether it is safe to enter a chamber that has been enclosed for some time without crew use, such as a module in ISS. In the other scenario, a background of clean air must be established to determine whether there has been a slow buildup of a contaminant. This second scenario is among the most likely for contamination of the air. Contaminants may build up slowly as offgassing, slow leaks in vapor and liquid containers, from inadequate air revitalization or filter breakthrough, and as human metabolic products such as methane or carbon dioxide. In both of these scenarios, a system by which a baseline of clean air can be established is necessary. Contamination from offgassing may be considered of minor importance for aircraft or automobile cabins because the air is exchanged frequently in the course of use and fresh air can be brought inside during use, but in cabins where air is not exchanged for several hours, the buildup can be considerable. Often the offgassed molecules are small, such as formaldehyde, and are not well scrubbed in the air revitalization system. In the space shuttle where air might not be exchanged for several days or, more importantly in ISS, where the air is not exchanged, offgassing becomes an important consideration. Flight qualification includes establishment that the offgassing rate of components be below a set level, but there are as yet no data for offgassing over periods of months to years, as will be found on ISS. The JPL electronic nose pneumatic system includes a diaphragm pump, which pulls atmosphere at 0.25 L/min over the sensors and two filters, an activated charcoal filter and a filter of inert material, before the sample chamber. The atmosphere to be analyzed travels through a filter that is selected by a solenoid valve, which switches between the two. During usual monitoring intervals, the air travels through the ‘dummy’ filter made of inert material to provide a pressure drop equivalent to the pressure drop across the charcoal filter. The charcoal filter cleans air without removing humidity, and a baseline of cleaned air can be constructed and used to determine the degree of baseline drift. The constructed baseline allows the analysis program to distinguish between drift and slow change in atmosphere. Figure 22.2 shows how drift and slow buildup can be distinguished after the charcoal filter is switched off; the sensor films respond by rising rapidly and creating a ‘virtual peak,’ and the sensor responses can then be analyzed against the cleaned air background. The analysis of the responses of the sensing array can then be used to determine whether the slow change in the atmosphere is caused by contamination. For the flight experiment, 6 days of continuous operation, the charcoal filter was switched on for 20 minutes out of every 210 minutes. This frequency was sufficient to determine the baseline in this application. If an electronic nose is to be used to determine whether a chamber is safe to enter after a closed period, the cleaned air baseline must be established for several minutes, and the virtual peak analyzed when the charcoal filter is turned off. A schedule for filter changeout must be estab- 22.5 Electronic Nose Operation in Spacecraft 531

Fig. 22.2 a) A virtual peak is created at time 21:08 when the airflow is switched from the charcoal filter, which determines the clean air baseline, to the inert filter that is used during normal measurements. The baseline drift can be determined by fitting the trend of the clean air baseline; in this case the virtual peak can be attributed to baseline drift. b) A virtual peak, which is not attributable to baseline drift, can be analyzed for the presence of hazardous materials

lished; for space shuttle air and no unusual events, changing the filter every 2–3 months is sufficient. If there has been an incident found by the filter, it should be changed after the cause of the incident has been fixed. In other applications, where the pressure and temperature are changing rapidly, or where the composition of the atmosphere changes frequently, the filters can be programmed to switch at different frequencies. In the passenger cabin of an aircraft, for example, filtering can be frequent during the loading and taxi stages, when the concentration of combustion products and of fuel can be high, and less frequent during cruise. The responses of the electronic nose were not influenced significantly by meals or activities in the crew quarters because the device was placed under the air intake vent for the entire cabin; odors were significantly diluted when they reached the sensors. This condition was chosen in order to monitor the average concentration in the cabin rather than localized concentrations. 532 22 Automotive and Aerospace Applications

22.5.1 The JPL Enose Flight Experiment

For the application of adverse event monitoring in the space shuttle, the JPL electronic nose was trained to respond to 12 compounds; 10 of these were compounds likely to leak or spill and the other two were humidity change and vapor from a medical swab (2- propanol and water), which was used daily to confirm that the device was operating. The electronic nose was trained to identify and quantify the 10 contaminant compounds at the 1-hour spacecraft maximum allowable concentration (SMAC) levels that are shown in the upper section of Table 22.1. The 10 contaminants were drawn from a list of compounds of concern and for which air samples are tested after a shuttle flight. In the second-generation device, now under development, there will be 10–12 additional compounds. The sensitivity required for the device was set at the 1 hour SMAC in the flight experiment, and is set at the 24 hour SMAC for the second-generation device. The upper section of Table 22.1 shows the 24-hour SMAC and the lowest level detected reliably by the first generation

Tab. 22.1 Upper Section: Compounds targeted in the first-generation electronic nose, with their 1-hour and 24-hour SMACs, and the lower level detected at JPL with that device. Lower Section: compounds considered for the second-generation electronic nose, with their 24-hour SMACs

Compound SMAC 1 hr (ppm) [**] SMAC 24 hr (ppm) [**] Detected at JPL (ppm)

Methanol 30 10 5 Ethanol 2000 500 50 2-Propanol 400 100 50 Methane 5300 5300 3000 Ammonia 30 20 20 Benzene 10 3 10 Formaldehyde 0.4 0.1 10 Freon 113 50 50 20 Indole 1 0.3 0.03 Toluene 16 16 15

Acetaldehyde 6 Acetone 270 Acetonitrile 4 2-Butanone 150 Chlorobenzene 10 Dichloromethane 35 Furan 0.1 Hexamethyltricyclosilane 25 Hydrazine 0.3 Methyl hydrazine 0.002 Tetrahydrofuran 40 1,1,1-Trichloroethane 11 o,p-Xylenes 100

* Source [11] 22.5 Electronic Nose Operation in Spacecraft 533 electronic nose at JPL, where lower levels were determined by SMACs and are not necessarily detection limts. The lower section of Table 22.1 shows a list of compounds considered for the second set and their 24-hour SMACs; sensor response data on these compounds are not yet available. As an event monitor, it is not necessary to be significantly more sensitive than the 24 hour SMAC level; when the concentration of a contaminant approaches 35 % of the SMAC, measures can be taken to remove the compound from the air and to take action on the source of the contamination. Further training of the software is possible in situ, but for accurate identification and quantification, the training must be done in an environment where it is possible to deliver precise concentrations of the compound in the range of interest. For all cases except formaldehyde, the electronic nose is able to detect the compound at or below the 1 hour SMAC. The sensitivity limit for formaldehyde in the flight experiment device is 10 ppm; by selection of a different polymer set with polymers more likely to sorb formaldehyde, it should be possible to detect that compound below the 24-hour SMAC level. The electronic nose is also able to deconvolute signals to identify and quantify mixtures of two compounds with moderate success (about 60 %). It is expected that with further training and a more selective group of polymers, it will be possible to detect lower concentrations of compounds and to deconvolute mixtures of three or four compounds.

22.5.2 Data Analysis

The data analysis software development portion of the JPL electronic nose flight experiment considered several different approaches. The primary constraint in software development was the requirement that gas events of single or mixed gases from the 10 target compounds be identified correctly and quantified accurately. The co-investigator in the flight experiment, Dr. John James of the Toxicology Branch at NASA-Johnson Space Center (JSC), defined accurate quantification as þ/ 50 % of the known concentration measured in the laboratory. This degree of error was defined based on the SMACs; the toxic level of most of the compounds is not known more accurately than þ/ 50 %, so the SMACs have been set at the lower end. For the flight experiment, constraints in telemetry and communication prevented real-time analysis, and so the development process did not include full capability for immediate resistance vs. time data analysis. A series of software routines was developed using MATLAB (from MathWorks, Inc.) as a programming tool. MATLAB is a flexible program, and thus appealing for development of software, though it runs relatively slowly. For future use, where real-time or quasi-real time analysis is called for, the routines can be translated into C and run on a desktop or laptop computer. For sensing media such as the conducting polymer/carbon films used in this program, relative response changes (in magnitude) have been found to be more reliable than the response shapes, especially at the low gas concentration range targeted in this program (1–100 ppm). Hence, the task of identifying and quantifying a gas event is roughly a two-step procedure: 534 22 Automotive and Aerospace Applications

1) Data pre-processing, to extract the response pattern of a gas event from raw time- series resistance data for subsequent analysis. 2) Pattern recognition, to identify and quantify a gas event based on the response pattern extracted.

22.5.2.1 Data Pre-Processing When presented with continuous monitoring data, a response pattern must be extracted by use of software. This process of extracting a response pattern from raw resistance data involves four sequential steps: 1) noise removal, 2) baseline drift accommodation, 3) gas event occurrence determination, and 4) resistance change calculation.

Noise removal Despite the best effort in choosing sensor films with the consideration of low noise level, the noise level can be quite large. Some polymer films were found to be noisier than others. The reasons one polymer/carbon composite film might be noisier than another are not well understood; noise may be attributed to high sensitivity of the polymer film to small changes in pressure caused by air flow, to differences in the carbon dispersion in the film, or to inhomogeneities in the thickness or even composition of the film itself. In general, the fluctuation in resistance (or noise) is fast compared to the response to a gas event. Therefore digital filtering may be used to filter out this high frequency fluctuation. The length of the filter may be different for different sensors and can be determined by analysis of the noise in each sensor.

Baseline drift accommodation Baseline drift is one of the most difficult problems to be solved in extracting electronic nose resistance data from the time data. The causes for baseline drift can be multiple, and include variations in temperature, humidity, pressure, aging of the sensors, and sensor saturation. However, at present there is no clear understanding of the underlying mechanism of each one of the causes, which makes attempts to compensate drift very difficult. Nevertheless, the baseline drift is generally slowly varying in nature compared to the response time of a detectable gas event. This difference in time scale enables us to use a long-length digital filter to determine the approximate baseline drift and then subtract it from the raw data. The result is further adjusted by piecewise fitting using the baseline information from the clean air reference cycles described above. Although this approach will not accommodate the drift fully, it will reduce the effect to a manageable degree. Figure 22.3 shows resistance data that has been processed. The dark, smooth trace in the upper plot shows the baseline variation determined through the use of low frequency filters. The gray, noisy trace in the lower plot is the data after baseline variation has been subtracted, and the dark line is the processed data, with baseline variation subtracted and after filtering for noise accommodation. 22.5 Electronic Nose Operation in Spacecraft 535

Fig. 22.3 a) Grey, noisy trace: raw resistance as recorded; dark line: baseline drift determined by low frequency digital filtering. b) Grey trace: resistance after baseline drift subtracted; dark line: processed data, resistance after noise accommodation by smoothing and high frequency filtering, and baseline drift corrected

Gas event occurrence determination Because data analysis in the flight experiment of the JPL electronic nose was not real- time owing to constraints unrelated to the technology development, it was not necessary for the analysis to be automatic, but a preliminary software routine for automated determination of whether and when a gas event occurs was developed. It is based primarily on threshold calculation, in which the resistance change over a certain time interval is calculated, and a time-stamp is registered if the change exceeds a pre-set threshold. This routine can detect most gas events; however, it was also found that it might identify noise, and sometimes baseline drift, as gas events. For the flight experiment, events identified by the automated routine were confirmed by visual inspection of the time domain data; future development of the data analysis software will refine the identification method.

Resistance change calculation Since the sensors’ relative responsiveness to a vapor determines the fingerprint of that gas – the response pattern – it is important to preserve this relative responsiveness. 536 22 Automotive and Aerospace Applications

This means any calculation method of the resistance change should be taken at the same time-stamp after the initial onset of a gas. Both the relative resistance change,

R=R0, and the fractional resistance change, (R R0Þ=R0 were tested, and the latter was adopted as it maximizes the difference between the signatures of different gas compounds.

22.5.3 Pattern Recognition Method

Although many pattern analysis methods exist in the general field of electronic nose and other array-based sensor data analysis [12; also see Chapter 6 of Part A, and Part C], no single method appears to be readily applicable to the task of identifying and quantifying single gases as well as mixtures of up to three of the 12 compounds (10 target compounds plus water, humidity change and the propanol wipe) at levels about 1– 100 ppm. Most of the widely used methods have demonstrated their effectiveness, but not to a combination of all three scenarios found here: a large number of target compounds, some of which are of very similar chemical structure (e.g., ethanol and methanol), low target concentrations, and both single gases and mixtures.

22.6 Method Development

For reasons stated above, three parallel approaches to electronic nose data analysis were used during the early stages of software development: discriminant function analysis (DFA), neural networks with back propagation (NNBP), and linear algebra (LA). Principal component analysis (PCA) was initially used, but was later replaced by DFA because DFA tends to do better at discriminating similar signatures that contain subtle, but possibly crucial, gas-discriminatory information. DFA is also better in class labeling than PCA. NNBP, or more specifically, multilayer perceptron (MLP), was selected as an approach because it has good generalization of functions to cases outside the training set, is capable of finding a best-fit function (linear or nonlinear; no models needed), and is also more suitable than DFA when the sensor signatures of two gases are not separable by a hyperplane (e.g., one gas has a signature surrounding the signatures of another gas). However, NNBP is inferior to DFA in classifying data sets that may overlap. The reason to use LA, which is not as commonly used as other methods, is that neither DFA nor NNBP were found to be well suited to recognizing the sensor signatures from combinations of more than one gas. This method tries to solve the equation x=Ac, where vector x is an observation (a response pattern), vector c is the cause for the observation (concentrations of a gas or combinations of gases), and matrix A describes system characteristics (gas signatures obtained from training data, or sensitivity coefficients). For electronic nose data analysis where the response pattern can 22.6 Method Development 537 be noise corrupted, which means there may exist no exact solution, least squares fitting is the preferred way to solve the equation [13, 14]. The idea of developing three parallel methods is that one can first use the LA method to deconvolute an unknown response pattern as a linear combination of target compounds; unknown compounds are expressed as a combination of up to four compounds. If a single compound is found, additional verification can then be carried out by NNBP and DFA methods for increased success rate and accuracy. However we have found the LA method to perform consistently best among the three methods even for single gases, while DFA was consistently the worst, which prompted us to discard the use of the two verification methods of NNBP and DFA in the process. LA is suitable only if the training data are linear, which is not the case for all sensors at the concentration ranges considered (see Table 22.2). For a nonlinear scenario, it is then reasonable to use some nonlinear least squares fitting methods such as that of Levenberg and Marquart (LM-NLS). This is the one of the two new methods that were investigated for nonlinear analysis. The other method, a differential evolution (DE) approach, was also investigated because it promises fast optimization (the LM-NLS method can be rather slow). DE represents some recently emerged so-called genetic algorithms [15]. It is a parallel direct search optimization tool, and begins with an initial randomly chosen population of parameter vectors, adding random vector differentials to the best-so-far solution in order to perturb it. A one-way crossover operation then replaces parameters in the targeted population vector with some (or all) of the parameter values from this ‘noisy’ best-so-far vector. In essence it imitates the principles of genetics and natural evolution by operating on a population of possible solutions using so-called genetic operators, recombination, inversion, mutation and selection. Various paths to the optimum solution are checked and information about them can be exchanged. The concept is simple, the convergence is fast and the required human interface is minimal: no more than three factors need be selected for a specific application. However the last advantage is also its disadvantage: limited control for electronic nose data analysis. Finally, the LM-NLS method was selected as the best tool for electronic nose data analysis.

22.6.1 Levenberg-Marquart Nonlinear Least Squares Method

For nonlinear models the technique of choice for least-squares fitting is the iterative damped least-square method LM-NLS. Similar to LA, LM-NLS tries to find the best-fit parameter vector c from an observation vector x, which is related to c through a known linear or nonlinear function, x ¼ f ðA; cÞ, where A covers system characteristics (sensitivity coefficients) obtained from training data. This method usually begins from a given starting point of c, and calculates the discrepancy of the fit:

residual ¼ðcomputed observedÞ=r; where r is the standard deviation, and updates with a better-fitted parameter c at each step. LM-NLS automatically adjusts the parameter step to assure a reduction in the 538 22 Automotive and Aerospace Applications

residual: increase damping (reduce step) for a highly nonlinear problem; decrease damping (increase step) for a linear problem. Because of this ability to adjust damping, LM-NLS is adaptive to both linear and nonlinear problems. How this method adjusts damping is discussed in detail elsewhere [16]. In the course of this work, it was found that the response of the films to the target compounds is linear with concentration only within a limited range. The nonlinearities in the training data generated are of low order, but successful identification and quantification of gas events must take the nonlinearities into account. To obtain sensor characteristics without further knowledge of sensor nonlinearities, a second-order polynomial fit was used to model the nonlinearities. For each sensor response to

each gas, the program finds the best-fit sensitivity coefficients A1 and A2 (in the least-squares sense) to the following equation:

2 resistance change ¼ A1c þ A2c

where c is gas concentration vector. The fit is constrained to pass through the origin.

A1 and A2 are 13 32 matrices characterizing the sensors’ response to ten targeted gases plus water, humidity change, and the propanol wipe. Several modifications were made to the standard LM-NLS method to suit the electronic nose data analysis problem. First, sets of starting points of vector c were used instead of a single set of starting points of vector c. The purpose of doing this is to avoid a local residual minimum, which is common in many optimization algorithms, including the LM-NLS method. These initial sets of vector c can be randomly assigned from within each element’s allowed range. The total number of initial sets will be determined by the speed desired and the complexity of local minimum problem. In our case, about 200 initial sets were found ( 15 N, where N ¼ 13 is the number of target compounds) to be a good compromise. Second, instead of always updating c for a smaller residual, we modify the update strategy to favor a smaller number of gases within certain ambiguity ranges of the residual. The reason is that signature patterns for a given gas compound generated by the electronic nose sensors have been observed to have large variations. The simple updating strategy tends to minimize the residual with a more than reasonable large number combination of gas when the residual is simply the variation in recorded response pattern itself and should be ignored. The amount of the final residual is an indicator of how large the fitting error is and the confidence level of the fitting. Finally, the sensors’ response pattern was weighted to maximize the difference between similar signatures. As seen in Fig. 22.4, which shows representative signatures of the ten target gas compounds plus the medical wipe at a median concentration level (because of the nonlinearity, there is no single signature for one gas at all concentrations), it is clear that ethanol and methanol have very similar signature patterns. Re- gression analysis also pointed out linear dependency to certain degrees. This means that the signature pattern of one gas could be expressed as a linear combination of the response pattern generated by some other target gases. To reduce this similarity, the sensors’ raw resistance responses must be modified by different weights in the data analysis procedure. 22.6 Method Development 539

Fig. 22.4 Representative signatures of ten targeted gas compounds plus wipe generated by electronic nose sensors. Notice the similarity between ethanol and methanol, and the significant difference between benzene and toluene

22.6.2 Single gases

For lab-controlled gas events, the overall success rate reaches 85 % for targeted singles where success is correct identification and quantification L. Broken down into individual singles, the successes are listed below in Table 22.2. The concentration ranges used in the training sets for each single gas are also given.

Tab. 22.2 Identification and quantification success rates for single gases. The ranges shown here are ranges used in LM-NLS analysis

Compound Concentration Range (ppm) Success Rate (%)

Ammonia 10–50 100 Benzene 20–150 88 Ethanol 10–130 87 Freon 113 50–525 80 Formaldehyde 50–510 100 Indole 0.006–0.06 80 Methane 3000–7000 75 Methanol 10–300 65 Propanol 75–180 80 Toluene 30–60 50 %Relative Humidity 5–65 100 Medical Wipe 500–4000 100 540 22 Automotive and Aerospace Applications

Considering that the raw data are often very noisy at low concentrations, nonlinear at high concentrations, highly correlated in some cases, and weakly additive in some mixtures, these results demonstrate that the LM-NLS method is an effective technique for analysis of an array of sensors. Future work on the electronic nose will attempt to remove many of the impediments to data analysis, with focus on noise and correlation. Correlation will be addressed in polymer film selection. The ability of the data analysis software to identify and quantify single and multiple gas events in clean air was tested in the laboratory. The targeted concentration range for quantification was 30 % to 300 % of the one hour SMAC for each compound. As can be seen from Table 22.2, in some cases it was possible to identify and quantify

Fig. 22.5 Identification and quantification of four single gases using LM-NLS. The shaded area is the target þ/ 50 % detection range

Fig. 22.6 Identification and quantification of three single gases using LM-NLS 22.6 Method Development 541 substantially below the 30 % SMAC concentration; however, in a few cases quantification was successful only as low as 100 % of the 1-hour SMAC. In one case, formaldehyde, we were unable to identify and quantify reliably below several times the 1-hour SMAC. Figures 22.5 and 22.6 show some results of single gas identification and quantification graphically.

22.6.3 Mixed Gases

Deconvolution for identification and quantification of mixtures relies on the additivity of the sensor responses. Here, additivity means that the strength of the response to a mixture of gas 1 at level c1 and gas 2 at level c2 equals the response of the single gas 1 at level 2 plus the response of the single gas 2 at level c2. Identification and quantification of mixtures in clean air was moderately successful. Additive linearity holds for some combinations in concentration ranges near the SMAC level of the lower SMAC compound. The success rate for double gases (about 60 %) was less than that of single gases, as would be expected. An exhaustive set of gas pairs was not run because of time constraints; only a selected group of mixture pairs were run to test the additivity. For this relatively small pool of data, additivity holds for the following gas combinations: methanol þ toluene ammonia þ benzene ethanol þ formaldehyde methanol þ benzene ammonia þ ethanol propanol þ benzene.

Although data obtained on some other combinations of gas compounds, e.g., {benzene þ formaldehyde} and {methanol þ propanol}, did not validate their additivity in these tests, this does not necessary mean the additivity does not hold for those gas combinations. In fact, in many of the gas combination tests, often one of the gases was run at a very low concentration and its response was overwhelmed by the other gas’s strong response. In other words, the detectable concentration of a gas might be higher if there exist other highly responsive gases.

22.6.4 STS-95 Flight Data Analysis Results

The resistance vs. time data that were returned from STS-95 showed that there were several gas events in addition to the daily marker. The daily marker, exposure to a propanol and water medical wipe, was added to the experiment so that operation of the device over the entire period could be confirmed. The initial analysis selected the daily markers and identified them as 2-propanol plus a humidity change. These identifications were confirmed by comparison of crew log times with the time of the event in the data. While the hope in an experiment such as this one is that there will be several events that test the ability of the device, such events would certainly be anomalous events in the space shuttle environment. Software analysis identifies all events 542 22 Automotive and Aerospace Applications

that were not propanol wipe events as humidity changes. Most of those changes can be well correlated with the humidity changes recorded by the independent humidity measurements provided to JPL by JSC. The events are not completely correlated in time because the humidity sensor was located on the stairway between the mid-deck and the flight deck, and the electronic nose was located in the mid-deck locker area near the air revitalization system intake. Those events identified as humidity changes but not correlated with cabin humidity change are likely to be caused by local humidity changes; that is, changes in humidity near the electronic nose that were not sufficient to cause a measurable change in cabin humidity. Figure 22.7 shows the correlation of cabin humidity with electronic nose response in several cases. There are visible dips in the traces at times 19:00, 20:52, and 0:07 CST, November 2–3, 1998. These dips are the changes in air composition, and thus resistance, during the baselining cycle, when air is directed through the charcoal filter. Piecewise baseline fitting is based on the resistance during the baselining cycle. Software analysis of the flight data did not identify any other target compounds as single gases or as mixtures. The independent analysis of collected air samples, in which the samples were analyzed at JSC by GC-MS, confirmed that no target compounds were found in the daily air samples in concentrations above the electronic nose detection threshold. It is not surprising that the only changes the electronic

Fig. 22.7 Sample data from STS-95 electronic nose flight experiment. Circles are the independent humidity measurements in the stairway from mid-deck to flight deck. Polymer sensor responses: (A) poly (2,4,6-tribromostyrene), (B) polyamide resin, (C) poly(ethylene oxide), (D) poly(4-vinylphenol) 22.7 Future Directions 543 nose saw were humidity changes, and it is because events were not expected that the experiment included the relatively uncontrolled daily marker events. There were no compounds that the electronic nose would have indicated as unidentified events present in the air samples.

22.7 Future Directions

22.7.1 Sensors

The number of sensors in the second-generation electronic nose will remain at 32. The number of polymers may be expanded beyond 16 in order to make sub-groups of polymers that have been selected for response to particular classes of compounds within the set of 32 sensors. To determine the set and sub-groups of polymers for the set of some 20 target compounds, a model of polymer-analyte interaction is under development. This model takes account of such parameters of equilibrium constant of solvation of the analyte in the film, analyte diffusion in the film, and the effect of the conductive medium. The model will be used to select polymer suites with maximum separation in patterns for particular analyte suites. This type of selection may result in using some subset of the 32 sensors for various patterns. It is possible that the use of carbon as the conductive medium is responsible for the nonlinearity of responses at low concentrations. Studies of the use of metals such as gold or oxides of transition metals as the conductive medium are underway. It has been found that alcohols and ketones desorb from metals more rapidly than they do from carbon.

22.7.2 Data Acquisition

Current research in data acquisition is investigating the use of frequency dependent methods for data acquisition. AC methods are generally more sensitive than DC methods of measurements; AC methods may allow the use of thinner, higher resistance films, thus increasing film sensitivity. Some sensors exhibit high frequency noise, which may be caused by local heating while resistance is measured, by inhomogeneously distributed carbon, or by variable thickness of the film. Thinner sensors could eliminate some sources of noise, and AC measurements may filter out some of the noise. To test whether high frequency noise can be filtered by AC methods, a single sensing film of polyethylene oxide/carbon was exposed to 2500 ppm methanol and the impedance measured at several frequencies, including DC resistance. As shown in Fig. 22.8, there is substantially less baseline drift when sensor response is plotted 544 22 Automotive and Aerospace Applications

Fig. 22.8 Response of a polymer/carbon film of polyethylene oxide to 2500 ppm of methanol, at three frequencies of impedance measurement and DC resistance measurement

as dI=I0 where I is the impedance, than there is in the same sensor measured at DC, but higher frequency noise is not diminished at the frequencies at which impedance was measured. The decision whether to change over to using AC measurement techniques will consider the efficiency of removing baseline drift through digital filtering in the data analysis process vs. the electronic requirements for AC measurements. It may be sufficient to measure DC resistance and remove the high frequency noise by increasing the number of signal averages from 16 to 32 or 64 and remove the low frequency noise by digital filtering in data processing, as described above.

22.7.3 Data Analysis

Though the data analysis software developed for this electronic nose program was highly successful for its application, several improvements can be made in the future. The overall approach to data analysis will not be modified in the second-generation device. The major change will be the addition of real-time or quasi-real-time analysis. For the flight experiment, data were stored and analyzed after the flight. For ground test experiments in which events are manufactured to challenge the electronic nose, the goal is to have data analyzed within minutes of detection. For faster data analysis, it will be necessary to implement a reliable automated event identification routine and to translate the identification and quantification routines from Matlab into C. There will also be some adjustments to the identification and quantification routines. First, the current data analysis software uses all 32 sensors’ responses as in- 22.8 Conclusion 545 put. Though each sensor’s response was weighted in the analysis in order to maximize the differences between similar signature patterns observed for different gas compounds, it was not done systematically and therefore was not necessarily optimal. In the second generation, the selection of the to-be-used sensor set and their corresponding weights will be optimized by maximizing distances between gas signatures.

The distance between the signatures for gas m and gas n, dmn, is defined as

XN 1 dmn ¼ jdRm;i dRn;ij N i where Rm;i is the ith sensor’s normalized (fractional) resistance change for the m th gas and the summation is over N numbers of sensors used. Second, the core of our data analysis software is the modified LM-NLS method, which is heavy with matrix operations and largely determines the entire data analysis speed. Matrix operation speed is known to be exponentially slower as the matrix size increases. One way to increase speed is to reduce the size of the matrix dynamically in operation by incorporating sensors’ characteristic response information, such as known negative or no responses to certain gas compounds. This characteristic response information can also be used for compounds that cannot be identified by the software; sensors which are known to respond or not to respond to particular functional groups can be sampled for a match. Thus, while it may not be possible to identify unexpected compounds, it will be possible to classify them by functional group. In the first generation electronic nose, data analysis is performed on the steady-state signal produced by changes in the atmosphere. For air quality monitoring, using the steady-state signal is, in general, acceptable, as a transient will not remain in the environment long enough to do harm. However, there are toxins that can be hazardous as transients. With automated event determination, analysis can begin as soon as the resistance measurement passes the preset threshold rather than waiting for steady- state to be reached. In addition, if desorption time is a function of the conductive medium, then it may be possible to use the kinetics of sensor film response for identification and quantification. Several compounds, such as ammonia, can be identified by the shape of the response curve upon visual inspection of the curve. Quantification of the kinetics of response may enable identification of transients.

22.8 Conclusion

The results of the flight experiment were somewhat disappointing to the experimenters, while satisfying to the crew. There were no anomalous events, and the electronic nose was not challenged to identify compounds for which it had been trained. Never- theless, the experiment was successful. The electronic nose detected changes in humidity and the presence of the daily marker, was able to identify and quantify the changes, and was able to use the training set made in the laboratory to do the data 546 22 Automotive and Aerospace Applications

analysis. Further work in development of the JPL electronic nose will involve substantial challenge to the device and to the analysis software, with blind testing, mixtures, and unknowns that can be identified by functional group.

Acknowledgements The research reported in this paper was carried out at the Jet Propulsion Laboratory, California Institute of Technology under a contract with the National Aeronautics and Space Administration, and was supported by NASA Code UL.

References

1 C. C. Chan, H. Ozkaynak, J. D. Spengler, L. tronic Nose’, Proc. National Academy of Sci- Sheldon. ‘Driver Exposure To Volatile Or- ence, 92, 2652, (1995).

ganic-Compounds, CO, Ozone, and NO2 8 M. C. Lonergan, E. J. Severin, B. J. Doleman, Under Different Driving Conditions,’ Envi- R. H. Grubbs, N. S. Lewis. ‘Array-Based ron. Sci. Technol., 25, 964 (1991). Sensing Using Chemically Sensitive, Car- 2 P. L. Leung, R. M. Harrison. ‘Roadside and bon Black-Polymer Resistors’, Chem. Mate- In-vehicle Concentrations of Monoaromatic rials, 8, 2298 (1996). Hydrocarbons,’ Atmospheric Environment, 9 E. J. Severin, B. J. Doleman, N. S. Lewis. ‘An 33, 191 (1999). Investigation of the Linearity and Response 3 M. A. Ryan, N. S. Lewis. ‘Low Power and to Mixtures of Carbon Black-Insulating Or- Lightweight Vapor Sensing Using Arrays of ganic Polymer Composite Vapor Detectors’, Conducting Polymer Composite Chemical- Anal. Chem., 72, 658 (2000). ly-Sensitive Resistors,’ Enantiomer, 6, 159 10 K. J. Albert, N. S. Lewis, C. L. Schauer, G. A. (2001). Sotzing, S. E. Stitzel, T. P. Vaid, D. R. Walt. 4 M. A. Ryan, M. L. Homer, M. G. Buehler, K. ‘Cross-Reactive Chemical Sensor Arrays,’ S. Manatt, F. Zee, J. Graf. ‘Monitoring the Chem. Rev., 2595 (2000). Air Quality in a Closed Chamber Using an 11 Spacecraft Maximum Allowable Concentrati- Electronic Nose,’ Proceedings of the 27th In- ons for Selected Airborne Contaminants, Vols. ternational Conference on Environmental Sy- 1 & 2, National Academy Press, Washington, stems, Society of Automotive Engineers, 97- DC (1994). ES84 (1997). 12 P. N. Bartlett, J. W. Gardner. Electronic Noses: 5 M. A. Ryan, M. L. Homer, M. G. Buehler, K. Principles and Applications, Oxford Univer- S. Manatt, B. Lau, D. Karmon, S. Jackson. sity Press, Oxford (1999). ‘Monitoring space shuttle Air for Selected 13 G. Stang. Linear Algebra and its applications, Contaminants Using an Electronic Nose,’ 2nd edition, Academic press, New York, Proceedings of the 28th International Conference 1980. on Environmental Systems, Society of Auto- 14 C. Lawson, R. Hanson. Solving Least Squares motive Engineers, 981564 (1998). Problems, S.I.A.M. Press, Philadephia, 1995. 6 M. A. Ryan, M. L. Homer, H. Zhou, K. S. 15 R. Storn. ‘On the usage of differential evo- Manatt, V. S. Ryan, S. P. Jackson. ‘Operation lution for function optimization,’ Biennial of an Electronic Nose Aboard the space Conference of the North American Fuzzy In- shuttle and Directions for Research for a formation Processing Society, NAFIPS, IEEE, Second Generation Device,’ Proceedings of the 519 (1996). 30th International Conference on Environmen- 16 M. Lampton. ‘Damping-Undamping Strate- tal Systems, Society of Automotive Engin- gies for the Levenberg-Marquart Nonlinear eers, 00ICES-259 (2000). Least-Squares Method,’ Comput. Phys., 11, 7 M. S. Freund, N. S. Lewis. ‘A Chemically 110 (1997). Diverse Conducting Polymer-Based Elec- 547

23 Detection of Explosives

Vamsee K. Pamula

Abstract Detection of explosives is one of the problems for which an electronic nose is the most appropriate technological solution. Currently, landmines are detected by dogs, which use their noses to sniff explosive vapors or particles. With the current technology it would take about a thousand years and hundreds of billions of dollars to clear all the mines in the world [1]. An electronic nose used in this context would save human lives, work round the clock without getting tired, and could improve security for all humans. In this chapter, a review of different state-of-the-art technologies developed for sensing explosives for the detection of landmines is presented. Various sensors are compared with respect to their detection limits of explosives such as trinitrotoluene and dinitrotoluene, because they are found to be the predominant explosives found in landmines. The system developed by Nomadics is identified to be the best of the currently available detection devices. Future success of the electronic nose in this area depends on the ability of these devices to outperform the dogs. Such systems will emerge within the next decade.

23.1 Introduction

There are some horrifying facts about landmines [1]. Around the world they claim the life of a victim or maim one victim every 22 minutes. There are about 120 million unexploded landmines lurking in 70 countries around the world. With the current technology, 4.6 square miles of landmine infested area can be cleared per year. For every mine that is cleared, 20 new mines are laid. The cost of a mine ranges from $3–$5, whereas clearing it costs $1000. On average, for every 5000 mines removed, one mine-clearer is killed and two others are injured. It would cost about $120 billion and take a thousand years to clear all the mines in the world with the current technology.

The insidious nature of mines has stimulated significant research – spanning over half a century – on techniques for mine detection, identification, and remediation. In the context of detection, technologies that have been explored include magnetic metal detectors, ground penetrating radars, optical, infrared, acoustic, X-ray, and thermal neutron analysis. The identification problem is even more daunting, requiring technology and algorithms that can selectively detect landmines among the clutter. Devel- opment of a sensor that is both sensitive and selective for all kinds of landmines under all conditions is almost impossible. Realizing that there is no single sensor that works, a multi-sensor approach needs to be taken for the detection of the mines. Currently, demining is done by humans with simple metal detectors or a human- dog team. Metal detectors have very high false alarm rates due to the metallic junk in a minefield. Also, they cannot detect plastic mines, which have almost no metallic content. Dogs have proven to be the most effective mine detectors to date, although they have limited attention spans measured in tens of minutes. Mine clearing by teams using metal detectors proceeds at 200 meters/day, whereas a human-dog team allows 2–4 kilometers/day to be cleared [2]. Dogs search by placing their noses close to the ground and inhaling vapors as well as solid particles of the material to be detected. It is not clearly understood whether they detect the pure explosive, some impurities associated with the explosive, or some signature of the odorant [3]. A sensor that combines both the vapor and particle detection will be the closest approximation to a dog’s nose. Such a sensor will work round the clock! One of the most important military and humanitarian applications of the electronic nose is to sniff out landmines. Most of this chapter will concentrate on various technologies developed to date for sensing explosive vapors in this context.

23.2 Previous Work

Semiconductor vapor sensors have been developed in the past [4, 5]. A complementary approach to obtaining increased sensitivity is to detect the particles of explosive residues in addition to the vapors around a landmine. This particle sensor, used in conjunction with the vapor sensor, would approximately mimic a dog’s functionality in detecting the landmines. Most of the commonly used landmines contain 2,4,6-trinitrotoluene (TNT) and/or 1,3,5-trinitro-1,3,5-triazocyclohexane (RDX) as the explosive charge. It has been observed from experiments that at least a few nanograms of TNT explosive particles are present in the vicinity of landmines. For a buried landmine, vapors of the explosive charge emanate from the casing of the mine into the soil and further into the air above it. Many explosives have very low vapor pressures, including TNT and RDX. The equilibrium vapor concentration of TNT is about 70 picograms/mL of air at 298 K [6]. Due to low vapor pressures of the explosives, the concentration of the vapors above a landmine are very low. Most of the contaminants present in TNT have a higher vapor pressure than TNT itself. For a particular mine, 2,4-dinitrotoluene (DNT) vapors were found to be 20 Â more concentrated than those of TNT vapors, even though DNT accounted for less than 1 % of the explosive by mass 23.3 State-of-the-art of Various Explosive Vapor Sensors 549

[7]. The mixture of the compounds escaping the landmine form a ‘chemical signature’ indicative of the explosive present in the landmine. Significant success has been reported by using trace explosive particles for substance identification [8]. Indeed, both RDX and TNT have been detected at higher levels than expected, when the vapor sampling system was augmented with a trace particle collector [9]. At Auburn University, the researchers found that dogs that were trained to detect TNT learn to use DNT as a detection odor signature. In their experiments the dogs were able to sense DNT in fractions of the parts per billion range. While evaluating the nature of olfaction to determine whether the particles or vapors play the main role in detection, they found that particles did not reach the olfactory epithelium of the dogs, which suggests that the particles may not be a likely basis for a detection scheme [10]. Researchers at Penn State University studied the flow patterns of air generated by a dog while sniffing. They observed that the sniffer must approach the scent source in close proximity to avoid dilution of the scent and disruption by wind. They also point out that particles may play a role in scenting as they observe that the particles on the surface become airborne while the dog is sniffing. Based on this research, the electronic dog’s nose should be aerodynamically designed to sniff efficiently [11]. A single solution does not exist for the landmine problem, therefore a variety of sensors would be needed to successfully replace dogs. Sandia National Labs’ studies [12, 13] indicate that the dogs seem to work better in wet conditions because water competes for soil sorption sites thereby enabling release of explosive vapors. Also, their experiments on a buried landmine made of TNT revealed that the vapor above the soil is that of DNT, and also that DNT passes through the mine casing more easily than TNT, therefore DNT ends up in a higher concentration on the surface of the soil.

23.3 State-of-the-art of Various Explosive Vapor Sensors

In this section, we will cover the work performed in developing electronic noses for explosive detection. In 1997, the Defense Advanced Research Projects Agency (DAR- PA) developed a high-risk technology development program to detect mines through their chemical signature. In view of the arguments presented in the previous section, a number of researchers concentrated their efforts in developing a sensor which mimics a dog’s nose, if not exactly at least functionally, for the detection of explosives. Research was performed on mammalian olfaction which stimulated new ideas for chemical sensing.

MIT Swager’s group from MIT has developed fluorescent conjugated polymer thin-films that have high affinities for DNT, TNT, and related compounds. The incorporation of rigid three-dimensional pentiptycene moieties in the conjugated polymer backbone prevent p-stacking or excimer formation, which allows the diffusion of analytes 550 23 Detection of Explosives

Fig. 23.1 Fluorescence quenching mechanism in polymer chemosensor films

into the dense polymer films. The fluorescence of the films reduces in a few seconds due to the vapors of TNT and DNT. The authors believe that the reduction in fluorescence is due to the exchange of the excited electrons of the polymer film with the electron-deficient DNT or TNT molecules [14]. In this process, TNT short circuits the migrating electron by allowing it to jump back to the valence band without the emission of light as shown in Fig. 23.1. Since the polymer molecules are wired serially, the TNT short circuit amplifies the reduction in fluorescence.

Duke University At Duke University, a microelectromechanical systems-based explosive particulate sensor was developed [15]. The purpose of this sensor is to complement the vapor sensors by detecting the explosive particles from the soil to aid more accurate detection. As mentioned earlier, a few nanograms of DNT and TNT are present on the surface of the soil near a buried landmine. The sensor comprises of a bimetallic gold (0.5-lm-thick)/polysilicon (1.5-lm-thick) surface micromachined cantilever. Due to a large difference in the thermal coefficients of expansion between gold and polysilicon, the cantilever deflects down upon heating.

Fig. 23.2 Schematic of a cantilever’s response to a deflagrating explosive particle 23.3 State-of-the-art of Various Explosive Vapor Sensors 551

A few nanograms of pure DNT was placed on the pad of the cantilever as shown in Fig. 23.2. When the cantilever is heated without an explosive particle, it deflects downwards monotonously. But when the cantilever is heated with the explosive particle, the cantilever’s deflection shows an additional dip around the temperature when the explosive particle disappears from the pad. The cantilever is heated at 6 8C/sec. It is assumed that the deflagrating explosive particle is generating this additional heat. The magnitude of the dip in deflection corresponds to the size of the DNT particle [16]. For nanograms of DNT particles, it was always observed that the particles release energy giving rise to the dip around 110 –120 8C.

Tufts Dickinson et al. from Tufts University have developed the first optical artificial nose [17]. As explained in Chapter 8 (Optical electronic noses), thousands of bead sensors are randomly dispersed across an etched fiber optic tip. Each bead sensor within the array is a porous silica bead impregnated with the environmentally sensitive dye, Nile Red, which is a solvatochromic dye (highly sensitive to the polarity of its local environment) as shown in Fig. 23.3. The sensor array is connected to a charge-coupled device (CCD) camera detector which monitors the fluorescence with an imaging system. On exposure to a particular vapor, the bead sensors undergo characteristic and reproducible fluorescence intensity and wavelength shifts that are used to generate time-dependent fluorescence response patterns. Each of these sensor beads is cross-reactive (not analyte specific and broadly selective) and produces a unique fluorescence signature in response to different analytes. These patterns can be used to train pattern recognition computational networks. On subsequent exposure to the same 552 23 Detection of Explosives

analyte vapor, the system identifies the vapor by the characteristic response pattern of the sensors. They have found that the surface chemistry of the sensor favors attraction between the electron-accepting nitroaromatic compounds such as DNT and TNT, and the highly adsorptive surface of the porous silica beads thus maximinzing the analyte- dye interactions [18, 19]. It was demonstrated that the detection limit can be enhanced due to the increase in the signal-to-noise ratio when the signal is collected over a thousand sensors and averaged as shown in Fig. 23.4. The sensors were able to respond to vapor concentrations of DNT and TNT up to tens of ppb (parts per billion).

Draper Laboratories Caltech Carbon black-insulating organic polymer composite films are employed in an array of vapor detectors. These vapor detectors are cross-responsive and respond by exhibiting

Fig. 23.5 Caltech/Draper sniffer assembly for landmine detection 23.3 State-of-the-art of Various Explosive Vapor Sensors 553 a change in their resistance on exposure to a particular vapor. Each element of the array contains a different organic polymer as the insulating material. The resistance between the electrodes of an element changes due the swelling of the polymer and varies due to the differing gas-solid partition coefficients for the various polymers of the detector array. No individual sensor is uniquely responsive to a given analyte, but the swelling pattern across all the elements of the array is unique for each odor. The response is matched to an existing pattern that aids in the classification and quantification of analytes in the vapor phase. The pattern type of the response allows identification of the vapor and the steady-state pattern height allows quantification of the analyte. In association with Draper Laboratories, Caltech’s vapor sensors were incorporated into a sniffer that collects the volume of air above a mine and delivers it to the sensor arrays. The sniffer head has two sensor chips opposite each other through which the sniffed vapor is investigated as shown in Fig. 23.5. They were able to detect DNT in the low ppb range in less than 5 seconds of exposure to the vapor [20].

Rockwell Science Center Rockwell Science Center developed a miniaturized mass detection system, which has an array of polymer-coated thin-film resonators (TFR) operating at 2 GHz as shown in Fig. 23.6. An array of eight TFR sensors, which change their resonance frequencies as a function of the mass of the vapor adsorbed in the polymer coatings, has been developed to detect vapors of TNT and its decomposition products for landmine detection. The surface coatings of these sensors preferentially adsorb specific types of chemical vapors. The TFRs were fabricated using AlN as the piezoelectric film with a thickness of 1.5 lm. The polymer coating was sprayed onto the TFR into thin films because thicker coatings degrade the quality of the acoustic resonance [21]. Out of the eight sensors, three were coated with polymers that have affinity for aromatic nitrates, one with affinity to water, three for varying degrees of adsorption of organic materials, and one left uncoated as a general reference [22]. The sensors recognize the target vapors and quantify their concentration by comparing the pattern of the response, which is based on the magnitude and time-dependence response of all the coated detectors, with a known pattern for that particular vapor. The system was able to detect DNT at few ppb concentration in air in the absence of large background levels of interference.

Fig. 23.6 Cross-section of the thin film resonator microbalance for vapor detection adapted from [22] 554 23 Detection of Explosives

Fig. 23.7 SPEC’s explosive particulate sampler

Naval Research Laboratory The US Naval Research Laboratory has developed polymer coatings for surface acoustic wave (SAW) sensors to be used for explosive vapor detection. SAW resonator devices (acquired from SAWTEK Inc, Orlando FL) were spray-coated with various polymer films to evaluate the most promising polymers for the vapor detection of nitroaromatic explosive compounds. As the coatings absorb the vapors, the resonance frequency of the polymer-coated SAW device decreases due to increased mass loading. Several hexafluorisopropanol-functionalized aromatic silicon-based polymers have been prepared and coated on the SAW devices to enhance the detection of nitroaromatic analytes. The polymers are strongly hydrogen-bond acidic which reversibly sorbs nitroaromatics and other hydrogen-bond basic vapors. They estimate that the detection limit for these sensors will be < 100 ppt (parts per trillion) for DNT [23].

Texas Instruments Texas Instruments’ Spreeta sensor, when used in conjunction with a sniffer from SPEC (Systems & Processes Engineering Corporation), closely mimics the dynamics of a dog’s nose. The SPEC sniffer has six exhaler orifices from where the particulates are stirred up and then drawn through the sampling orifice, as shown in Fig. 23.7. These particulates then impinge on a membrane. TNT and DNT from the sample dissolve into the membrane and rapidly diffuse to the liquid side. An automated microfluidics system mixes the sample with antibodies, which can then be delivered to the Spreeta sensor for analysis using a bioassay. The Spreeta sensor is based on the principle of surface plasmon resonance (SPR). SPR can be employed to study the kinetics of molecular binding events in real-time. 23.3 State-of-the-art of Various Explosive Vapor Sensors 555

Fig. 23.8 SPR sensor for detecting DNT and TNT dissolved in a liquid. Adapted from [25]

On an active gold surface, the refractive index changes can be monitored by SPR. As shown in Fig. 23.8, the liquid sample containing dissolved TNT and DNT is delivered to the top of the gold’s surface through a flow cell. An AlGaAs infrared light-emitting diode with a transverse magnetic polarizer excites the surface plasmons in the gold film at the liquid interface. The reflected light is captured on a photodiode, which yields the refractive index of the liquid on the sensor. It also has a built-in temperature sensor because refractive indices vary with temperature [24]. A glass slide coated with gold is used as the SPR active surface. The slides were coated with trinitrobenzene (TNB) and bovine serum albumin (BSA), which serves as an immobilized ligand. Since TNT is a small molecule and SPR detects changes in the surface mass concentrations of an analyte, a competition assay is used. A TNT antibody with a large molecular weight is used, the binding of which is competitively prevented by free TNT in the solution. When the TNT antibody binds to the TNB-BSA groups in the gold surface, an increase in refractive index is observed. When TNT is present in the solution, however it reduces the rate of antibody adsorption leading to a reduced value of the refractive index. The limit of detection of this sensor is 1 ppm (parts per million) of TNT (1 mg of TNT/1 kg of soil) [26].

ETC Laboratories EIC Laboratories have made a vapor sensor based on surface-enhanced raman spectroscopy. A laser interrogates an area of a microscopically roughened metal for adsorbed analytes. The vibrational modes of the analyte adsorbed on the metal are enhanced compared to their nonresonant Raman intensities. The metal surface can be made to selectively adsorb compounds of similar chemical structure by choosing a combination of the metal surface, the degree of roughness, the degree of oxidation of the surface, and other factors. The Raman spectra are collected using an echelle spectrograph coupled to an air-cooled CCD camera. The raw vapor spectra are presented to a software algorithm which creates a curve fit and compares it to the anticipated curve for DNT, and therefore ascertains the presence or absence of the analyte. The laser signal is delivered through a fiber-optic probe and the spectrometer was packaged for use in the minefield [27]. 556 23 Detection of Explosives

Fig. 23.9 Raman spectra of (a) TNT, (b) DNT, and (c) DNB. Raman spectra were normalized. The intensity axis was not plotted for illustrative purposes [29]. Reproduced with permission from Anal. Chem. (2000) 72, 5834–5840. Copyright 2000 Am. Chem. Soc

Figure 23.9 shows the Raman spectra of TNT and its primary impurities, DNT and DNB. EIC sensors were able to detect the presence of sub-ppb concentration of DNT over aqueous solutions.

Quantum Magnetics Quantum Magnetics (QM) has developed a sensor based on quadrupole resonance, which is similar to the magnetic resonance imaging technique used in the medical industry. QM in a subsidiary In Vision Technologies, which produces X-ray computed-tomography machines for scanning airport cargo and baggage. The QM instrument is not specifically an electronc nose in that it does not detect the explosives through vapors or particles, but it is chemically specific enough to detect explosives. The device sends short pulses of radio waves at specific frequencies that reso- nate with the atomic nuclei of the explosive molecules. At the end of the pulsing, the nuclei send out a weak radio signal. Out of 10 000 compounds studied, there has not been an overlap in the responses. 14N nuclei gives the characteristic signal in the case of TNT and RDX. There are no false alarms due to other nitrogen-containing compounds available in the background because the signal is either not given or is given at a sufficiently different resonance frequency. This signal depends on the molecular structure of the atoms, which is analyzed by a computer to identify the material [28]. 23.4 Case Study 557

Field tests performed to detect RDX- and TNT-based nonmetallic antitank and antipersonnel mines, yielded 100 % probability of detection with very low false alarm rates [29].

Nomadics Another scheme for detecting explosives is with electrochemical sensors, which yield qualitative information about the presence of these compounds. In voltammetry, the potential of a sensor is held constant and the sensor detects the current resulting from electrochemical oxidation or reduction. In this case, the signal may be disturbed due to the presence of other substances, or adsorbates may form on the electrode surface rendering the sensor less sensitive over time. To overcome these problems, cyclic voltammetry was employed where a time-varying potential was applied on a gold electrode in sulfuric acid and the resulting current recorded as a function of the potential. This sensor is in its early stages and the detection was demonstrated only for TNT in the gaseous phase [30].

23.4 Case Study

Nomadics Inc., in Stillwater, Oklahoma, is developing a highly sensitive and selective landmine detector based on the detection of the trace amounts of TNT vapors emanating from a landmine. Nomadics’ Fido (Fluorescence Impersonating Dog Olfaction)

Fig. 23.10 Nomadics’ Fido landmine detector [31] 558 23 Detection of Explosives

Tab. 23.1 Vapor Detection limits of various systems

Detection Method Limits (/mL)

High Performance Liquid Chromatography Ultraviolet (HPLC-UV) 1 nanogram (ng) Mass Spectrometer 800 picogram (pg) High Performance Liquid Chromatography 600 pg Electrochemical (HPLC-EC) Thermal Energy Analysis (TEA) 30 –50 pg Mass Spec – Chemical Ionization (MS-CI) 20 pg Airport Sniffers 20 pg Electron Capture Detector (ECD) 10 pg Micro Electron Capture Detector (lECD) 1 pg Ion Mobility Spectrometer (IMS) 50 –100 femtogram (fg) Nomadics Amplifying Fluorescent Polymer 1 fg

landmine detector has demonstrated the ability to detect landmines under field conditions, and is perhaps one of the most promising explosive detection technologies on the market. It is based on the fluorescent polymer beads developed by Swager’s group at MIT as detailed in 23.3 above [14]. Nomadics’ landmine detector uses the same technology as shown in Fig. 23.0, which is based on amplifying fluorescent polymers. The fluorescence of many polymers decrease when a single molecule of the nitroaromatic compound binds to a polymer. In its handheld configuration, the system consists of a small sensor module, detector electronics, operator display/control panel, battery pack, and mounting arm. As shown in Fig. 23.10, a blue-light fluorescence excitation laser is collimated and filtered to pass a narrow band of light around 405 nm. This beam is normally incident through two borosilicate glass substrates coated on the surfaces with spin-cast thin films of the pentiptycene polymer. The coated substrates are held in a cassette that can be easily removed from the device to facilitate the replacement of the polymer films. A small gap is maintained between the two substrates by a thin-U-shaped spacer. The spacer forms a seal along three edges of the polymer-coated inner faces of the substrates. The substrates are not sealed along the fourth edge. This opening serves as a sample inlet. Vapor is drawn through the inlet into the sampling volume between the two substrates by a small pump. The pump is connected to an exit port bored through the spacer on the side opposite the inlet. Transmitted incident light, along with the emitted fluorescent light, is passed through a filter which passes only the fluorescence signal at 460 nm. The intensity of the emission from the films is then measured with a photomultiplier tube (PMT) [32]. Seventy one soil and water samples containing landmine explosives with potential interferants and blanks were presented to this detector, which has successfully identified each of them without any single error in the laboratory conditions. Blind field testing was performed by DARPA at Fort Leonard Wood test-field over real landmines. The probability of detection was 0.89 with a probability of false alarms of 0.27. Noma- dics soon hopes to be in full production of field-deployable Fido landmine detectors. The current sensor prototype can instantly detect in the parts per quadrillion range, 23.6 Future Directions 559 which is better than most of the current explosive detection methods (Table 23.1). To the author’s knowledge, this is the first sniffer capable of detecting landmines in the field with performance comparable to that of dogs.

23.5 Conclusions

In conclusion, we have presented an overview of sensors that detect either the vapors or particles of the explosives commonly found in landmines. The sensors are made in a variety of technologies, each having their own advantages and disadvantages for field deployment. Most of the sensors presented in this chapter are beyond the proof-of- concept stage and many are driven by the industry for commercialization, though there are no commercial products available in the market yet that can sniff landmines in the field. Among the electronic noses made for explosive detection, currently No- madics’ FIDO landmine detector has shown capabilities that match those of dogs. The successful detector will have characteristics such as portability, high sensitivity to the explosive vapors and selectivity to detect only those vapors among clutter, a friendly interface for the deminers, very low false alarm rates with low maintenance, and will be very robust.

23.6 Future Directions

Once a commercially viable electronic nose for landmine explosive detection is available, the potential customers include professional deminers, humanitarian demining groups like the United Nations and the International Committee of the Red Cross, various non governmental organizations, land and economic developers, and govern- ments of countries affected by landmines. Currently, there are about 120 million mines deployed around the world, which would cost about 120 billion dollars to de- mine. This presents a huge market opportunity for any company that comes up with a suitable solution. Some research groups have already demonstrated that the elctronic noses developed by them have comparable sensitivity to that of a dog’s nose. With the increasing awareness of the landmine problem and various companies and university- based research groups around the world working on the problem, it may not be long before a commercial electronic nose, which provides a better solution than a dog in many ways other than just the nose aspect, will successfully emerge. An electronic nose to detect landmines and explosives would be required to operate in situations that will be dangerous to human life. One that is integrated with tele- operation capability or a robot will be far more attractive in such situations, but such an autonomous electronic nose is still years away from becoming a reality. 560 23 Detection of Explosives

References

1 http://www. icrc. org/eng/mines 20 S. M. Briglin, M, C, Burl, M. S. Freund, N. S. 2 A. M. Prestrude, J. W. Ternes. Proc. of SPIE, Lewis, A. Matzger, D. N. Ortiz, P. Toku- 2093, (1994) 633–643. maru. Proc. of SPIE, 4038, (2000), 530–538. 3 L. J. Myers. Critical Reviews. CR42, (1994), 21 P. Kobrin, C. Seabury, C. Linnen, A. Harker, 93–103. R. Chung, R. A. McGill, P. Matthews. Proc. 4 R. T. Howe, R. S. Muller. IEEE Transactions of SPIE, 3392, (1998), 418–423. on Electron Devices, ED-33, 18–19; (1986) 22 C. Linnen, P. Kobrin, C. Seabury, A. B. 499–506. Harker, R. A. McGill, E. J. Houser, R. 5 A. Wilson, M. Tamizi, J. D. Wright. Sensors Chung, R. Weber, T. Swager. Proc. of SPIE, and Actuators B (Chemical), 18–19, (1994) 3710, ( 1999), 328–334. 511–514. 23 E. J. Houser, R. A. McGill, V. K. Nguyen, R. 6 J. Yinon, S. Zitrin. Modern Methods and Chung, D. W. Weir. Proc. of SPIE, 4038, Applications in Analysis of Explosives, John (2000), 504–510. Wiley and Sons, New York, (1993). 24 J. Mendelez, R. Carr, D. U. Bartholomew, K. 7 V. George, T. F. Jenkins, J. M. Phelan, D. C. Kukanskis, J. Elkind, S. Yee, C. Furlong, R. Leggett, J. Oxley, S. W. Webb, P. H. Miyares, Woodbury. Sensors and Actuators B, 35–36, J. H. Cragin, J. Smith, T. E. Berry. Proc. of (1996), 212–216. SPIE, 3710, (1999) 258–269. 25 R. G. Woodbury, C. Wendin, J. Clenden- 8 T. G. Sheldon, R. J. Lacey, G. M. Smith, P. J. ning, J. Mendelez, J, Elkind, D. U. Bartho- Moore, L. Head. Proc. of SPIE, 2092, (1994), lomew, S. Brown, C. Furlong. Biosensors 145–160. and Bioelectronics, 13, (1998), 1117–1126. 9 W. R. Davidson, W. Scott. Proc. of SPIE, 26 A. A. Strong, D. I. Stimpson, D. U. Bar- 2092, (1994), 108–119. tholomew, T. F. Jenkins, J. Elkind. Proc. of 10 J. M. Johnston, M. Williams, L. P. Waggo- SPIE, 3710, (1999), 362–372. ner, C. C. Edge, R. E. Dugan, S. F. Hallowell. 27 J. M. Sylvia, J. A. Janni, J. D. Klein, K. M. Proc. of SPIE, 3392, (1998), 490–501. Spencer. Anal. Chem., 72, (2000), 5834– 11 G. S. Settles, D. A. Kester. Proc. of SPIE, 5840. 4394, (2001). 28 A. D. Hibbs, G. A. Barrall, P. V. Czipott, A. J. 12 S. W. Webb, J. M. Phelan. Proc. of SPIE, Drew, D. Gregory, D. K. Lathrop, Y. K. Lee, 4394, (2001), 474–488. E. E. Magnuson, R. Matthews, D. C. Skvo- 13 V. George, T. F. Jenkins, J. M. Phelan, D. C. retz, S. A. Vierkotter, D. O. Walsh. Proc. of Leggett, J. Oxley, S. W. Webb, P. H. Miyares, SPIE, 3710, (1999), 454–463. J. H. Cragin, J. Smith, T. E. Berry. Proc. of 29 A. D. Hibbs, G. A. Barrall, S. Beevor, L. J. SPIE, 4038, (2000), 590–601. Burnett, K. Derby, A. J. Drew, D, Gregory, C. 14 J.-S. Yang, T. M. Swager. J. Am. Chem. Soc, S Hawkins, S. Huo, A. Karunaratne, D. K. 120, (1998), 11864–11873. Lathrop, Y. K. Lee, R. Matthews, S. Milber- 15 V. K. Pamula. Ph. D. Thesis. Department of ger, B. Oehmen, T. Petrov, D. C. Skvoretz, S. Electrical and Computer Engineering, Duke A. Vierkotter, D. O. Walsh, C. Wu. Proc. of University, (2001). SPIE, 4038, (2000), 564–571. 16 V. K. Pamula, R. B. Fair. Proc. of SPIE, 4038, 30 T. Berger, H. Ziegler, M. Krausa. Proc. of (2000), 547–552 SPIE, 4038, (2000), 452–461. 17 J. White, J. S. Krauer, T. A. Dickinson, D. R. 31 M. la Grone, C. Cumming, M. Fisher, D. Walt. Nature, 382, (1996), 697–700. Reust, R. Taylor. Proc. of SPIE, 3710, (1999), 18 K. J. Albert, D. R. Walt. Anal. Chem., 72, 409–420 (2000), 1947–1955. 32 M. la Grone, C. Cumming, M. Fisher, M. 19 T. A. Dickinson, K. L. Michael, J. S. Krauer, Fox, S. Jacob, D. Reust, M. Rockley, E. D. R. Walt. Anal. Chem., 71, (1999), 2192– Towers. Proc. of SPIE, 4038, (2000), 553– 2198. 562. 561

24 Cosmetics and Fragrances

P. A. Rodriguez, T. T. Tan, and H. Gygax

Abstract The use of electronic noses in the cosmetic and fragrance industry appears limited when compared to other industries and areas of application, such as the food and beverage industry, or the chemical, polymer, and plastic industries, or in environmental and medical applications. However, the literature and the work we present in this chapter show that, with optimization, many challenging problems in the cosmetic and fragrance industry can be successfully addressed using electronic nose technology. In this chapter we describe key challenges and limitations of analytical instruments expected when correlating their output with the human response to perfume-related samples. We also include two industrial applications addressed by the use of commercially available instruments; one based on a chemical sensor, the other on a mass spectrometer. They provide insights into the ability of electronic noses to match and mimic the perception of odor by humans, as well as their ability to compete with well-established analytical methods. Good sensitivity, selectivity, and reproducibility were obtained in the two cases presented here.

24.1 Introduction

Perfumes, derived from plants or flowers, have been used for millennia as a means to enhance the quality of life. Today, perfumes are ubiquitous in society, we encounter them in cosmetics, in the home environment, and in virtually every cleaning product. As a consequence, perfumery has become a global, multibillion-dollar industry. Although the industry employs modern, sophisticated analytical tools to ensure the quality of their products, the creation of a winning fragrance is still an art. Skilled perfume designers (also known as perfumers), rely on intuition, market research, and knowledge of raw materials to create perfumes designed to meet the requirements of a particular product. An important requirement, in addition to meeting cost constraints, is to deliver a fragrance that reinforces the product image. Thus, if you are developing a cleaning product, the perfume is likely to be required to deliver a ‘clean’ fragrance.

Translating words like ‘clean’ and ‘fresh’ into chemical formulas useful in a product is part of the art of perfumery. But, unlike music or painting, the art of perfumery has fewer standardized tools to accomplish the objective of creating a masterpiece. The difficulty in developing the necessary tools is a consequence of our inadequacy in using words to define odors. The closest we come to defining an odor is through the description of how it resembles other familiar, well-known odors. Interestingly, although there have been many attempts to describe or classify odors, no scheme has survived the test of time. Nevertheless, a successful classification scheme would provide a useful framework for understanding odors, and would facilitate efforts aimed at identifying elusive, primary odorants. In addition to its academic importance, the need to describe and classify odors has enormous economic consequences. This is true because description and classification would provide a link to understanding preference. Importantly, consumers throughout the world often use odor preference to discriminate between products that otherwise offer similar price/performance attributes. Description, classification, and understanding preference are areas where ‘electronic noses’ or ‘chemical classification tools’ could make unique contributions to perfumery. In this chapter we discuss the requirements, characteristics, and usage of commercial electronic noses in the perfume and cosmetics industry. A comparison with a gas chromatography (GC)-based approach is also presented. The chapter ends with an assessment of the technology for future applications in this market segment.

24.2 The Case for an Electronic Nose in Perfumery

Perfumes are complex mixtures of volatile and semi-volatile organic compounds [1]. Today, it is not uncommon for a commercial perfume to be prepared by mixing fifty to one hundred or even more perfume raw materials (PRMs). Furthermore, PRMs are not pure chemical compounds. Because many are obtained by complex processes and/ or derived from complex raw materials, PRMs may contain many isomers or even compounds unrelated to the main odorant in the PRM. For example, the main odorant in Galbanum PRM is a pyrazine accounting for less than 0.01 % of the total mass in this PRM. As a consequence, we find it is not uncommon for a finished perfume to contain hundreds or even thousands of distinct chemical compounds. Unfortunately, perfume complexity quickly adds to the perfume cost. This is true because in addition to inventory costs, specifications for each PRM must be established and confirmed by analysis, and safety evaluations must be performed on the many possible compounds present above a certain percentage. Presently, the chemical complexity of perfumes is mainly a consequence of the desire by perfume designers to deliver perceptual complexity to consumers. It is the richness of the perception that makes perfumes so attractive to humans. Unfortu- nately, the link between chemical complexity and perceptual complexity has not been thoroughly examined. As a consequence, the optimum number of compounds in a perfume has not been established. Recent work suggests it is possible to reduce 24.3 Current Challenges and Limitations of Electronic Noses 563 the chemical complexity of existing perfumes without any measurable reduction in perceptual complexity [2]. Reductions in the number and quantity of chemicals can be achieved through the use of psychophysical principles and the use of humans to establish the relative importance of individual odorous compounds to the overall perfume fragrance. Although reductions of 20–40 % in the mass and number of PRMs can be achieved when examining existing perfumes, the process is not straightforward and requires a number of iterations. Because the reasons to do the work are so compelling, i.e. cost reductions, raw material inventory simplification, and the elimination of tens of thousands of metric tons of materials from the environment, the industry and the planet would certainly benefit from a rapid and simple process to do the work. There are reasons to be optimistic about the use of an electronic nose as a tool to help simplify existing perfumes and help design new, cost- and material-efficient fragrances. Although the initial report by Axel and Buck [3] on the identification of human genes that code for olfaction, suggested the existence of perhaps 1000 such genes, or corresponding ‘sensor molecules’, recent work suggests that number is significantly smaller. If true, the number of required sensors may approach the number of ‘sensing channels’ measurable with a mass spectrometer as discrete ions, and reported as mass/charge (m/z). In addition, our work to understand the relative importance of perfume odorants suggests that only 10 to 15 compounds contribute most of the intensity and character to any given perfume. We could call those compounds ‘principal odorants’, and although they would be different for each perfume, we find that most perfumes contain many of the same compounds as principal odorants. Thus, if sensors could be developed to be quasi-selective for those compounds, we would expect the resulting electronic nose to have near-perfect correlation with humans judging variations in perfumes.

24.3 Current Challenges and Limitations of Electronic Noses

Humans are highly sensitive and selective sensors of perfume components. For example, odor detection thresholds (ODTs) are in the low- or sub-part-per-billion (ppb (volume/volume, v/v)) range for many compounds used as principal odorants in current perfumes. In addition, the human selectivity for certain odorous materials allows perception of those odorants when in the presence of much higher concentrations of other compounds. For example, 10 ppb (v/v) of a jasmonate in the headspace of a product would deliver a fresh, floral fragrance to consumers. Humans would perceive the jasmonate fragrance even in the presence of 1000-fold excess (10 ppm (v/v)) of limonene (or orange terpenes), a widely used PRM. The jasmonate/limonene example is by no means a rare case or exception. Indeed, key odorous compounds classified as principal odorants of a perfume often account for a small fraction of the product headspace composition. The reality is that humans have no problem ignoring the bulk of the compounds in the headspace and sensing principal odorants in product. Perfumers take advantage of human selectivity towards the 564 24 Cosmetics and Fragrances

principal (and other) odorants to deliver desirable fragrances to products. Indeed, human selectivity is of paramount importance to perfumery. Sensitivity and selectivity are also important to efforts aimed at developing an electronic nose for perfumery. State-of-the-art electronic noses employing a few sensors or sensing strategies, with selectivities vastly different from those of the human, are likely to be limited to perfumes where the principal odorants are a major fraction of the headspace composition.

24.4 Literature Review of Electronic Noses in Perfumery and Cosmetics

The use of electronic noses in cosmetics and perfumery appears limited compared to other areas, e.g. food, beverages, chemicals, polymers, and plastics. The limited use is reflected in the number of publications. We found about twenty publications specifically addressing cosmetics and perfumery, while the published reports in other areas reach into the hundreds. Interestingly, references including perfumery applications often include the development of new chemical sensors designed to enhance sensitivity and selectivity. For example, Kusumoputro and Rivai of Indonesia University [4] used quartz resonator crystals with lipid membranes to discriminate fragrance odor. Quartz resonators are also known as quartz microbalances (QMB) or quartz crystal microbalances (QCM). Using those sensors and an artificial neural network, they achieved high recognition accuracy when determining the correct percentage of aroma from Martha Tilaar cosmetics products and five flavors from Splash Cologne products. Byfield et al. [5] also demonstrated the use of quartz crystal resonators in the fragrance and petrochemical industries, and in another case [6] demonstrated chiral discrimination with a QMB sensor. This development is especially important to perfumery were optical isomers may have clear differences in odor. Chiral recognition was achieved by coating the crystals with compounds such as heptakis (2,3,6-tri-o-methyl)-beta-cyclodextrin, and octakis (6-o-methyl 2, 3-di-o-pen- tyl) gamma-cyclodextrin dissolved (as 50 % and 20 % (w/w) solutions) in OV1701, a widely used stationary phase in GC. The sensors showed preferential binding for enantiomers of a- and b-pinene and cis- and trans-pinane. By comparing to elution time in gas chromatography, the observed separation factor was seen to be dependent upon the chiral stationary phase concentration. The results suggest that on-line determination of enantiomeric excess and concentration of certain monoterpenes is possible at room temperature using QMB sensors coated with chiral GC stationary phases. Cao et al. [7] and Yokoyama and Ebisawa [8] have published results related to the development of QMB sensors for use in the fragrance and perfume industry. Both groups concluded that their sensors could correlate with sensory perception and discriminate between different fragrances. Alternative approaches to the use of QMB have also been reported. Hyung-Ki-Hong et al. [9] developed an electronic nose with a micro gas-sensor array. The chemical 24.4 Literature Review of Electronic Noses in Perfumery and Cosmetics 565 sensors were made using thin-film metal oxides. As with the work discussed above, good discrimination between samples was reported for both flavor and fragrances. Two fragrances, a women’s perfume (eau de cologne) and a man’s perfume (eau de toilette) were correctly identified. Penza et al. [10] classified food, beverages, and perfumes using an electronic nose based on the use of a thin-film sensor array and pattern recognition. Using tungsten oxide (WO) with different catalysts, e.g. Pd, Au, Bi, and Sb, good selectivity and sensitivity were obtained to correctly classify the samples in question. The authors concluded the arrays show promise for use in a variety of industries and applications. Letant et al. [11] used porous silicon chips in an electronic nose designed to measure a series of solvent vapors, ethyl esters, and perfumes. The chemical information from the porous silicon sensors was obtained by measuring changes in reflectivity and photoluminescence. Good reversibility and reproducibility were obtained. They also compared results with those obtained using metal-oxide sensors. Recently, a new technique to discriminate Yves Saint Laurent (YSL) perfumes by means of an electronic nose was described by Carrasco et al. [12]. The authors addressed an off-odor problem reported by an expert panel at Sanofi Beaute´. Three YSL perfumes, Paris eau de toilette, Paris eau de toilette with an off-odor and Opium eau de toilette were analyzed. The differences between samples were also apparent in their GC profiles. However, to meet the needs of a perfume quality control laboratory, where the analysis would need to be faster than possible by GC, GC-mass spectrometry (GC-MS) and/or sensory analysis, an electronic nose was considered. The methodology included the use of Fox4000 Electronic nose (Alpha MOS, France) and an autosampler. The system was equipped with 18 metal-oxide sensors. The only sample preparation technique used was to allow alcohol evaporation before analysis, because the sensors are sensitive to alcohol. The procedure allowed 35 ll of eau de toilette samples, deposited onto a 2-cm2 paper strip placed inside a 10 ml headspace vial, to evaporate in air. The authors concluded that the electronic nose could correctly identify 100 % of all the samples in their respective perfume families, within 30 min, and without using elaborate sample preparation techniques. They also recommended that the electronic nose be considered, along with classical techniques such as GC-MS or infrared spectroscopy, as another useful tool for studying perfume volatiles. Feldoff et al. [13] studied the use of electronic noses with metal-oxide sensors and MS-based sensors as tools for the discrimination of diesel fuels. No sample preparation other than the use of a static autosampler was necessary for both the chemical and MS-based sensors. Good correlation was found between the samples, which corresponded to the origin of the fuel for both types of instrument. In this particular application, data obtained with the MS-based sensor was reported to be easier to obtain, and more reproducible, compared to data obtained by the use of chemical sensors. In summary, the literature review reveals a number of approaches, ranging from the use of QMB, metal-oxide semiconductors and new sensor types, in conjunction with the use of a number of pattern recognition and sample preparation methods, have been used in the perfume and cosmetic industry. In general, good correlations are 566 24 Cosmetics and Fragrances

reported between analytical data, obtained by means of a number of sensor-based strategies – including MS, and the human response to the sample odor. The use of autosamplers simplifies the tasks within quality control laboratories, while helping to achieve good reproducibility. Although some of the applications reported in the literature are based on the use of experimental sensors or sensing strategies, it is clear that commercially available instrumentation may offer a viable alternative to those sensors or strategies. In addition, commercial instrumentation may offer a viable alternative, or be a powerful adjunct, to conventional GC and or GC-MS analysis, as illustrated by the work of Carrasco et al. [12] and Feldoff et al. [13].

24.5 Special Considerations for using Electronic Noses to Classify and Judge Quality of Perfumes, PRMs, and Products

Today, the human nose is the ultimate judge of the quality of a perfume, PRM or product. This is true even after samples are examined by high-resolution, multi-dimensional chromatographic tools, such as capillary GC/FID/MS (FID – flame ionization detector) or GC/MS/IRD (IRD-inhared detection). The primary reason for the use of humans as judges is that, as mentioned before, they have exquisite sensitivity and selectivity towards certain odorous compounds. Thus, a peak seemingly insignificant in a chromatogram, e.g. the pyrazine in Galbanum, may be the most important odorous compound in a perfume or PRM. As a consequence, it is not uncommon for a perfume or PRM to meet analytical specifications and fail sensory evaluation, or vice versa. Therefore, to successfully address odor issues in the perfume and cosmetic industry it is essential to combine results from analytical and sensory measurements [14]. Unfortunately, human judgment is subjective and somewhat variable. In addition, for any given odorant, a fraction of the population would have ODTs significantly higher/lower (> 10–100 ) than the average population. Thus, to use humans as an analytical tool to judge perfumes one must go through a process designed to:

* select humans for their ability to smell, * teach how to scale intensity and name odorants, * calibrate people over time and correct for ‘drift’.

Such a process is often used to identify and train a number of judges who work individually or as a group, i.e. as in an ‘expert panel’. As a consequence, developing and maintaining expert judges and expert panels is an expensive, laborious, and time-intensive activity. In addition, human fatigue (adaptation) and habituation require special attention be given to testing protocols. Thus, even under the best of circumstances, it is possible to encounter artifacts that hinder the human capacity to judge odors. For those reasons, there is great interest in developing alternatives to the use of expert judges or expert panels in perfumery. 24.6 Case Study 1: Use in Classification of PRMs with Different Odor Character but of Similar Composition 567

Although we addressed the selectivity and sensitivity requirements above, we have not addressed the instrumental analogs of human drift, fatigue, and habituation. First, we must define the terms. Panel drift is a change in panel judgment towards a given, standard stimulus presented over time. Detector drift is a change in output in the absence of an input. However, in addition to detector drift, there is an instrumental ‘classification drift’ similar to that experienced by expert panels. Human fatigue (adaptation) is a decrease in perceived odor intensity as a result of exposure to a constant odorant concentration. In addition to a decrease in perceived intensity, fatigue may also produce changes in the perceived character of an odor when a complex mixture of odorants is used. Fatigue in an instrumental detector is a change (typically a decrease) in output when the device is exposed to a constant input. Detector fatigue is a unique function of detector design, sensitivity and selectivity. It may be the primary factor responsible for instrumental ‘classification drift’. Habituation in humans, as is also true for fatigue, is a decrease in perceived intensity as the human brain grows accustomed to a constant stimulus. Because it is strictly a consequence of how the human brain processes stimuli, it has no corresponding instrumental-sensor analog. Thus, for successful use of electronic noses in perfumery, the detectors must have adequate sensitivity and selectivity, have minimum drift and fatigue, and the signal- processing package must address the problem of ‘classification drift’. To measure how well those requirements are met by available electronic noses, analysts typically use a training set consisting of samples selected to encompass the range of odors expected. The number of samples to be used depends on the ability of the electronic nose to differentiate between extremes, e.g. best-worst, or most similar from most different odor. The following two case studies were selected to illustrate the use of electronic noses and other classification tools to address perfume-related questions. In both cases, special care was given to the sample-introduction phase of the measurement. Autosamplers were used to ensure high reproducibility in generating headspace and introducing the sample into the different detectors. The second case study describes the use of the electronic nose within a production environment whereby results obtained were compared to the current quality methods being used.

24.6 Case Study 1: Use in Classification of PRMs with Different Odor Character but of Similar Composition

24.6.1 The Problem

Because of their high selectivity, humans may perceive odorous compounds in the presence of 103,106 or even larger excess of other non-odorous compounds in air. In other words, the human response towards odorous compounds may exceed the response towards non-odorous compounds by many orders of magnitude. This is in contrast to two common analytical detectors, the FID and the MS in electron ioniza- 568 24 Cosmetics and Fragrances

tion (EI) mode, which would have roughly the same response factor for odorous and non-odorous compounds. Thus, we reasoned that electronic noses and other classification tools, utilizing detectors having roughly comparable sensitivity towards organic compounds, would have problems dealing with samples that have very different odors but have similar bulk chemical composition. An experiment to assess this perceived limitation was designed by C. L. Eddy of The Procter and Gamble Co. For the experiment, eight PRM samples with distinct odor characters but similar bulk composition were selected: bergamot, clementine, grapefruit, lime, lemon, mandarin, orange, and tangerine. Typically, the samples contained >85 % D-limonene. For two samples, orange and grapefruit, limonene together with myrcene, and a-, and b- pinene, accounted for 99 þ % and 96 þ % of the mass, respectively. Importantly, the relative abundance of those four compounds is virtually identical in the two samples. Therefore, we would expect that those two samples would be the most difficult to distinguish. Samples were analyzed by means of an HP 4440 (Hewlett-Packard) chemical sensor and by capillary GC-FID. Results obtained with the HP 4440 were provided by D. R. White Jr., and GC-FID data analysis was performed by K. D. Juhlin, both of The Procter and Gamble Co.

24.6.2 Methods

The HP4440 is a device that combines a headspace analyzer and a bench-top MS. To perform an analysis, the PRM headspace was injected directly into the MS where it was subjected to EI. A mass range was rapidly scanned, and ion currents at each m/z were summed over the duration of the run time, e.g. 1 min. Data were analyzed by tools in the Pirouette suite of chemometric methods (Infometrix, Inc.). For the GC analysis, we chose to analyze the samples as neat oils, using a conventional HP-GC equipped with an autosampler for liquids. We justified this choice, as opposed to using headspace, because the PRM samples were similar in composition and volatility. Analysis time was kept at ca. 15 min, although it could certainly be decreased if desired. A 30-m DB- 1, 0.5-lm-thick, 0.32 mm id column and FID were used to separate and detect the compounds.

24.6.3 Results

We compared results obtained by the two approaches. The HP4440 discriminated the PRMs, with the exception of some overlap of orange and grapefruit, as shown in the dendrogram in Fig. 24.1. Repeat analysis on Day 2 showed good reproducibility. Both SIMCA and K-nearest-neighbors (KNN) classification models predicted Day 2 samples with 100 % accuracy. Mass fragments, (m/z) in decreasing order of discrimination power (DP, a ratio of between-class to within-class variances), are listed as m/z of 24.6 Case Study 1: Use in Classification of PRMs with Different Odor Character but of Similar Composition 569

Fig. 24.1 HCA cluster dendrogram of training set (Day 1) of eight PRMs. Data autoscaled the ion followed by (DP), as follows: 43(4320), 90(2853), 154(2759), 150(2554), 69(1727), 68(1696), 70(1677), 67(1629), 41(1067), 89(1003). Those ions are characteristic of terpene-like compounds, the most likely class of compounds responsible for the odor of the PRMs. As expected, the most prominent ions in the raw data are due to limonene, as this compound accounts for most of the mass in the headspace. Therefore, the general look of the raw data resembles the limonene mass spectrum. Limonene would have a mass spectrum with prominent ions at (listed as m/z followed by relative abundance in parenthesis) 136(25), 121(23), 107(22), 94(27), 93(70), 92(22), 91(18), 79(31), 68(100), 67(63). Therefore, those ions would not be expected to be among the list of ions with high discrimination power. Surprisingly, ions at m/z 67 and 68 appear to have high discrimination (DPs 1629 and 1696, respectively) probably because their relative abundances are a sensitive function of terpene structure. As expected, the highest discrimination power was exhibited by ions of low abundance or absent from the limonene mass spectrum. Thus, two ions at m/z 150 and 154 should not be present in limonene (MW 136), while ions at m/z 90 and 43, if present, should be low abundance ions, i.e. 1 %. Consequently, it may be possible to enhance discrimination between orange and grapefruit PRMs by selecting ions with the highest discrimination power for the analysis. GC-FID chromatograms of the orange and grapefruit PRMs are shown in Fig. 24.2. Peaks labeled ‘A’ are virtually superimposable in the two samples. They correspond to limonene (the largest peak in the chromatogram), myrcene, and a-, and b-pinene. However, a number of other peaks, labeled ‘B’, represent peaks distinctly different in the two samples. Because each peak can be viewed as an independent measure- 570 24 Cosmetics and Fragrances

Fig. 24.2 GC-FID chromatograms of samples of orange (red trace) and grapefruit (blue trace) PRMs. Peaks labeled “A” are nearly identical in the two samples and account for most of the mass under the peaks. Peaks labeled “B” differ significantly between the two samples

ment of a compound in a given PRM, the discrimination of those two samples is simple. The classification results are shown in the dendrogram in Fig. 24.3. Having 20–30 peaks, representing independent variables (because the peaks are a measure of virtually ‘pure’ compounds), probably over-defines this system. This is in contrast to the use of the total ion current at a given m/z, which may depend on the presence of interfering, structurally related compounds in the sample.

24.6.4 Conclusions for Case Study 1

Two instrumental approaches, GC- and MS-based, were used to successfully classify and differentiate odorous samples of similar chemical composition but different odor character. Because the samples were chosen to challenge instrumental capabilities to match the odor recognition abilities of humans, we conclude that the future is indeed bright for instrumentally based approaches to evaluate and mimic the perception of odors by humans.

24.7 Case Study 2: Use in Judging the Odor Quality of a Sunscreen Product

24.7.1 Background

Established practice in the industry requires the use of various analytical measurements to ensure the quality of every aspect of a perfumed product. On delivery of 24.7 Case Study 2: Use in Judging the Odor Quality of a Sunscreen Product 571

Fig. 24.3 HCA cluster dendrogram based on all 185 time points in the chromatograms of eight PRMs. A principal components analysis (PCA) of the 185 points produced 12 factors and explained 89 % of the variance. The first two principal components separate most of the PRMs, however, to separate orange from grapefruit we needed to go to principal components 4 through 6 572 24 Cosmetics and Fragrances

the product, specifications and measurements are attached as a written record. How- ever, as mentioned before, it is not uncommon for a perfume, PRM, or product to meet analytical specifications and fail sensory evaluation (or vice versa). Thus, sensory evaluation is an important additional quality control task, most often performed by an expert panel. Unfortunately, as discussed previously, development and maintenance of expert panels are costly and time consuming. Thus, we wanted to establish the use of an electronic nose as a tool to qualify the sensory properties of a product.

24.7.2 The Problem

We wanted to establish if an objective and sensitive electronic nose could free expert panels from tedious quality control activities, thus freeing them to judge more difficult samples. In our example, the best product would be judged to be ‘odorless’. The product samples would have already passed analytical tests prior to undergoing sensory evaluation. We used a Fox4000 electronic nose with 18 chemical sensors for correlation with sensory evaluations. Expert panel evaluations were made on 150 samples judged to fall in three categories: A: does not meet odor standard quality, but it is sufficiently good to be used as ‘diluent’ when adjusting bulk quality B: good (BON) odor quality, meets sensory standard M: rejected quality (MAUvais) To demonstrate the ability of the electronic nose Fox4000 to function in both research and development and production environments, two systems were evaluated. To function in both environments the electronic nose must be:

* As sensitive as the expert panel * Selective * Reproducible over time (short- and long-term, to allow the generation of databases) * Reproducible following sensor exchange or array replacement (to allow transferability of databases) * Robust, and simple to use and maintain

The following experiments were carried out to evaluate the performance of the electronic nose on a compound designed to serve as a sunscreen. The work was carried out over a six-month period in parallel with the standard quality control operating procedures at Givaudan Vernier. The initial work was carried out at the research facility located at Givaudan Du¨bendorf. 24.7 Case Study 2: Use in Judging the Odor Quality of a Sunscreen Product 573

Fig. 24.4 Alpha MOS Fox4000 electronic nose

24.7.3 Equipment and Methods

24.7.3.1 Equipment

Supplied and manufactured by Alpha MOS 2 Fox4000 electronic nose units (18 sensors), operated with zero-grade air. ACU500 humidifier, operated with HPLC-grade water. Fox4000 software. (Calibration methods) HS100 headspace autosampler.

Fox4000 EN The system used for this study at Givaudan R&D Du¨bendorf was a Fox4000 electronic nose with three metal oxide sensor chambers (18 sensors). The equipment is shown in Figure 24.4. All the chambers had their temperature controlled at 55 0.1 8C. The carrier gas was synthetic air (P ¼ 5 psi) and humidity was controlled by an ACU500 (RH ¼ 20 %, T ¼ 36 8C) using pure water. The samples were injected to the Fox by an autosampler from 10 ml sealed vials, the acquisition time and time between subsequent analyses were 120 s and 20 min, respectively, and the flow rate was kept at 300 mL min1. The second Fox4000 was used in a factory environment at Givaudan Vernier, using the database developed at the R&D facility in Du¨- bendorf.

Specific parameters for oil injections: * Headspace generation time: 20 min at 100 8C. * Injection volume: 2500 lL. * Volume of sample: 2 mL. 574 24 Cosmetics and Fragrances

Specific parameters for injection of standards:

* Headspace generation time: 2 min at 60 8C. * Injection volume: 100 lL. * Volume of sample: 1 mL.

Most of the standards used as calibration products were chosen from selected control sensory samples that were used as odor standards. Selected samples were also used as reference compounds.

24.7.4 Results

24.7.4.1 Sensory Correlation and Long Term Repeatability Analytical results are shown in Fig. 24.5. The PCA clustering of good (red) and rejected (blue) samples shows an excellent correlation with the expert panel judgments. Only three sensors (out of eighteen) were required to achieve those results demonstrating sufficient sensitivity and selectivity. Importantly, ten weeks later it was necessary to address calibration drift to interpret sample quality. This was achieved using a built-in calibration option available in the standard instrument software. This option allows the acquisition of data on new standard samples selected to track the drift and compensate for it. As a consequence, good results were obtained over a six month period when comparing electronic nose results with those obtained by standard sensory methods.

24.7.4.2 Database transfer from Du¨bendorf to Vernier Database transfer from Du¨bendorf to Vernier was carried out with help from Alpha MOS Toulouse. At the present stage of development, the successful use of the software

Fig. 24.5 (a) The PCA-clustering of good (red) and rejected (blue) samples shows an excellent correlation with the assessment of the sensory expert panel. Only three sensors are necessary to achieve this discrimination model. (b) The discriminant function model is capable identifying all unknowns 24.8 Conclusions 575 used to address calibration drift required help from Alpha MOS. While good results were obtained, the present methodology is not plug-and-play. New developments, aim- ing at addressing this difficult problem, are in progress.

24.7.5 Conclusions for Case Study 2

The six-month evaluation of the electronic nose in quality control (Vernier facility), in parallel with standard sensory evaluations by an expert panel, demonstrated the ability of the Fox4000 electronic nose to carry out sensory analyses. Over the study period, the system accurately classified ‘good’ and ‘bad’ batches of the tested product. Although this was a remarkable result, further improvements would have to be made to justify replacing current practice. Some of the improvements include a reduction of capital investment for the plant and a simpler software calibration option (i.e. a ‘plug-and-play’ software) as well as a significant reduction in the required measurement time. Since this work was performed, a number of improvements have been made available by the manufacturer. The improvements include faster sample throughput (5 min), and a significant reduction in the level of expertise and labor required to run the instruments. Finally, there is a ’plug-and-play’ database transferability between units.

24.8 Conclusions

The exquisite sensitivity and selectivity humans exhibit towards ‘key’ components of perfumes presents a challenging problem when attempting to predict human perception based on data derived from instrumental measurements. Ideally, to predict the human response to perfumes our instruments would need to approach the sensitivity and selectivity exhibited by humans. However, while state-of-the-art electronic noses may differ from humans in both selectivity and sensitivity, they can be trained to perform the function of a highly skilled sensory panel. Furthermore, there may only be a few hundred ’key’ compounds we would need to measure to obtain near-perfect correlations with the human response to virtually all perfumes. The number would drop to less than fifty within any given perfume family. Those key compounds, and their respective concentration, could be measured by high- resolution techniques such as GC-FID. Alternatively, markers of those compounds could well serve the purpose. This could be done, without prior separation, by monitoring key ions with a mass spectrometer or by the use of quasi-selective sensors. General-purpose instruments and sensors would work in cases where the bulk gas- phase composition is determined by key compounds or marker compounds. This is often the case in samples expected to have little or no odor, such as bases for cosmetics and raw materials used in the industry, e.g. plastics. A number of studies, reflecting the status of the field, are listed under references. 576 24 Cosmetics and Fragrances

24.9 Future Directions

The sensitivity of analytical instruments has increased dramatically over time. Thus, analytical detection limits reported using electronic-based instruments have dropped by roughly 1000-fold per decade starting in the 1970s. In that decade, concentration units (or mass) reported in the literature, and instrumental specifications were in parts-per-million (or micrograms). The literature and specifications changed to parts-per-billion (or nanograms) in the 1980s, and parts-per-trillion (or picograms) in the 1990s. The increased sensitivity is, to a large extent, a consequence of developments in the semiconductor and computer industries and their application to analytical chemistry. This trend is likely to continue in the foreseeable future. Importantly, those advances in sensitivity often translated into advances in selectivity. Today, the selectivity of state-of-the-art GC-MS instrumentation equipped with large-volume-injection systems can be used to identify and measure hundreds of perfume compounds present in the gas-phase at, or above 1 ppb (v/v). This capability can be used to identify ‘key’ compounds in perfumes and should facilitate the development of new, highly sensitive quasi-selective sensors. Advances in solid-state chemistry and ionization mechanisms, as well as advances in microfabrication techniques are likely to produce large detector arrays with enhanced sensitivity and selectivity. Those advances, coupled with the low power requirements of small arrays should produce portable electronic noses with capabilities comparable to those of humans.›

References

1 R. R. Calkin, J. S. Jellinek. ‘Perfumery 6 M. P. Byfield, M. Lindstrom, L. F. Wunsche. practice and principles’, Wiley & Sons, 1994. Chiral discrimination using a quartz crystal 2 A. Jinks, D. Laing. Perception 28: 395–404 microbalance and comparison with gas 1999. chromatographic retention data, Chirality 3 L. Buck, R. Axel. A novel multigene family 1997. may encode odor recognition: a molecular 7 Z. Cao, H. G. Lin, B. F. Wang, D. Xu, R. Q. basis for odor recognition, Cell 65: 175 1991. Yu. A perfume odor-sensing system using 4 B. Kusumoputro, M. Rivai. ‘Discrimination an array of piezoelectric crystal sensors with of fragrance odor by arrayed quartz resona- plasticized PVC coatings, Fresenius Journal of tor and a neural network’. Proceedings of Analytical Chemistry 355 (2): 194–199 1996. International Conference on Computational 8 K. Yokoyama, F. Ebisawa. Detection and Intelligence and Multimedia Applications evaluation of fragrances by human reactions (Eds. H. Selvaraj, B. Verma), Gippsland, using a chemical sensor based adsorbate Victoria, Australia, 1998, pp.264–269. detection, Analytical Chemistry 65 (6): 673– 5 M. P. Byfield, L. Wunsche, C. R. Vuil- 677 1993. leumier. ‘Development and applications of 9 Hyung-Ki-Hong, Hyun-Woo-Shin, Dong- an electronic nose based on arrays of pie- Hyun-Yun, Seung-Ryeol-Kim, Chul-Han- zoelectric sensors’. Proceedings of the Se- Kwon, Kyuchung-Lee, T. Moriizumi-T. venth Conference on Sensors and their Electronic nose system with micro gas sen- Applications. (Ed. A.T. Augousti) Institute of sor array, Sensors and Actuators B (Chemical) Physics Publishing, Bristol, UK, 1995, 36 (1–3): 338–341 1996. pp.52–57. 24.9 Future Directions 577

10 M. Penza, G. Cassano, F. Tortorella, G. 13 R. Feldhoff, C. A. Saby, P. Bernadet. De- Zaccaria. Classification of food, beverages tection of perfumes in diesel fuels with se- and perfumes by WO thin-film sensors array miconductor and mass spectrometry-based and pattern recognition techniques, Sensors electronic noses, Flavour and Fragrance and Actuators B (Chemical) 73 (1): 76–87 Journal 15 (4): 215–222 2000. 2001. 14 (a) N. Neuner-Jehle, F. Etzweiler. in ‘Per- 11 S. E. Letant, S. Content, Tze-Tsung-Tan, F. fumes art, science and technology’, (Eds. P. Zenhausern, M. J. Sailor. Integration of M. Mu¨ller, D. Lamparsky), Elsevier, London, porous silicon chips in an electronic artificial New York, 1991, p.153. Updated in: (b) H. nose, Sensors and Actuators B (Chemical) 69 Gygax, H. Koch, Chimia 55 (5): 401 2001. (1–2): 193–198 2000. 12 A. Carrasco, C. Saby, P. Bernadet. Discri- mination of Yves Saint Laurent perfumes by an electronic nose, Flavour and Fragrance Journal 13 (5): 335–348 1998. Chapter 27: Automotive and Aerospace Applications

M. A. Ryan and Hanying Zhou Jet Propulsion Laboratory, California Institute of Technology Pasadena CA 91109

INTRODUCTION

By the nature of their trainability to a broad range of compounds, electronic noses are a good choice for air quality monitoring in an environment where the possible contaminants are known. The trainability of an electronic nose, along with the ability to select sensors for response to a suite of compounds has made this type of device useful in several applications; in this chapter we will discuss its application to monitoring the breathing air in an enclosed space for the presence of hazardous compounds. The application of an electronic nose as an air quality monitor is as an event monitor, where events of low concentration which do not present a hazard are not reported, but events of concentration approaching a hazardous level are reported so remedial action can be taken. The electronic nose used in these applications is not an analytical device which analyzes the air for all compounds present, but neither is it an alarm which sounds at the presence of any change in the atmosphere. The device described here was used as an air quality monitor in an experiment aboard NASA’s Space Shuttle Flight STS-95, and was designed to fill the gap between an alarm with no ability to distinguish among compounds and an analytical instrument.

AUTOMOTIVE APPLICATIONS Use of an electronic nose in the automotive industry is primarily conceptual today,

but there are several areas in which such a device can be used. These include

monitoring the exhaust for combustion efficiency, monitoring the cabin air for passenger

safety, and monitoring the engine compartment for other conditions such as leaking oil

or other fluids. Owing to offgassing of fabrics and materials (“new car smell”), to leaks

of coolant from the air-conditioning system, and intake of air from the roadway and the

engine compartment, the passenger cabin of an automobile can be significantly more

hazardous to human health than the outside air [Chan, Leung]. Improvement of the air

quality in an automobile cabin can be accomplished rather simply, but as cabins will

remain well-sealed for climate control and energy conservation, a need to monitor the

interior will remain. As environmental concerns spur development of more efficient

combustion, it will be useful to monitor the exhaust for combustion products as well.

Several automobile manufacturers have discussed the possibility of using an ENose in

a system in which the exhaust is monitored for the presence of compounds indicative of

incomplete combustion, and feedback to the engine will adjust engine settings to

improve combustion efficiency.

AEROSPACE APPLICATIONS

Electronic Noses have been proposed for many applications in aerospace; some

of those applications are realistic within the limits today‘s technology, and some will require more development. In the area of space exploration, electronic noses have been proposed for planetary atmospheric studies on landers. This application varies from addition of an electronic nose to a rover to study the atmosphere as the rover moves, to stationary devices which will study the variations in atmosphere over days or seasons. In the search for evidence of life on other planets, electronic noses have been proposed as desirable sensors because the sensing media in the array can be selected to make it possible to distinguish isomers and enantiomers [Ryan and Lewis,

2001], and because the sensor array can be configured to span a broad range of compounds. These applications require development of methods which will allow the electronic nose to deconvolute target vapors from an unknown background; work to develop devices with these capabilities is underway at the Jet Propulsion Laboratory

(JPL).

An immediate, and perhaps the most important, application is monitoring air quality in human habitats. The ability to monitor the recycled breathing air in a closed chamber is important to NASA for use in enclosed environments such as the crew quarters in the Space Shuttle and the International Space Station (ISS). Today, air quality in the Space Shuttle is generally determined anecdotally by crew members’ reports, and is determined after flight by collecting an end-of-mission sample and analyzing it in an analytical laboratory using gas chromatography-mass spectrometry

(GC-MS). The availability of a miniature, low power instrument capable of identifying contaminants in the breathing environment at part-per-million (ppm) and sub-ppm levels would enhance the capability to monitor the quality of recycled air and thus to protect crew health. Such an instrument is envisioned for use as an incident monitor, to notify the crew of the presence of potentially dangerous substances from spills and leaks and to provide early warning of heating in electrical components which could lead to a fire.

In addition to notification of events, it is necessary to have a reliable method by which judgements on the use of breathing apparatus can be made; if the crew has put on breathing apparatus while repairing a leak or cleaning a spill, it is necessary to know whether it is safe to remove the apparatus. These needs have led to the development of an electronic nose at JPL [Ryan et al., 1997, 1998, 2000], with ultimate application to

ISS intended and experiments on the Space Shuttle in the near-term.

The qualities required for an incident monitor to be used in spacecraft are that it be capable of identifying and quantifying target compounds at determined levels in a fairly wide range (see Table I), that it be a low mass and volume device which uses low power, and that it require little crew time for maintenance, calibration and air analysis.

There are several possible sensing devices which could be used in the Space Shuttle or ISS, but all have limitations in terms of the requirements. These devices include GC-

MS, Volatile Organic Carbon Analyzer, Flame Ionization Detectors, and Smoke Alarms.

Of these, only GC-MS discriminates among compounds; it also has the greatest sensitivity. However, it generally requires crew time in sample preparation, maintenance and calibration. An electronic nose does not, in general, have the sensitivity of GC-MS; however, for most target compounds ppm and sub-ppm sensitivity is required, but not the parts per trillion level found with GC-MS.

An electronic nose meets the requirements for an incident monitor. It can identify and quantify compounds in its target set with a dynamic range of about 0.01 to 10,000 ppm, depending on the compound, it lends itself to miniaturization, and because it measures deviation from a background it does not require frequent calibration and maintenance. The electronic nose developed at JPL was designed to detect a suite of

compounds and is suitable for use in the crew habitat of a spacecraft. The habitat is an

enclosed space where air is recycled and where it is unlikely that unknown and

unexpected vapors will be released into the air. It can be assumed that the air is clean at the beginning of a period of enclosure, and it is deviations from that state that the electronic nose will monitor; thus, it is not necessary to have detailed knowledge of the constituents of the air to start. In addition, the contaminants which are likely to be present and for which it is important to monitor are well known, the number of compounds is not large (50 or so), and the probability of mixtures of 5 or more such compounds appearing at one time is small. It is possible, then, to design and train a device to monitor the air for deviation from a clean baseline and to analyze those deviations for the appearance of a set of target compounds.

The air quality conditions in the crew quarters of a spacecraft are not radically

different from the conditions in an aircraft cabin, or in the passenger cabin of a bus or

automobile. In all those cases, it is reasonable to assume the air is clean at the

beginning of a monitoring period, and there is a set of contaminants of concern for

which to monitor. With such conditions in mind, the JPL electronic nose was designed

for a flight experiment where the crew habitat in the Space Shuttle was monitored

continuously for six days.

The JPL ENose is a low power, miniature device which, in its current

experimental design, has the capability to distinguish among, identify and quantify 10 common contaminants which may be present as a spill or leak in the recirculated

breathing air of the space shuttle or space station. It has as its basis an array of conductometric chemical sensors made from polymer/carbon composite sensing films developed at Caltech [Freund; Lonergan]. It is an array of 32 sensors, coated with 16 polymers/carbon composites. The polymers were selected by analyzing polymer responses to the target compounds and selecting those which gave the most distinct fingerprints for the target analytes. The JPL development model was used in a flight experiment on the Space Shuttle flight STS-95 (October-November 1998) to determine whether it could be used as a continuous air quality monitor. A block diagram and photo of the JPL ENose are shown in Figure 27.1. The device used in the flight experiment has a volume of 2000 mL and a mass of 1.4 kg including the HP200 LX computer used for control and data acquisition, and uses 1.5 W average power. The mass and volume were determined primarily by the spaceflight-qualified container required for the device to be used in an experiment; the volume and mass can be reduced by a factor of 4 with no modifications to the sensor head or the electronics and minor modifications to the pneumatic system.

POLYMER COMPOSITE FILMS

The polymer-carbon composite films developed at Caltech are the sensing media used in the JPL ENose [Freund; Lonergan; Severin; Albert]. These films are made from insulating polymers loaded with a conductive medium such as carbon to make resistive films. When a polymer film is exposed to a vapor, some of the vapor partitions into the film and causes the film to swell; the degree of swelling is proportional to the change in resistance in the film because the swelling decreases the number of connected pathways of the conducting component of the composite material [Freund]. The electrical resistance of each sensor is then and the response of each sensor in the array is expressed as the change in resistance, dR.

Using commercially available organic insulating polymers as the basis for

conductometric sensing films allows ready incorporation of broad chemical diversity into

the sensing array. The sensors respond differently to different vapors, based on the

differences in such properties as polarizability, dipolarity, basicity or acidity, and

molecular size of the polymer and the vapor.

The polymer-carbon composite sensing films are sensitive to temperature and

pressure change as well as to change in the composition of the atmosphere. In a

measuring mode where the device is sniffing the atmosphere and comparing it to a

clean background with measurements of each a few minutes apart, temperature

changes are generally not significant. However, in the case of continuous monitoring

over several hours or days, both temperature and pressure changes will influence the

location of the baseline, and it is necessary to distinguish among temperature and/or

pressure change, slow buildup of compounds, and baseline drift. All of these issues

were addressed in the device developed at JPL. Neither changes in pressure nor

humidity which might be found in normal habitat have a significant effect on the

differential sensor response, but temperature changes greater than 4 – 8 oC influence

the magnitude of response across the sensing array as well as the fingerprint of

individual analytes. While it is possible to measure temperature, pressure and humidity

and to subtract any effect of changes in these conditions from the sensor response

data, the JPL ENose was built with the capability to control temperature, and pressure

and humidity were measured separately. Temperature was controlled on the sensor substrates to stay constant at 28, 32 or 34 oC, both to eliminate apparent baseline drift

(film resistance changes) caused by temperature change and to aid the sensing process. Temperatures around 30 oC will assist the process of desorption of analytes

from the films and will prevent hydrogen bonds from forming between analytes and the

polymers.

ENOSE OPERATION IN SPACECRAFT

While it is reasonable to assume clean air at the beginning of an enclosed period

in the Space Shuttle, there are two scenarios in which a clean air baseline must be

established. In one scenario, the ENose might be used to determine whether it is safe

to enter a chamber that has been enclosed for some time without crew use, such as a

module in ISS. In the other scenario, a background of clean air must be established to

determine whether there has been a slow buildup of a contaminant. This second

scenario, slow buildup of a compound, is among the likely scenarios for contamination

of the air. Contaminants may build up slowly as offgassing, slow leaks in vapor and

liquid containers, from inadequate air revitalization or filter breakthrough, and as human

metabolic products such as methane or carbon dioxide. In both of these scenarios, a

system by which a baseline of clean air can be established is necessary.

Contamination from offgassing may be considered of minor importance for

aircraft or automobile cabins because the air is exchanged frequently in the course of

use and fresh air can be brought inside during use, but in cabins where air is not

exchanged for several hours, the buildup can be considerable. Often the offgassed

molecules are small, such as formaldehyde, and are not well scrubbed in the air revitalization system. In the Space Shuttle where air might not be exchanged for several days or, more importantly in ISS, where the air is not exchanged, offgassing becomes an important consideration. Flight qualification includes establishment that the offgassing rate of components be below a set level, but there are as yet no data for offgassing over periods of months to years, as will be found on ISS.

The JPL ENose pneumatic system includes a diaphragm pump which pulls atmosphere over the sensors at 0.25 L/min and two filters, an activated charcoal filter and a filter of inert material, before the sample chamber. The atmosphere to be analyzed travels through a filter that is selected by a solenoid valve which switches between the two. During usual monitoring intervals, the air travels through the “dummy” filter made of inert material to provide a pressure drop equivalent to the pressure drop across the charcoal filter. The charcoal filter cleans air without removing humidity, and a baseline of cleaned air can be constructed and used to determine the degree of baseline drift. The constructed baseline allows the analysis program to distinguish between drift and slow change in atmosphere. Figure 27.2 shows how drift and slow buildup can be distinguished after the charcoal filter is switched off; the sensor films respond by rising rapidly and creating a “virtual peak,” and the sensor responses can then be analyzed against the cleaned air background. The analysis of the responses of the sensing array can then be used to determine whether the slow change in the atmosphere is caused by contamination.

For the flight experiment, 6 days of continuous operation, the charcoal filter was switched on for 20 minutes out of every 210 minutes. This frequency was sufficient to determine the baseline in this application. If an electronic nose is to be used to determine whether a chamber is safe to enter after a closed period, the cleaned air baseline must be established for several minutes, and the virtual peak analyzed when the charcoal filter is turned off. A schedule for filter changeout must be established; for

Space Shuttle air and no events, changing the filter every 20-30 days is sufficient. If there has been an incident found by the filter, it should be changed after the cause of the incident has been fixed.

In other applications, where the pressure and temperature are changing rapidly, or where the composition of the atmosphere changes frequently, the filters can be programmed to switch at different frequencies. In the passenger cabin of an aircraft, for example, filtering can be frequent during the loading and taxi stages, when the concentration of combustion products and of fuel can be high, and less frequent during cruise.

The responses of the ENose were not influenced significantly by meals or activities in the crew quarters because the device was placed under the air intake vent for the entire cabin; odors were significantly diluted when they reached the sensors.

This condition was chosen in order to monitor the average concentration in the cabin rather than localized concentrations.

THE JPL ENOSE FLIGHT EXPERIMENT

For the application of adverse event monitoring in the Space Shuttle, the JPL

ENose was trained to 12 compounds; 10 of these were compounds likely to leak or spill and the other two were humidity change and vapor from a medical swab (2-propanol and water), which was used daily to confirm that the device was operating. The ENose was trained to identify and quantify the 10 contaminating compounds at the 1 hour

Spacecraft Maximum Allowable Concentration (SMAC) levels which are shown in the upper section of Table I.

The 10 contaminants were drawn from a list of compounds of concern and for which air samples are tested after a shuttle flight. In the second generation device now under development, there will be 10-12 additional compounds. The sensitivity required for the device was set at the 1 hour SMAC in the Flight Experiment and is set at the 24 hour SMAC for the second generation device. The upper section of Table I shows the

24 hour SMAC and the lowest level detected reliably by the first generation ENose at

JPL. The lower section of Table I shows a list of compounds considered for the second set and their 24 hour SMACs. As an event monitor, it is not necessary to be significantly more sensitive than the 24 hour SMAC level; when the concentration of a contaminant approaches ~35% of the SMAC, measures can be taken to remove the compound from the air and to take action on the source of the contamination. Further training of the software is possible in situ, but for accurate identification and quantification, the training must be done in an environment where it is possible to deliver precise concentrations of the compound in the range of interest. Table I Upper Section: Compounds targeted in the first generation ENose, with their 1-hour and 24-hour SMACs, and the lower level detected at JPL with that device. Lower Section: compounds considered for the second generation ENose, with their 24 hour SMACs.

Compound SMAC 1hr SMAC 24 hr Detected at JPL (ppm) [**] (ppm) [**] (ppm)

alcohols methanol 30 10 5 ethanol 2000 500 50 2-propanol 400 100 50 methane 5300 5300 3000 ammonia 30 20 20 benzene 10 3 10 formaldehyde 0.4 0.1 10 Freon 113 50 50 20 indole 1 0.3 0.03 toluene 16 16 15

acetaldehyde 6 acetone 270 acetonitrile 4 2-butanone 150 chlorobenzene 10 dichloromethane 35 furan 0.1 hexamethyltricyclosilane 25 hydrazine 0.3 methyl hydrazine 0.002 tetrahydrofuran 40 1,1,1-trichloroethane 11 o,p-xylenes 100

** source: Spacecraft Maximum Allowable Concentrations for Selected Airborne Contaminants; Space Physiology and Medicine

For all cases except formaldehyde, the ENose is able to detect the compound at or below the 1 hour SMAC. The sensitivity limit for formaldehyde in the flight experiment device is 10 ppm; by selection of a different polymer set with polymers more likely to sorb formaldehyde, it should be possible to detect that compound below the 24- hour SMAC level. The ENose is also able to deconvolute signals to identify and quantify mixtures of two compounds with moderate success (about 60%). It is expected that with further training and a more selective group of polymers, it will be possible to detect lower concentrations of compounds and to deconvolute mixtures of three or four compounds.

DATA ANALYSIS

The data analysis software development portion of the JPL ENose flight experiment considered several different approaches. The primary constraint in software development was the requirement that gas events of single or mixed gases from the 10 target compounds be identified correctly and quantified accurately. The co-investigator in the flight experiment, Dr. John James of the Toxicology Branch at NASA-Johnson Space

Center (JSC), defined accurate quantification as +/- 50% of the known concentration measured in the laboratory. This degree of error was defined based on the SMACs; the toxic level of most of the compounds is not known more accurately than +/- 50%, so the

SMACs have been set at the lower end. For the flight experiment, constraints in telemetry and communication prevented real-time analysis, and so the development process did not include full capability for immediate resistance vs. time data analysis.

A series of software routines was developed using MATLAB (from MathWorks,

Inc.) as a programming tool. MATLAB is a flexible program, and thus appealing for development of software, though it runs relatively slowly. For future use, where real- time or quasi-real time analysis is called for, the routines can be translated into C and run on a desk top or lap-top computer.

For sensing media such as the conducting polymer/carbon films used in this program, relative response changes (in magnitude) have been found to be more reliable than the response shapes, especially at the low gas concentration range targeted in this program (1–100 ppm). Hence, the task of identifying and quantifying a gas event is roughly a two-step procedure:

1) Data pre-processing, to extract the response pattern of a gas event from raw

time-series resistance data for subsequent analysis, and

2) Pattern recognition, to identify and quantify a gas event based on the response

pattern extracted.

Data Pre-processing

When presented with continuous monitoring data, a response pattern must be extracted by use of software. This process of extracting a response pattern from raw resistance data involves four sequential steps: 1) Noise removal, 2) Baseline drift accommodation, 3) Gas event occurrence determination, and 4) Resistance change calculation.

Noise removal Despite the best effort in choosing sensor films with the consideration of low noise level, fluctuation in the sensors’ responses are still seen to be quite large. Some polymer films were found to be noisier than others. The reasons one polymer-carbon composite film might be noisier than another are not well understood; noise may be attributed to high sensitivity of the polymer film to small changes in pressure caused by air flow, to differences in the carbon dispersion in the film, or to inhomogeneities in the thickness or even composition of the film itself. In general, the fluctuation in resistance (or noise) is fast compared to the response to a gas event. Therefore digital filtering may be used to filter out this high frequency fluctuation. The length of the filter may be different for different sensors and can be determined by analysis of the noise in each sensor.

Baseline drift accommodation Baseline drift is one of the most difficult problems to be solved in extracting ENose resistance data from the time data. The causes for baseline drift can be multiple, and include variations in temperature, humidity, pressure, aging of the sensors, and sensor saturation. However, at present there is no clear understanding of the underlying mechanism of each one of the causes, which makes attempts to compensate drift very difficult. Nevertheless, the baseline drift is generally slowly-varying in nature compared to the response time of a detectable gas event. This difference in time scale enables us to use a long-length digital filter to determine the approximate baseline drift and then subtract it from the raw data. The result is further adjusted by piecewise fitting using the baseline information from the clean air reference cycles described above. Although this approach will not accommodate the drift fully, it will reduce the effect to a manageable degree. Figure

27.3 shows resistance data which has been processed. The dark, smooth trace in the upper plot shows the baseline variation determined through the use of low frequecy filters. The grey, noisy trace in the lower plot is the data after baseline variation has been subtracted, and the dark line is the processed data, with baseline variation subtracted and after filtering for noise accomodation. Gas event occurrence determination Because data analysis in the flight

experiment of the JPL ENose was not real-time because of constraints unrelated to the

technology development, it was not necessary for the analysis to be automatic, but a

preliminary software routine for automated determination of whether and when a gas

event occurs was developed. It is based primarily on threshold calculation, in which the

resistance change over a certain time interval is calculated, and a time-stamp is

registered if the change exceeds a pre-set threshold. This routine can detect most gas

events; however, it was also found that it may identify noise, and sometimes baseline

drift, as gas events. For the flight experiment, events identified by the automated

routine were confirmed by visual inspection of the time domain data; future

development of the data analysis software will refine the identification method.

Resistance change calculation Since the sensors’ relative responsiveness to

a vapor determines the fingerprint of that gas, the response pattern, it is important to

preserve this relative responsiveness. This means any calculation method of the

resistance change should be taken at the same time stamp after the initial onset of a

gas. Both relative resistance change, R/Ro, and fractional resistance change, (R-Ro)/Ro

were tested, and the latter was adopted as it maximizes the difference between the

signatures of different gas compounds.

Pattern Recognition Method

Although many pattern analysis methods exist in the general field of electronic nose and other array-based sensor data analysis [Bartlett, 1999; also see Chapter 8 of

Part B, and Part C], no single method appears to be readily applicable to the task of identifying and quantifying single gases as well as mixtures of up to three of the 12 compounds (10 target compound plus water, humidity change and the propanol-wipe) at levels about 1 –100 ppm. Most of the widely used methods have demonstrated their effectiveness, but not to a combination of all three scenarios found here: a large number of target compounds, some of which are of very similar chemical structure

(e.g., ethanol and methanol), low target concentrations, and both single gases and mixtures.

METHOD(S) DEVELOPMENT

For reasons stated above, three parallel approaches to ENose data analysis were used during the early stages of software development: Discriminant Function

Analysis (DFA), Neural Networks with Back Propagation (NNBP), and Linear Algebra

(LA). Principal Component Analysis (PCA) was initially used , but was later replaced by

DFA because DFA tends to do better at discriminating similar signatures that contain subtle, but possibly crucial, gas-discriminatory information. DFA is also better in class labeling than PCA.

NNBP, or more specifically, multilayer perceptron (MLP), was selected as an approach because it has good generalization of functions to cases outside the training set, is capable of finding a best-fit function (linear or nonlinear; no models needed), and is also more suitable than DFA when the sensor signatures of two gases are not separable by a hyperplane (e.g. one gas has a signature surrounding the signatures of another gas). However, NNBP is inferior to DFA in classifying data sets which may overlap. The reason to use LA, which is not as commonly used as other methods, is that neither DFA nor NNBP were found to be well suited to recognizing the sensor signatures from combinations of more than one gas. This method tries to solve the equation x=Ac, where vector x is an observation (a response pattern), vector c is the cause for the observation (concentrations of a gas or combinations of gases), and matrix A describes system characteristics (gas signatures obtained from training data, or sensitivity coefficients). For ENose data analysis where the response pattern can be noise corrupted, which means there may exist no exact solution, least squares fitting is the preferred way to solve the equation [Stang; Lawson].

The idea of developing three parallel methods is that one can first use the LA method to deconvolute an unknown response pattern as a linear combination of target compounds; unknown compounds are expressed as a combination of up to four compounds. If a single compound is found, additional verification can be then carried out by NNBP and DFA methods for increased success rate and accuracy. However we have found the LA method to perform consistently best among the three methods even for single gases, while DFA was consistently the worst, which prompted us to discard the use of the two verification methods of NNBP and DFA in the process.

Linear Algebra is suitable only if the training data are linear, which is not the case for all sensors at the concentration ranges considered (see Table II) . For a nonlinear scenario, it is then reasonable to use some Nonlinear Least Squares fitting methods such as that of Levenberg and Marquart (LM-NLS). This is the one of the two new methods which were investigated for non-linear analysis. The other method, a

Differential Evolution (DE) approach, was also investigated because it promises fast optimization (the LM-NLS method can be rather slow). DE represents some recently emerged so-called genetic algorithms [Storn; also see Chapter 15 of Part C]. It is a parallel direct search optimization tool. It begins with an initial randomly-chosen population of parameter vectors, adding random vector differentials to the best-so-far solution in order to perturb it. A one-way crossover operation then replaces parameters in the targeted population vector with some (or all) of the parameter values from this

“noisy” best-so-far vector. In essence it imitates the principles of genetics and natural evolution by operating on a population of possible solutions using so-called genetic operators, recombination, inversion, mutation and selection. Various paths to the optimum solution are checked and information about them can be exchanged. The concept is simple, the convergence is fast and the required human interface is minimal: no more than three factors need be selected for a specific application. However the last advantage is also its disadvantage: limited control for ENose data analysis. Finally, the LM-NLS method was selected as the best tool for ENose data analysis.

Levenberg-Marquart Nonlinear Least Squares Method

For nonlinear models the technique of choice for least-squares fitting is the iterative damped least-square method of Levenberg and Marquart (LM-NLS). Similar to

LA, LM-NLS tries to find the best-fit parameter vector c from an observation vector x, which is related to c through a known linear or nonlinear function, x=f(A,c), where A is system characteristics (sensitivity coefficients) obtained from training data. This method usually begins from a given starting point of c, calculates the discrepancy of the fit: residual =(computed-observed)/,

where is the standard deviation, and updates with a better-fitted parameter c at each

step. LM-NLS automatically adjusts the parameter step to assure a reduction in the

residual: increase damping (reduce step) for a highly nonlinear problem, decrease

damping (increase step) for a linear problem. Because of this ability to adjust damping,

LM-NLS is adaptive to both linear and nonlinear problems. How this method adjusts

damping is discussed in detail elsewhere [Lampton].

In the course of this work, it was found that the response of the films to the target compounds is linear with concentration only within a limited range. The nonlinearities in the training data generated are of low order, but successful identification and quantification of gas events must take the nonlinearities into account. To obtain sensor characteristics without further knowledge of sensor nonlinearities, a second order polynomial fit was used to model the nonlinearities. For each sensor response to each gas, the program finds the best-fit sensitivity coefficients A1 and A2 (in the least-squares

sense) to the following equation:

2 resistance change = A1c + A2c

where c is gas concentration vector. The fit is constrained to pass through the origin.

A1 and A2 are 13x32 matrices characterizing the sensors’ response to ten targeted

gases plus water, humidity change, and the propanol wipe.

Several modifications were made to the standard LM-NLS method to suit the

ENose data analysis problem. First, sets of starting points of vector c were used

instead of a single set of starting points of vector c. The purpose of doing this is to

avoid a local residual minimum, which is common in many optimization algorithms, including the LM-NLS method. These initial sets of vector c can be randomly assigned from within each element’s allowed range. The total number of initial sets will be determined by the speed desired and the complexity of local minimum problem. In our case, about 200 initial sets were found (~15N, where N=13 is the number of target compounds) to be a good compromise.

Second, instead of always updating c for a smaller residual, we modify the update strategy to favor a smaller number of gases within certain ambiguity ranges of the residual. The reason is that signature patterns for a given gas compound generated by the ENose sensors have been observed to have large variations. The simple updating strategy tends to minimize the residual with a more-than-reasonable- large-number combination of gas when the residual is simply the variation in recorded response pattern itself and should be ignored. The amount of the final residual is an indicator of how large the fitting error is and the confidence level of the fitting.

Finally, the sensors’ response pattern was weighted to maximize the difference between similar signatures. As seen in Figure 27.4, which shows representative signatures of the ten target gas compounds plus the medical wipe at a median concentration level (because of the nonlinearity, there is no single signature for one gas at all concentrations), it is clear that ethanol and methanol have very similar signature patterns. Regression analysis also pointed out linear dependency to certain degrees.

This means that the signature pattern of one gas could be expressed as a linear combination of the response pattern generated by some other target gases. To reduce this similarity, the sensors’ raw resistance responses must be modified by different weights in the data analysis procedure.

Single gases

For lab-controlled gas events, the overall success rate reaches ~85% for targeted singles. Broken down into individual singles, the success are listed below in Table II.

The concentration ranges used in the training sets for each single gas are also given.

Table II Identification and quantification success rates for single gases. The ranges shown here are ranges used in LMNLS analysis.

Compound Concentration Success Range (ppm) Rate (%)

Ammonia 10 – 50 100 Benzene 20 – 150 88 Ethanol 10 – 130 87 Freon 113 50 – 525 80 Formaldehyde 50 – 510 100 Indole .006 – 0.06 80 Methane 3000 – 7000 75 Methanol 10 – 300 65 Propanol 75 – 180 80 Toluene 30 – 60 50 %Relative Humidity 5 – 65 100 Medical Wipe 500 – 4000 100

Considering that the raw data are often very noisy at low concentrations, nonlinear at high concentrations, highly correlated in some cases, and weakly additive in some mixtures, these results demonstrate that the LM-NLS method is an effective technique for analysis of an array of sensors. Future work on the ENose will attempt to remove many of the impediments to data analysis, with focus on noise and correlation. Correlation will be addressed in polymer film selection. The ability of the data analysis software to identify and quantify single and multiple gas events in clean air was tested in the laboratory. The targeted concentrations range for quantification was 30% to 300% of the one hour SMAC for each compound. As can be seen from Table II, in some cases it was possible to identify and quantify substantially below the 30% SMAC concentration; however, in a few cases quantification was successful only as low as 100% of the one-hour SMAC.

In one case, formaldehyde, we were unable to identify and quantify reliably below several times the one-hour SMAC. Figures 27.5 and 27.6 show some results of single gas identification and quantification graphically.

Mixed gases

Identification and quantification of mixtures in clean air was moderately successful. Additive linearity holds for some combinations in concentration ranges near the SMAC level of the lower SMAC-compound. The success rate for double gases

(about 60%) was less than that of single gases, as would be expected. An exhaustive set of gas pairs was not run because of time constraints; only a selected group of mixture pairs were run to test the additivity. For this relatively small pool of data, additivity holds for the following gas combinations:

methanol + toluene ammonia + benzene ethanol + formaldehyde

methanol + benzene ammonia + ethanol propanol + benzene.

Although data obtained on some other combinations of gas compounds, e.g.,

{benzene + formaldehyde} and {methanol + propanol}, did not validate their additivity in

these tests, this does not necessary mean the additivity does not hold for those gas

combinations. In fact, in many of those gas combination tests, often one of the gases

was run at a very low concentration and its response was overwhelmed by the other

gas’s strong response. In other words, the detectable concentration of a gas might be

higher if there exist other highly responsive gases.

STS-95 Flight Data Analysis Results

The Resistance vs. Time data that were returned from STS-95 showed that there were several gas events in addition to the daily marker. The daily marker, exposure to a propanol and water medical wipe, was added to the experiment so that operation of the device over the entire period could be confirmed. The initial analysis selected the daily markers and identified them as 2-propanol plus a humidity change. These identifications were confirmed by comparison of crew log times with the time of the event in the data.

While the hope in an experiment such as this one is that there will be several events which test the ability of the device, such events would certainly be anomalous events in the space shuttle environment. Software analysis identifies all events which were not propanol wipe events as humidity changes. Most of those changes can be well correlated with the humidity changes recorded by the independent humidity measurements provided to JPL by JSC. The events are not completely correlated in time because the humidity sensor was located on the stairway between the mid-deck and the flight deck, and the ENose was located in the mid-deck locker area near the air revitalization system intake. Those events identified as humidity changes but not correlated with cabin humidity change are likely to be caused by local humidity changes; that is, changes in humidity near the ENose which were not sufficient to cause a measurable change in cabin humidity.

Figure 27.7 shows the correlation of cabin humidity with ENose response in several cases. There are visible dips in the traces at times 19:00, 20:52, and 0:07 CST,

Nov 2 -3, 1998. These dips are the changes in air composition, and thus resistance, during the baselining cycle, when air is directed through the charcoal filter. Piecewise baseline fitting is based on the resistance during the baselining cycle.

Software analysis of the flight data did not identify any other target compounds, as single gases or as mixtures. The independent analysis of collected air samples, in which the samples were analyzed at JSC by GC-MS, confirmed that no target compounds were found in the daily It is not surprising that the only changes the ENose saw were humidity changes, and it is because events were not expected that the experiment included the relatively uncontrolled daily marker events. air samples in concentrations above the

ENose detection threshold. There were no compounds that the ENose would have indicated as unidentified events present in the air samples.

FUTURE DIRECTIONS

Sensors

The number of sensors in the Second Generation ENose will remain at 32. The number of polymers may be expanded beyond 16 in order to make sub-groups of polymers which have been selected for response to particular classes of compounds within the set of 32 sensors.

To determine the set and sub-groups of polymers for the set of some 20 target compounds, a model of polymer-analyte interaction is under development. This model takes account of such parameters of equilibrium constant of solvation of the analyte in the film, analyte diffusion in the film, and the effect of the conductive medium. The model will be used to select polymer suites with maximum separation in patterns for particular analyte suites. This type of selection may result in using some subset of the

32 sensors for various patterns.

It is possible that the use of carbon as the conductive medium is responsible for the non-linearity of responses at low concentrations. Studies of the use of metals such as gold or oxides of transition metals as the conductive medium is underway. It has been found that alcohols and ketones desorb from metals more rapidly than they do from carbon.

Data Acquisition

Current research in data acquisition is investigating the use of frequency dependent methods for data acquisition. AC methods are generally more sensitive than DC methods of measurements; AC methods may allow the use of thinner, higher resistance films, thus increasing film sensitivity. Some sensors exhibit high frequency noise which may be caused by local heating while resistance is measured, by inhomogeneously distributed carbon, or variable thickness in the film. Thinner sensors could eliminate some sources of noise, and AC measurements may filter out some of the noise.

To test whether high frequency noise can be filtered by AC methods, a single

sensing film of polyethylene oxide/carbon was exposed to 2500 ppm methanol and the

impedance measured at several frequencies, including DC resistance. As shown in

Figure 27.8, there is substantially less baseline drift when sensor response is plotted as

dI/I0 where I is the impedance, than there is in the same sensor measured at DC, but

higher frequency noise is not diminished at the frequencies at which impedance was

measured. As would be expected, the magnitude of the response is not substantially

different when measured by AC or DC methods, as the film is equally thick and

probably equally sensitive. The decision whether to change over to using AC

measurement techniques will consider the efficiency of removing baseline drift through

digital filtering in the data analysis process vs. the electronic requirements for AC

measurements. It may be sufficient to measure DC resistance and remove the high

frequency noise by increasing the number of signal averages from 16 to 32 or 64 and

remove the low frequency noise by digital filtering in data processing, as described

above.

Data Analysis Though the data analysis software developed for this ENose program was highly successful for its application, several improvements can be made in the future. The overall approach to data analysis will not be modified in the Second Generation device.

The major change will be the addition of real time or quasi-real time analysis. For the flight experiment, data were stored and analyzed after the flight. For ground test experiments in which events are manufactured to challenge the ENose, the goal is to have data analyzed within minutes of detection. For faster data analysis, it will be necessary to implement a reliable automated event identification routine and to translate the identification and quantification routines from Matlab into C.

There will also be some adjustments to the identification and quantification routines.

First, the current data analysis software uses all 32 sensors’ responses as input.

Though each sensor’s response was weighted in the analysis in order to maximize the differences between similar signature patterns observed for different gas compounds, it was not done systematically and therefore was not necessarily optimal. In the second generation, the selection of the to-be-used sensor set and their corresponding weights will be optimized by maximizing distances between gas signatures. The distance between the signatures for gas m and gas n, dmn , is defined as

N 1 d mn RRm,,i n i N i where Rm,i is the ith sensor’s normalized (fractional) resistance change for the mth gas and the summation is over N sensors used. Second, the core of our data analysis software is the modified LM-NLS method, which is heavy with matrix operations and largely determines the entire data analysis speed. Matrix operation speed is known to be exponentially slower as the matrix size increases. One way to increase speed is to reduce the size of the matrix dynamically in operation by incorporating sensors’ characteristic response information, such as known negative or no responses to certain gas compounds.

This characteristic response information can also be used for compounds which cannot be identified by the software; sensors which are known to respond or not to respond to particular functional groups can be sampled for a match. Thus, while it may not be possible to identify unexpected compounds, it will be possible to classify them by functional group.

In the First Generation ENose, data analysis is performed on the steady state signal produced by changes in the atmosphere. For air quality monitoring, using the steady state signal is, in general, acceptable, as a transient will not remain in the environment long enough to do harm. However, there are toxins which can be hazardous as transients. With automated event determination, analysis can begin as soon as the resistance measurement passes the pre-set threshold rather than waiting for steady state to be reached. In addition, if desorption time is a function of conductive medium, then it may be possible to use the kinetics of sensor film response for identification and quantification. Several compounds, such as ammonia, can be identified by the shape of the response curve upon visual inspection of the curve.

Quantification of the kinetics of response may enable identification of transients.

CONCLUSION

The results of the flight experiment were somewhat disappointing to the experimenters, while satisfying to the crew. There were no anomalous events, and the

ENose was not challenged to identify compounds to which it had been trained.

Nevertheless, the experiment was successful. The ENose detected changes in humidity and the presence of the daily marker, was able to identify and quantify the changes, and was able to use the training set made in the laboratory to do the data analysis. Further work in development of the JPL ENose will involve substantial challenge to the device and to the analysis software, with blind testing, mixtures, and unknowns which can be identified by functional group.

REFERENCES

K. J. Albert, N. S. Lewis, C. L. Schauer, G. A. Sotzing, S. E. Stitzel, T. P. Vaid and D. R.

Walt, “Cross-Reactive Chemical Sensor Arrays,” Chem. Rev., 2595-2626 (2000).

P. N. Bartlett and J.W. Gardner, Electronic Noses : Principles and Applications, Oxford

Univ Press, Oxford (1999).

M. G. Buehler and M.A. Ryan, “Temperature and Humidity Dependence of a Polymer-

Based Gas Sensor,” Proc. SPIE Conf. on Electro-Optical Tech. for Chemical

Detection, (1997).

C. C. Chan, H. Ozkaynak, J. D.Spengler and L. Sheldon, “Driver Exposure To Volatile

Organic-Compounds, CO, Ozone, and NO2 Under Different Driving Conditions,”

Environmental Science & Technology, 25, 964 (1991).

M. S. Freund and N. S. Lewis, “A Chemically Diverse Conducting Polymer-Based

“Electronic Nose”, Proc. National Academy of Science, 92, 2652, (1995).

J. T. James, T.F. Limero, H.J. Leano, et al., “Volatile Organic Contaminants Found in

the Habitable Environment of the Space-Shuttle: STS-26 TO STS-55,” Aviation,

Space Environ. Med., 65, 851 (1994).

M. Lampton, "Damping-Undamping Strategies for the Levenberg-Marquart Nonlinear

Least-Squares Method," Computers in Physics, 11, 110 (1997).

C. Lawson and R. Hanson, Solving Least Squares Problems, S.I.A.M. Press,

Philadephia, 1995.

P .L. Leung and R. M. Harrison, “Roadside and In-vehicle Concentrations of

Monoaromatic Hydrocarbons,” Atmospheric Environment, 33, 191 (1999). M. C. Lonergan, E. J. Severin, B. J. Doleman, R. H. Grubbs and N. S. Lewis “Array-

Based Sensing Using Chemically Sensitive, Carbon Black-Polymer Resistors”,

Chem. Materials, 8, 2298 (1996).

M. A. Ryan, M.L. Homer, M.G. Buehler, K.S. Manatt, F. Zee, and J. Graf, “Monitoring

the Air Quality in a Closed Chamber Using an Electronic Nose,” Proceedings of the

27th International Conference on Environmental Systems, Society of Automotive

Engineers, 97-ES84 (1997).

M. A. Ryan, M. L. Homer, M. G. Buehler, K. S. Manatt, B. Lau, D. Karmon and S.

Jackson, “Monitoring Space Shuttle Air for Selected Contaminants Using an

Electronic Nose,” Proceedings of the 28th International Conference on Environmental

Systems, Society of Automotive Engineers, 981564 (1998).

M. A. Ryan, M. L. Homer, H. Zhou, K. S. Manatt, V. S. Ryan and S. P. Jackson,

“Operation of an Electronic Nose Aboard the Space Shuttle and Directions for

Research for a Second Generation Device,” Proceedings of the 30th International

Conference on Environmental Systems, Society of Automotive Engineers, 00ICES-

259 (2000).

M. A. Ryan and N. S. Lewis, “Low Power, Lightweight Vapor Sensing Using Arrays of

Conducting Polymer Composite Chemically-Sensitive Resistors,” Enantiomer, in

press

E. J. Severin, B. J. Doleman and N. S. Lewis, “An Investigation of the Linearity and

Response to Mixtures of Carbon Black-Insulating Organic Polymer Composite

Vapor Detectors”, Anal. Chem., 72, 658 (2000). Spacecraft Maximum Allowable Concentrations for Selected Airborne Contaminants,

Vols. 1 & 2, National Academy Press, Washington, DC (1994).

Space Physiology and Medicine, A.E. Nicagossian, C.L. Hunton & S.L. Pool, eds., Lea

and Febiger, Philadelphia (1994).

G. Stang, Linear Algebra and its applications, 2nd ed, Academic press, New York,

1980.

R. Storn, “On the usage of differential evolution for function optimization,” Biennial

Conference of the North American Fuzzy Information Processing Society, NAFIPS,

IEEE, 519 (1996).

ACKNOWLEDGEMENTS

The research reported in this paper was carried out at the Jet Propulsion

Laboratory, California Institute of Technology under a contract with the National

Aeronautics and Space Administration, and was supported by NASA Code UL. FIGURE CAPTIONS

Figure 27.1 The JPL Electronic Nose used in the flight experiment on STS-95 is shown

as a block diagram and as a photo. The developmental device occupies a volume

of 2000 mL and has a mass of 1.4 kg, including the HP 200 LX computer.

Figure 27.2: a) A virtual peak is created at time 21:08 when the air flow is switched

from the charcoal filter, which determines the clean air baseline, to the inert filter

which is used during normal measurements. The baseline drift can be determined

by fitting the trend of the clean air baseline; in this case the virtual peak can be

attributed to baseline drift.

b) A virtual peak which is not attributable to baseline drift can be analyzed for the

presence of hazardous materials.

Figure 27.3: a) Grey, noisy trace: raw resistance as recorded; dark line: baseline drift

determined by low frequency digital filtering.

b) Grey trace: resistance after baseline drift subtracted; dark line: Processed data,

resistance after noise accomodation by smoothing and high frequency filtering, and

baseline drift corrected.

Figure 27.4 Representative signatures of ten targeted gas compounds plus wipe

generated by ENose sensors. Notice the similarity between ethanol and methanol,

and the significant difference between benzene and toluene.

Figure 27.5 Identification and quantification of four single gases using LM-NLS. The

shaded area is the target +/- 50% detection range.

Figure 27.6 Identification and quantification of three single gases using LM-NLS

Figure 27.7 Sample data from STS-95 ENose Flight Experiment. Circles are the plot of independent humidity measurements in the stairway from mid-deck to flight deck.

Polymer sensor responses: (A) Poly (2, 4, 6-tribromostyrene), (B) Polyamide resin (C)

Poly(ethylene oxide), (D) Poly(4-vinylphenol).

Figure 27.8 Response of a polymer/carbon film of polyethylene oxide to 2500 ppm of

methanol, at three frequencies of impedance measurement and DC resistance

measurement. AIR OUT

Pump ENose (250 mL/min) Chamber

8-bit teflon filter Microcontroller Computer carbon filter for pressure Data Acquisition HP200 LX for baselining equalization Subystem

solenoid valve (choose filter) DC 28 V in Power switch

AIR IN

Figure 27.1

Photo of ENose See jpg file Poly(2, 4, 6-tribromostyrene) 0.002 Poly(4-vinylphenol) Poly(ethylene oxide) Polyamide resin 0.001 0 / R

R 0.000 lta de -0.001

-0.002 20:00 20:30 21:00 21:30 22:00 Time of Day (Nov 2, 1998)

0.030 poly(ethylene oxide) 0.025 poly(styrene -co-allyl alcohol) 0.020 poly(ethylene-co-vinyl acetate) 0 0.015 poly(styrene -co-maleic / R

anhydride) R

a 0.010 lt e d 0.005

0.000

-0.005 15:45 16:00 16:15 16:30 16:45 17:00 17:15 17:30

Time of Day (Feb 25, 1997) Figure 27.2 a, b

Figure 27.3

Figure 27.4

200 180 2-propanol methanol 160 ethanol (ppm) 140 benzene

cted 120 te 100 80 60 40 oncentration de

C 20 0 0 20 40 60 80 100 120 140 160 180 200 Concentration delivered (ppm)

Figure 27.5

100

indole (x 103)

80 NH3 pm) toluene (p

n detected 40

centratio 20 Con

0 0 20 40 60 80 100 Concentration delivered (ppm)

Figure 27.6

Figure 27.7

0.015

0.01

0.005

0 -0.005 R / -0.01 dR

-0.015 1 kHz, dt = 5 sec 2 kHz, dt = 5 sec -0.02 6 kHz, dt = 1 sec DC, dt = 1 sec -0.025

-0.03 0 5 10 15 20 25 30 35 40 time (min)

Figure 27.8