International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Conference Proceedings Book ISBN: 978-605-68970-1-6

Editors : Assoc. Prof. Dr. Yakup Kutlu Assoc. Prof. Dr. Sertan Alkan Assist. Prof. Dr. Yaşar Daşdemir Assist. Prof. Dr. İpek Abasıkeleş Turgut Assist. Prof. Dr. Ahmet Gökçen Assist. Prof. Dr. Gökhan Altan

Technical Editors: Merve Nilay Aydın Handan Gürsoy Demir Ezgi Zorarpacı Hüseyin Atasoy Kadir Tohma Mehmet Sarıgül Süleyman Serhan Narlı Kerim Melih Çimrin

2

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) has been organized in Iskenderun, Hatay/Turkey on 15-17 November 2018. The main objective of ICAII4.0 is to present the latest research and results of scientists related to Artificial Intelligence, Industry 4.0 and all sub-disciplines of Computer Engineering. This conference provides opportunities for the different areas delegates to exchange new ideas and application experiences face to face, to establish business or research relations and to find global partners for future collaboration. We hope that the conference results provide significant contribution to the knowledge in this scientific field. The organizing committee of conference is pleased to invite prospective authors to submit their original manuscripts to ICAII4.0. All paper submissions will be double-blind and peer-reviewed and evaluated based on originality, technical and/or research content/depth, correctness, relevance to conference, contributions, and readability. Selected papers presented in the conference that match with the topics of the journals will be published in the following journals:

• Natural and Engineering Sciences • Communication in Mathematical Modeling and Applications • Modelling & Application & Theory • International Journal of Intelligent Computing and Cybernetics • Journal of Intelligent Systems with Applications

In particular we would like to thank Prof.Dr. Türkay Dereli, Rector of Iskenderun Technical University; Prof.Dr. Mehmet Tümay, Rector of Adana Science ve Technology University; Natural and Engineering Sciences, Academic Publisher; Communication in Mathematical Modeling and Applications; Modelling & Application & Theory; International Journal of Intelligent Computing and Cybernetics; Journal of Intelligent Systems with Applications. They have made a crucial contribution towards the success of this conference. Our thanks also go to the colleagues in our conference office. Looking forward to see you in next Conference.

Conference Organizing Committee

3

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

SUPPORTERS

Adana Science and Technology Iskenderun Technical University University www.iste.edu.tr www.adanabtu.edu.tr

Payas Municipality Iskenderun Municipality www.payas.bel.tr www.iskenderun.bel.tr

T.R. Eastren Mediterranean Development Agency www.dogaka.gov.tr

4

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

UnipaStore Casmin Hotel www.unipastore.com www.casmin.com.tr

Communication in Mathematical Natural and Engineering Sciences Modeling and Applications www.nesciences.com ntmsci.com/cmma

Inbox www.inboxmailmarketing.com

5

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

COMMITTEES

HONOUR COMMITTEE

Prof. Dr. Türkay DERELİ Iskenderun Technical University Rector

Prof. Dr. Mehmet TÜMAY Adana Science and Technology University Rector ORGANISATION COMMITTEE

Assoc. Prof. Dr. Yakup KUTLU Iskenderun Technical University

Assoc. Prof. Dr. Sertan ALKAN Iskenderun Technical University

Assoc. Prof. Dr. Serdar YILDIRIM Adana Science and Technology University

Assoc. Prof. Dr. Esen YILDIRIM Adana Science and Technology University

Dr. Yaşar DAŞDEMİR Iskenderun Technical University

Dr. Ahmet GÖKÇEN Iskenderun Technical University

Dr. İpek ABASIKELEŞ TURGUT Iskenderun Technical University

Dr. Gökhan ALTAN Iskenderun Technical University

Onur YILDIZ Eastern Mediterranean Development Agency SCIENTIFIC COMMITTEE

Prof. Dr. Adil BAYKASOĞLU Dokuz Eylul University

Prof. Dr. Ahmet YAPICI Iskenderun Technical University

Prof. Dr. Alpaslan FIĞLALI Kocaeli University

Prof. Dr. Cemalettin KUBAT Sakarya University

Prof. Dr. Dardan KLIMENTA University of Pristina

Prof. Dr. Hadi GÖKÇEN Gazi University

Prof. Dr. Harun TAŞKIN Sakarya University

Prof. Dr. Mustafa BAYRAM Istanbul Gelisim University

Prof. Dr. Nilgün FIĞLALI Kocaeli University

Prof. Dr. Novruz ALLAHVERDI KTO Karatay University

Prof. Dr. Zerrin ALADAĞ Kocaeli University

Assoc. Prof. Dr. Ali KIRÇAY Harran University

Assoc. Prof. Dr. Adem GÖLEÇ Erciyes University

Assoc. Prof. Dr. Aydın SEÇER Yildiz Technical University

Assoc. Prof. Dr. Darius ANDRIUKAITIS Kaunas University of Technology

Assoc. Prof. Dr. Emin ÜNAL Iskenderun Technical University

6

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Assoc. Prof. Dr. Eren ÖZCEYLAN Gaziantep University

Assoc. Prof. Dr. Esen YILDIRIM Adana Science and Technology University

Assoc. Prof. Dr. Önder TUTSOY Adana Science and Technology University

Assoc. Prof. Dr. Selçuk MISTIKOĞLU Iskenderun Technical University

Assoc. Prof. Dr. Serdar YILDIRIM Adana Science and Technology University

Assoc. Prof. Dr. Serkan ALTUNTAŞ Yildiz Technical University

Dr. Alptekin DURMUŞOĞLU Gaziantep University

Dr. Atakan ALKAN Kocaeli University

Dr. Cansu DAĞSUYU Adana Science and Technology University

Dr. Celal ÖZKALE Kocaeli University

Dr. Çağlar CONKER Iskenderun Technical University

Dr. Çiğdem ACI Mersin University

Dr. Ercan AVŞAR Cukurova University

Dr. Esra SARAÇ Adana Science and Technology University

Dr. Fatih KILIÇ Adana Science and Technology University

Dr. Fei HE Imperial College London

Dr. Hatice Başak YILDIRIM Adana Science and Technology University

Dr. Houman DALLALI California State University

Dr. Hüseyin Turan ARAT Iskenderun Technical University

Dr. Koray ALTUN Bursa Technical University

Dr. Mehmet ACI Mersin University

Dr. Mehmet Hakan DEMİR Iskenderun Technical University

Dr. Mahmut SİNECAN Adnan Menderes University

Dr. Mustafa İNCİ Iskenderun Technical University

Dr. Mustafa Kaan BALTACIOĞLU Iskenderun Technical University

Dr. Mustafa YENİAD Ankara Yildirim Beyazit University

Dr. Salvador Pacheco-Gutierrez University of Manchester

Dr. Tahsin KÖROĞLU Adana Science and Technology University

Dr. Yalçın İŞLER Izmir Katip Celebi University

Dr. Yildiz ŞAHİN Kocaeli University

Dr. Yunus EROĞLU Iskenderun Technical University

Dr. Yusuf KUVVETLİ Cukurova University

7

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

INVITED SPEAKERS

Prof. Dr. Adil BAYKASOGLU Dokuz Eylul University “Dynamic Optimization in a Dynamic Industry”

Prof. Dr. Ercan OZTEMEL Marmara University “Artificial Intelligence and its Effect on Social Transformation”

Prof. Dr. Novruz ALLAHVERDI KTO Karatay University “Fuzzy Logic and its Applications in Medicine”

Dr. Yalcin ISLER İzmir Katip Celebi University “Data Security and Privacy Issues of Implantable Medical Devices”

8

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Content Agent Based Modeling in the Fuzzy Cognitive Mapping Literature ...... 11 An Overall Equipment Efficiency Study and Improvements in Ulus Metal Company ...... 20 Application of Tree-Seed Algorithm to the P-Median Problem ...... 32 Artificial Intelligence And Social Transformation ...... 37 Brain Computer Interface Chess Game Platform ...... 39 Competition in Statistical Software Market for ‘AI’ Studies...... 48 Comparison of Investment Options and an Application in Industry 4.0 ...... 55 Computer-Aided Detection of Pneumonia Disease in Chest X-Rays Using Deep Convolutional Neural Networks ...... 65 Convolutional Neural Network for Environmental Sound Event Classification ...... 70 Current Situation of Gaziantep Industry in the Perspective of Industry 4.0: Development of a Maturity Index ...... 77 Demands in Wireless Power Transfer of both Artificial Intelligence and Industry 4.0 for Greater Autonomy ...... 89 Detection and 3D Modeling of Brain Tumors Using Image Segmentation Methods and Volume Rendering ...... 99 Determination of Groundwater Level Fluctuations by Artificial Neural Networks ...... 106 Dynamic Scheduling in Flexible Manufacturing Processes and an Application ...... 113 Dynamic Optimization in A Dynamic Industry ...... 123 Estimation of Daily and Monthly Global Solar Radiation with Regression and Multi Regression Analysis for Iskenderun Region ...... 124 Gender Estimation from Iris Images Using Tissue Analysis Techniques ...... 133 Interactive Temporal Erasable Itemset Mining ...... 140 Jute Yarn Consumption Prediction by Artificial Neural Network and Multilinear Regression ...... 149 Latest Trends in Textile-Based IoT Applications ...... 159 Localization and Point Cloud Based 3D Mapping With Autonomous Robots ...... 164 NoSQL Database Systems: Review and Comparision ...... 172 Response of Twitter Users to Earthquakes in Turkey ...... 178 Review of Trust Parameters in Secure Wireless Sensor Networks ...... 187 Superiorities of Deep Extreme Learning Machines against Convolutional Neural Networks ...... 192 The Changing ERP System Requirements in Industry 4.0 Environment: Enterprise Resource Planning 4.0 ...... 197 The Effect of Different Stimulus on Emotion Estimation with Deep Learning ...... 202

9

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

The Investigation of The Wind Energy Potential of The Belen Region and The Comparison of The Wind Turbine with The Production Values ...... 209 The Effects of Sodium Nitrite on Corrosion Resistance Ofsteel Reinforcement in Concreta 221 The Evaluation and Comparison of Daily Reference Evapotranspiration with ANN and Empirical Methods ...... 226 The Importance of Business Intelligence Solutions in The Industry 4.0 Concept ...... 236 Turkish Abusive Message Detection with Methods of Classification Algorithms ...... 242

10

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Agent Based Modeling in the Fuzzy Cognitive Mapping Literature

Zeynep D. Unutmaz Durmuşoğlu, Pınar Kocabey Çiftçi

Department of Industrial Engineering, Gaziantep University, Turkey [email protected] [email protected]

Abstract

Agent Based Modeling (ABM) and Fuzzy Cognitive Mapping (FCM) are two important techniques for helping researchers to overcome the complexities of the real life systems/problems. Although the structures of them differ from each other, they can provide beneficial solutions according to the problem contents and types in their own ways. The utilities of using these techniques in complex and dynamic problems have taken the interests of modelers and researchers since the first studies of them was appeared in the literature. These interests have yielded vast bodies of literature for both techniques separately. On the other hand, it is possible to find some studies that use these two techniques in the same study to obtain more advanced and useful way to solve problems. In this study, the FCM literature and FCM with ABM related literature were quantitatively examined in order to observe the current trends in the academic publications. In addition, the contents of FCM with ABM literature were reviewed to search the benefits of using both methodologies together. Thomson Reuters Web of Knowledge was used as the database of this study. A total of 896 publications related to FCM were found. However, only limited number of these publications used both ABM and FCM together. This study showed that there still may be an important gap on this topic in the literature.

Keywords: Agent Based Modeling, Fuzzy Cognitive Mapping, Quantitative Analysis

1. Introduction

Researchers of today's world need to deal with many complex real life problems. These problems mostly consist of several different variables that can interact and affect each other under changing external and internal situations. Reaching the optimal solutions for them can become an impossible task. That is why; searching and finding a possible solution for them can bring to mind a question like "Is there any better way for solving this problem?". The idea of "searching for better" leads the world to improve the current technologies, systems, and methodologies. Thus, it is possible to find several different approaches that can be implemented to solve similar real life problems in the literature. Among them, Agent Based Modeling (ABM) and Fuzzy Cognitive Mapping (FCM) are two useful techniques to overcome the complexities of these problems. ABM is a powerful simulation method (Bonabeau, 2002) that consists of autonomous and interacting agents (Macal & North, 2005). It is relatively new when compared to Discrete Event Simulation and System Dynamics (Borshchev & Filippov, 2004). In the ABM, the studied system is basically modeled as a collection of autonomous agents that may imitate various behaviors of the system that represent such as producing, selling, consuming etc.(Bonabeau, 2002). ABM defines the behaviors at individual level and that is why; it is called as bottom up approach (Borshchev & Filippov, 2004). Thus, encoding micro-rules of behavior and measuring the emergent macro level results can be called as the major power of the ABM (Rand & Rust, 2011). On the other hand, FCM is an extended version of cognitive mapping (CM). CM is a directed graph that gives information about perceptions of individuals relevant to a specific topic at a particular point of time (Elsawah, Mclucas & Mazanov, 2013). CMs consist two main elements: nodes (concepts) that are

11

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY used to represent the variables to describe the belief system of an individual and casual relationship (casual belief) represents causal dependencies between variables (Bertolini, 2007). Although CMs help researchers to analyze the complex relationship among studied issues, they are not sufficient to represent uncertainty involved in the casual relations among nodes in a map (Lee, Lee & Lee, 2012). In order to overcome this limitation of CMs , an extension of the CM was proposed by Kosko (Baykasoglu, Durmusoglu & Kaplanoglu, 2011).This new version of the CMs is called as Fuzzy CMs (FCM) that allow researchers to analyze the strength of the impacts of a concept node to other concepts (Lee, Lee & Lee, 2012). The benefits of both of these methodologies in complex and dynamic real life problems have attracted the interests of researchers since the first studies of them were appeared in the literature. These interests have yielded vast bodies of literature for both techniques separately. However, the possible utilities of using these two methodologies cannot be underestimated. Thus, studies that focus on ABM and FCM in one study started to appear in the literature, too. In this context, the major objective of the presented paper is examining quantitatively the place of ABM in the literature of FCM and searching for the research trends in FCM with ABM studies. Academic publications are considered one of the most important resources for this type of analysis (Durmusoglu, 2016). Thus, bibliometric analysis was used for performing a quantitative analysis. As the result of the analysis, the points such as the change in publication counts for FCM literature and FCM with ABM related literature, mostly studied research areas, most productive countries, most productive authors and etc. are presented. The studies that used both methodologies in the same study were also examined to reveal how intelligent agents were used in FCM studies and what kinds of areas were chosen to implement this type of analysis. The details of the data retrieval procedure are explained in the methodology section. The findings of the study are presented in section 3. Finally, section 4 provides a conclusion. 2. Methodology

A bibliometric analysis was performed on Fuzzy Cognitive Map field using the academic publications. Bibliometric analysis is an important technique which uses statistical methods in order to determine the characteristics of publications according to features such as field, source, topic, author, country and etc (Abejon & Garea, 2015, Durmusoglu & Çiftçi, 2017a). It is also a beneficial method for obtaining a clear picture of the current state of the scientific researches in special fields (Battisti & Salini, 2012, Zyoud, Al-Jabi & Sweileh, 2014). It has been implemented for several different topics due to the practicability of it ( Durmusoglu & Çiftçi, 2017b). Thus, it can be found hundreds of bibliometric analysis in the literature on different fields. However, there is still a gap on the comparison of the FCM studies and FCM with ABM studies quantitatively to our best knowledge. In this study, the publications related to FCM were retrieved from the Thomson Reuters Web of Knowledge (WOK) all databases. There are several reasons of studying with WOK: • It includes a wide range of bibliographic databases, citations and scientific publication references (Sanchez, Cruz Del Rio Rama & Garcia, 2017). • It covers the publications indexed by SCI-SSCI-SCIE published by 500 publishers (Durmusoglu, 2016). • It includes more than 50,000,000 articles and 15,000 journals that have high quality standards (Hew, 2017, Merigo, Mas-Tur, Roig-Rierno & Riberio-Soriano, 2015). Different search terms were determined for FCM and ABM related FCM. For the basic FCM field, the search terms were as "fuzzy cognitive map", "fuzzy cognitive mapping", "fuzzy" and "cognitive mapping", "fuzzy" and "cognitive map". For the ABM related FCM field, the same search terms were used with "agent based" search term.

12

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

The topic search field of WOK was used in order to perform a broad analysis on the determined field. There was no limitation on the publication year, document types, and languages. The content of the FCM with ABM studies were also reviewed in detail in order to understand how the agents can be used in FCM studies and the possible application areas. The results of the performed analysis are presented in the next section in detail.

3. Results

The results of the presented study are collected under two titles. The findings of the bibliometric analysis are given in the first sub-section while the findings of the review on FCM with ABM studies are provided in the second sub-section.

a. The Results of the Quantitative Analysis

A total of 896 publications related to FCM were found in the WOK all databases for the determined search terms. Figure 1 represents the distribution of the FCM related publications by years. The first studies of the field started to be seen in 1980s. The number of publications generally showed an increasing trend starting from 1995. It passed the ten publications after 2001. A fluctuating trend can be seen after 2001. In some years, there was an important increase while decreasing in some years. It reached the peak point in 2015 and started to decrease after 2015. The number of publications in 2018 is not completed since the year has not ended yet. On the other hand, the quantitative analysis of publications which uses ABM and FCM in one study showed that, there is still an important gap about this area because only 12 of the 896 FCM related publications mentioned about ABM, too. It is almost 1.34% of all publications. The first publication of the FCM with ABM studies was released in 1999. It can be seen that the literature is lacking on this area.

101 97 90

74 70 57 52 42 36 37 35 32 28 22 25 24 13 9 9 9 6 5 6 7

1 1 3 1 4

1991 1989 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 1984 Figure 1 – Number of publications of FCM by years

In the FCM related publications, 490 (54.6%) of all FCM were articles. Meeting, patents, review were followed the articles respectively. In addition, 868 (96.87%) of all documents were prepared in English while the remaining part of them were in Korean, Russian, Spanish and etc. On the other hand, all of the ABM with FCM related studies (12 documents) were prepared as articles and written in English.

13

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

The major research areas of the retrieved publications were also analyzed. The WOK categorizes the research areas of the publications in six major categories. These categories are social sciences, life sciences and biomedicine, technology, science and technology, physical sciences, and arts humanities. More than one research area can be assigned to the papers published in WOK. Thus, a paper can be related with different research areas at the same time. That is why; if a paper has more than one research area, they were also counted for the related category in this study. Figure 2 represents the distribution of the FCM related publications according to the WOK major research area categories. The vast majority of the FCM related publications (95.20% of all) can be categorized in science and technology while the technology research areas followed it with 85.94%. The physical sciences took the third place with the 48.21% of all publications. In the FCM with ABM field, all publications were related with science and technology research area as seen in figure 4. Technology field followed it with 85.94%. Similar to the previous analysis, physical sciences took the third place for these studies.

Life Sciences Arts Biomedicine; Humanities; Social 24,33% 3,46% Sciences; Science 37,28% Technology; 95,20%

Physical Sciences; 48,21%

Technology; 85,94%

Figure 2 – Major research areas of FCM studies

The countries that provided most contribution to the studied field were also examined. Table 1 represents the publication counts and their percentages according to the countries that were declared by authors to WOK all databases as the address. One publication can be written by authors from different countries. In this case, this publication was counted for each country.

14

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Life Sciences Arts Biomedicine; Humanities; 50,00% Science 0,00% Technology; 100,00% Social Sciences; 58,33%

Physical Sciences; Technology; 66,67% 83,33%

Figure 3 – Major research areas of FCM with ABM studies

Greece was the pioneer country in the FCM related studies because 16.07% of the 896 publications had at least one author from Greece. People’s Republic of China followed Greece with 13.95% of the publications. USA took the third place. There is a dramatic decrease in the number of publications after USA. Australia, England, and Canada contributed around 5% of the studies separately.

Table 1 – Most productive countries on the studied field ONLY FCM RELATED STUDIES FCM with ABM RELATED STUDIES Country Count Percentage Country Count Percentage Greece 144 16,07% USA 4 33,33% Peoples R China 125 13,95% England 3 25,00% Peoples Republic USA 115 12,83% China 3 25,00% Australia 49 5,47% Singapore 3 25,00% England 48 5,36% South Korea 3 25,00% Canada 46 5,13% Netherlands 2 16,67% India 42 4,69% Costa Rica 1 8,33% Turkey 41 4,58% Croatia 1 8,33% South Korea 38 4,24% France 1 8,33% Iran 37 4,13% India 1 8,33% Poland 37 4,13% Russia 1 8,33%

On the other hand, USA was the pioneer country in the FCM with ABM related studies because 33.33% of the studies were prepared by authors from USA. England, Peoples R China, Singapore and South Korea took the second place with the same contribution percentage (25%) while Netherlands took the third place with 16.67%. Papageorgiou EI is the most productive author for FCM field because Papageorgiou EI prepared more than 60 of all documents. Stylios CD followed her with more than 20 papers and took the second place. On the other hand, Lee H, Lee KC, and Lee N contributed to the FCM with ABM literature mostly compared to others with 3 articles of each.

15

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

b. The Review on FCM with ABM Studies

A total of 12 of the retrieved articles mentioned FCM and ABM terms together. The two of them do not focus on combination of both methodologies. They mentioned each methodology separately. For that reason, they were not included in this review. The major objective, major usage of ABM and FCM, and the application areas of the remaining 10 studies were explained chronologically below. The first study in literature for the related topic was performed by Chiu et al. (1999). The authors presented a methodology to acquire user's problem solving skills during human computer interaction using agent based modeling and fuzzy cognitive map. Agents basically used to eliminate the noise information and obtain the critical information for valid judgment to derive a more accurate understanding about a users' problem solving knowledge. Thus, a prototype reasoning agent was setup and was verified with the dialog events during UNIX operations. The second study in the literature was conducted by Miao et al. (2002). They proposed a model with intelligent software agent who has the ability to model, reason and make decision on behalf of human beings using the theory of FCM. These agents have the ability to handle different types of complex fuzzy information and make intelligent decisions in complex situations. A simple case study was also illustrated to make the intelligent agent's action more clear. In this case, an intelligent software agent was created to accomplish the task of purchasing computer chips concerning their price, quality, and reliability. The agent mainly creates the related concepts and weight for related topic, gets the real input values, and constructs the maps. If the agent is satisfied with the product based on the analysis, it performs the purchase action. In another study that was performed in 2007, Miao et al. combined the intelligent software agent and fuzzy cognitive map by creating fuzzy cognitive agents for e-commerce and e-business sites. Fuzzy cognitive agents were used to provide personalized recommendations based on the personal preferences of the users, other user's common preferences, and expert knowledge. They are able to communicate with users, perceive environment, represent knowledge, learn from behaviors of users, make inference based on knowledge, and make personalized recommendation to individual users. The proposed agent approach was supported with a case study in which agents acted as personal assistants for providing personalized suggestions to users in a used car electronic market over the Internet. In 2010, Stula et al. developed an agent based fuzzy cognitive map by embedding the multi agent system into the fuzzy cognitive map. This new approach basically focused on the fuzzy cognitive map in which each node was mapped into the agent. The classical FCM uses the same inference algorithm for the generation of each concept. However, the presented agent based fuzzy cognitive map provides the ability to use different algorithm for generation of each concept. The presented approach was supported with two different systems applications; an inverted pendulum and a three coupled tanks systems. The author Kun Chang Lee and his friends contributed to the studied area significantly because they performed three of the following studies respectively. Firstly, a new approach called as the multiple agent based knowledge integration mechanism (MAKIM) was proposed in the study of Lee et al.(2012a). This approach mainly utilizes from fuzzy cognitive map, particle swarm optimization and agent based modeling. FCM was used to provide the knowledge of various experts on the problem. Particle swarm optimization was used to train the constructed FCM. In this study, each node of the FCM was modeled as an agent that can communicate with other node agents. Thus, each agent can have results of inference process and can make autonomous decisions in order to reduce inference time. The proposed approach was implemented on an IT project risk assessment to test the validity of the approach. In their second study, Lee et al.(2012b) focused on building an agent based mobile negotiation mechanism in order to perform personalized pricing of last minute theatre tickets considering time remaining to the performance and the locations of the potential customers. The sellers who make decision to maximize the profits and buyers who maximize the gains are defined as agents in the presented model. The casual relationships between the factors were investigated using fuzzy cognitive map. In their last study, Lee et al. (2013) basically combined fuzzy cognitive map and agent based

16

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY modeling for industrial marketing planning. This study explained drawbacks of classical FCM such as fixed relationship among factors, lacking a time concept, difficulty to represent high number of nodes and etc. These drawbacks were eliminated using agent technology. The nodes of the FCM represented as agents in the proposed approach. Also the agents such as interface agent, coordination agent were also created to develop an interference of an FCM and set up a network of node agents. This study is an important study to understand how to overcome the drawbacks of a classic FCM. In 2014, Mei et al. combined both methodology to investigate on an health topic. They used FCM in order to examine how the factors of individual emotions and cognition influence each other. Emotions were represented by emotional agents who could interact with an environment and other agents. These agents mainly used to generate human like emotion outputs that affect the decision making process. Considering the individual emotions and cognition of individuals and expert knowledge, they performed FCM for infectious disease simulation. In 2015, Chen performed an analysis to generate an intelligent autonomous tracing system based on internet of things with the help of FCM. This system was generated for tracing food product usage life cycle. Intelligent software agents helped to reduce tracing time and perform autonomous operation for food product problem tracing system. Agents and FCM contributed in assessing the most critical factors for backward design process such as internet of things based tracing system. The proposed method was applied on agriculture food product life cycle. In 2017, Giabbanelli et al. searched the two ways of using agents in FCM studies. In the first one, the FCM can be embedded within each agent. In the second one, agents can be created within one FCM. In the first one, each agent basically has its own mental FCM to run. In the second one, there is one FCM (macro level) and agents are identified in this macro FCM. The authors discussed the details of these two different approaches in detail to the related study. The review basically indicated that, researchers mostly implements both methodology to together in order to reduce the major drawbacks of the FCM and obtain more intelligent systems that can provide the planned interaction for the studied system. Agents in the studies mostly created as intelligent software agents or concepts of the examined FCM. They provide the autonomous behavior that the researchers searched for their studies. On the other hand, FCM used to represent the causal relations. The usage of FCM generally helped to reduce the limitations of the FCM.

4. Conclusions

This paper quantitatively examined the scientific papers related to FCM. It also searched for the current trend in the publications that uses ABM and FCM methodologies in the same studies. Web of Knowledge was used as the main database. The results showed that a total of 896 studies were found related to FCM by using the relevant search terms. The considerable amount of publications proved that FCM has built up a vast body of literature. The main focal point of the study is to find the place of ABM in FCM publications. The findings of the study indicated that only 12 of the 896 FCM related publications were related with ABM. The country statistics on the retrieved studies showed that Greece, Peoples Republic of China, and USA are the pioneering countries that contributed the examined FCM literature mostly. On the other hand, USA, England, Peoples Republic of China, Singapore, and South Korea contributed to the ABM and FCM related literature. Although both methodologies are accepted as useful techniques for social science, the vast majority of the publications were related to science and technology field and technology field. The results of the review indicated that, ABM can provide important advantages to the FCM studies. The structure of agent can help to make autonomous decisions. Also, it helped to reduce the limitations of FCM. This study focused on the numeric data supplied by WOK all databases. Future studies can consider other databases and also the contents of the each study in order to perform more detailed study.

17

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

References

Abejon, R. & Garea, A. (2015). A bibliometric analysis of research on arsenic in drinking water during the 1992–2012 period: An outlook to treatment alternatives for arsenic removal. Journal of Water Process Engineering, 6, 105-119.

Battisti, F. D. & Salini, S. (2012). Robust analysis of bibliometric data. Statistical Methods & Applications, 22, 269-283.

Baykasoglu, A., Durmusoglu, Z. D. U. & Kaplanoglu, V. (2011). Training Fuzzy Cognitive Maps via Extended Great Deluge Algorithm with applications. Compuers in Industry, 62, 187-195.

Bertolini, M. (2007). Assessment of human reliability factors: A fuzzy cognitive maps approach. International Journal of Industrial Ergonomics, 37,405-413.

Bonabeau, E.(2002). Agent-based modeling: Methods and techniques for simulating human systems. Proceedings of the National Academy of Science, 99, 7280-7287.

Borshchev, A. & Filippov, A. (2004). From System Dynamics and Discrete Event to Practical Agent Based Modeling: Reasons, Techniques, Tools. 22nd International Conference on System Dyamics Society.

Chen, R. Y. (2015). Autonomous tracing system for backward design in food supply chain. Food Control, 51, 70-84.

Chiu, C. C, Hsu, P. L. & Norcio, A. F. (1999). A fuzzy reasoning approach to support knowledge acquisition. International Journal of General Systems, 28, 227-241.

Durmusoglu, A. (2016). A pre-assessment of past research on the topic of environmental-friendly electronics. Journal of Clean Production, 129, 305-314.

Durmusoglu, Z. D. U. & Çiftçi, P. K. (2017a). An analysis of trends in publications on ‘tobacco control’. Health Education Journal, 76, 544-556.

Durmusoglu, Z. D. U. & Çiftçi, P. K. (2017b). Bibliometric Analysis on Scientific Research on Innovation Diffusion. TOJSAT,7, 5-11.

ElSawah, S., Mclucas, A. & Mazanov J. (2013). Using a Cognitive Mapping Approach to Frame the Perceptions of Water Users About Managing Water Resources: A Case Study in the Australian Capital Territory. Water Resource Management, 27, 3441-3456.

Giabbanelli, P. J., Gray, S. A. & Aminpour, P. (2017). Combining fuzzy cognitive maps with agent-based modeling: Frameworks and pitfalls of a powerful hybrid modeling approach to understand human-environment interactions. Environmental Modelling and Software, 95, 320-325.

Hew, J. J. (2017). Hall of fame for mobile commerce and its applications: A bibliometric evaluation of a decade and a half (2000–2015). Telematics and Informatics, 24,43-66.

Lee, K. C., Lee, N. & Lee, H. (2012a). Multi-agent knowledge integration mechanism using particle swarm optimization. Technological Forecasting and Social Change, 79, 469-484.

Lee, K. C., Lee, N. & Lee, H. (2012b). Agent based mobile negotiation for personalized pricing of last minute theatre tickets. Expert Systems with Applications, 39, 9255-9263.

Lee, K. C., Lee, H., Lee, N. & Lee, J. (2013). An agent-based fuzzy cognitive map approach to the strategic marketing planning for industrial firms. Industrial Marketing Management, 42, 552-563.

Macal, C. M. & North, M. J. (2005). Tutorial on agent-based modeling and simulation. in Proceedings of the Winter Simulation Conference 2005, 14.

Mei, S. et al. (2014). Individual Decision Making Can Drive Epidemics: A Fuzzy Cognitive Map Study. IEEE Transactions Fuzzy Systems, 22, 264-273.

18

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Merigo, J. M., Mas-Tur, A., Roig-Tierno, N. & Ribeiro-Soriano, D. (2015). A bibliometric overview of the Journal of Business Research between 1973 and 2014. Journal of Business Research, 68, 2645-2653.

Miao, C. Y., Goh, A., Miao, Y. & Yang, Z. H. (2002). Agent that models, reasons and makes decisions. Knowledge Based Systems, 15,203-211.

Miao, C.,Yang, Q., Fang, H. & Goh, A. (2007). A cognitive approach for agent-based personalized recommendation. Knowledge Based Systems, 20, 397-405.

Rand, W. & Rust, R. T. (2011). Agent-based modeling in marketing: Guidelines for rigor. International Journal of Research in Marketing, 28,181-193.

Sanchez, A. D, de la Cruz Del Rio Rama, M. & Garcia, J. A. (2017). Bibliometric analysis of publications on wine tourism in the databases Scopus and WoS. European Reearch on Management and Business Economics, 23, 8-15.

Stula, M., Stipanicev, D. & Bodrozic, L. (2010). Intelligent Modeling with Agent-Based Fuzzy Cognitive Map. International Journal of Intelligent Systems, 25, 981-1004.

Zyoud, S. H., Al-Jabi, S. W. & Sweileh, W. M. (2014). Bibliometric analysis of scientific publications on waterpipe (narghile, shisha, hookah) tobacco smoking during the period 2003-2012. Tobacco Induced Disease, 12, 7.

19

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

An Overall Equipment Efficiency Study and Improvements in Ulus Metal Company Selen Eligüzel Yenice1, Bülent Sezen2

Department of Management, Gebze Technical University, Turkey [email protected] [email protected]

Abstract

This paper presents a comparison of 2 different calculation methodology of OEE in a cold forming sheet metal manufacturer. The comparisons demonstrate that automatical data input systems which is working on cloud systems can determine the most realistic data for continuous improvement studies. The fourth stage of industrial revolutions, provides smart solutions and analysis to all enterprises. At the beginning of this revolution, the enterprises should aim to redesign their own system accordingly for being more competitive.

Keywords: OEE, Industry 4.0, cloud, continuous improvement

1.Introduction At the present time, all enterprises are trying to continue their lives with their national and international competitors in a very challenging competitive environment. Although the quality - cost - time trio is important for customers, the first point of attention is always the cost and the suppliers who have the lower cost from the level demanded by the customer, or even lower than it, are becoming superior to their competitors. In today's world where the distances are lost, the prices determined by the customer may not seem to be realisible according to the current production system. The way to achieve this pass through the improvement activities. Even though it is a very large scope when thinking about improvement and it doesn't give any idea about what, where, how, how much and who is going to improve. At this point, it is very important to focus on a improvement topic which provides the effective results in a short time. Because when enterprises solve 20% of their basic problems, they can provide 80% of improvement. Overall Equipment Efficiency (OEE) is one of the common used method to analyze the current situation, find out the factors which are causing loss and compare with the improved values. OEE is an efficiency enhancement tool developed by Seiichi Nkajima in the 1960s to determine how effectively machines are used (Lahri and Pathak, 2015). The main feature that differentiates OEE from other productivity measurement tools is consisting of 3 different components as availability, performance, quality and shows that which one should be focused on.

In the light of the correct data collected by the correct method, OEE shows the correct results for companies and sheds light on the processes which should be improved. If the method determined in the data collection is reliable and devoid of intervention of any person, the collected data are realistic information. Industry 4.0 which is the fourth stage of industrial revolutions, is expected to lead to a change in the current world order which labor costs and productivity make a difference. Nowadays, one of the working areas of the smart factory systems made within the scope of Industry 4.0, works on continuous monitoring of machines with established systems, reporting downtime, monitoring machine operation

20

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY according to defined cycle time and storing all these information on the cloud. Accessing the information instantaneously when needed, preparing reports without requiring any extra work and accessing this information from anywhere & anytime you want, makes all these calculation and tracking transactions much easier.

The aim of this study is comparing 2 different data collection methods for calculating OEE in Ulus Metal San.Tic.A.S.. First method is collecting data with paper forms and second method is collecting data by using a hardware and software program called Prowmes. Prowmes's hardware connects to machine’s PLC system and import data to software.

Ulus Metal is a medium-sized family company where located in Gebze Organized Industrial Zone with 248 employees. The company is one of the best skilled cold forming sheet metal parts producer in Turkey and serve to the automotive and compressor industry.

2. Overall Equipment Efficiency

Overall Equipment Efficiency (OEE) is a calculation tool that shows the utilization rates of equipment and is made up of 3 main components; availability, performance and quality. For a better understanding of these 3 components, please see Graphic 1 and then examine the below definitions.

Total Available Time: The total time which the machine can produce parts. This time contains all planned break time and planned production time. Planned Production Time: If all holidays and planned break time are took out of total available time, then planned production time is calculated. Operating Time: If all time losses like setup time, supply&filling time and waiting time are took out of planned production time, then residual net time is operating time. Net Operating Time: When performance&speed losses are took out of operating time, calculated net production time is called net operating time. Productive Time: The time which is used for producing good parts.

Availability = Operating Time/Planned Production Time Performance = Parts Produced/(Operating Time/Cycle Time) Quality = Good Parts/Parts Produced OEE = Availability*Performance*Quality

Graphic 1–OEE Categories (http://www.acusys.co.za/en/others/introduction-to-oee 29.10.2018)

21

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

OEE is determined 85% according to world standards and to achieve this target; the availability should be 90%, the performance should be 95% and the quality should be 99%. In Turkey, manufacturers’ OEE value is around 50% - 60% (Maraşlı ve Kemahlı, 2013).

2. Literature Review When the literature is examined, it is seen that OEE value is used as a measurement and comparison tool in order to increase efficiency and eliminate waste. When collecting data for calculating OEE, at the same time downtime is determining which has a key role either to calculate the availability or to determine the best improvement study. If Industry 4.0 applications are used in data collection, the enterprises will spend less time and less manpower however they win through a better result. Singh et al (2012) used OEE to measure the efficiency of the study about the Total Productive Maintenance in a company which is operating in the automotive sector. While they were examining the main issues affecting the productive maintenance, they recorded all the downtime and production quantities which were affected to OEE. After that they made improvement studies. When they compared OEE value before and after the study, they noticed that it is increased from 63% to 79%. Naik et al (2015) developed a program in their simulation study at production lines to calculate OEE value and compare with the world class OEE. In the study, they explained 6 main factors which caused low OEE value. They determined that one cost-optimised line in the company has 80% OEE because that line’s machine downtime had been minimised at the optimising study. Polat (2014) worked on making a contribution to energy efficiency by establishing a system to follow up OEE value online. He explained that when employees collected data and reported them, there would have been either loss of or missing data. After the system creation, automatically collected & analyzed data with the calculated OEE value are shown online on a screen to the relevant personnels in the company. By this online data transfer of time and quality loss problems, determining and interfering losses online were provided 30% of increasement at the OEE value and also provided electric energy savings. Yaşin (2014) used OEE to measure the performance of door production line in a medium-sized wood factory. He followed up production line for 3 months to determine existing OEE value. During this period, line stoppages, which were caused a decreasement in the OEE value, were taken under

22

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY record and the improvement studies were planned by focusing on the determined stoppages. He tried to bring the OEE value to 85% by eliminating the existing stoppages. Alkan (2015) used OEE to calculate efficiency value of 2 different grinding lines in a micronized mine grinding plant. The project took 6 months and in the light of the data collected in the first 3-months period, it was determined that the average OEE value was 52%. According to unplanned downtime report for the first 3 months, he made improvement studies to eliminate the downtime. In the second 3 months period, OEE value is become to 70%. Temiz et al (2010) emphasized the importance of keeping OEE value in a database for regular evaluations and determining basic improvement subjects. They created a data collection system to calculate and report OEE. They collected data for 8 days and determined that the OEE value of company was 56%. Sohal et al (2010) made a study to identify difficulties of OEE applications at six different company which carried business in six different sectors. When they found the main difficulty was culture of the company, they remarked that the only way of getting over from this problem was collecting right data, analyzing and evaluating with the support of management. It was explained that the key of succes was training the operators and creating observable targets in the light of training and intellection. It is emphasized that if the companies managed these recommendations, OEE would be used for improvement studies as a main source. Sahu et al. (2015) analyzed the relationship between OEE - 5S and production efficiency by developing 3 different hypothesis. According to their research, 5S was not only improved the work area but also increased product quality, production process performance and reduced production costs. In the conceptual model they established in line of these results, they showed that 5S was increased OEE and the productivity but at the same time increased OEE value would be increased production efficiency. Iannone and Nenni (2013) started their study with the strategy of Lord Kelvin’s “if you can not measure it, you can not improve it”. They researched about OEE as a efficiency measurement tool which calculated by deviding actual produciton to planned production. The other efficiency measurement tools such as TEEP, PEE, OAE, OFE, OPE were also explained. In the case study, it is determined that if OEE calculated with the correct methods and followed up regularly, the enterprises would find 6 main losses, observe significant improvements. Pamuk and Soysal (2018), in their study about Industry 4.0, have identified the internet of objects, cyber-physical system, large data&data analytics and smart factories as the main sources used to reach Industry 4.0 level. In addition to this, they determined that the applications of Industry 4.0 aim to increase the flexibility, efficiency, speed and quality in production. At the same time they explained that daily life will change with the adaptation of these applications. Eldem (2017)explained Industry 4.0 in the form of 9 main subjects as Internet of objects, simulation, autonomous robots, layered production, augmented reality, cloud computing, cyber security, big data&analysis, horizontal&vertical system integration. In addition to this, he emphasized that Industry 4.0 was a necessity to produce value-added products and our country should be adapted to this change as soon as possible. 3. Method and Mothodology In this study, P64, which is in Ulus Metal automatic production lines, was followed up for 1 week with two diffrent ways at the same time period. First way is collecting the data with a downtime follow-up form which is given to the production operator of P64 (see Table 1 below). The second way is using a device by Prowmes which was connected to P64’s PLC system.

23

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

The downtime follow-up form is simply designed. The operator wrote the date of production, the material number of producing part, the reason of downtime and the total downtime for the relevant loss (see Form 1). The downtime follow-up form and how it would be used were explained to the operator. The given forms were taken from the operator at the end of the shift and the data on the tables entered into the OEE calculation table (see Table 2). The cycle time of the parts required which was needed for the performance calculations are taken from the cycle time studies that had done before and processed in the table. Daily part production quantities and scrap quantities are taken from the ERP program used by the company. Table 1 – Downtime Follow-Up Form for Operators

The device supplied from the Prowmes has been connected to the PLC system of P64 and the signals were tested when the machine was running and stopped. Then, the part numbers, cycle time and weekly production plan were introduced to the program from the administrator screeen on the internet. According to the data obtained from past time studies conducted by the company, the most common downtime codes were introduced to the system and the operator was only allowed to select when there were any downtime (see Figure 1). When an unspecified downtime was encountered, the operator was required to enter the downtime description. After the system definitions and tests completed, the operator screen created and the operator was trained for using the system.

Figure 1 – Operator Data Input Screen for the Reason of Downtime

24

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

In both systems, it was explained to the operator that was not expected to do any entry for small stoppages called chokote which is below 3 minutes and causes performance loss. Due to the working principle of Prowmes, due to the fact that every time losses that exceed the cycle time of the part is called as a stoppage, it was added to the program's fiction under a separate title called chokote in the downtime analysis graph. The main reason why the company was not interested in focusing on the stoppages under 3 minutes was that the firm knew that there were many long downtime types and they would like to see the analyse of long downtime. When the Prowmes device did not receive a signal that the machine was working, then the system created a stoppage and because of the program is running over the cloud, operator alerted that a downtime code needed to be selected for reporting the cause of the stop. All the users of the program, even if they were operators or administrators, could see whether the machine was working or had a stoppage, the total downtime of the day, the reasons of downtime, OEE values. If they had an internet access, they followed these datas anywhere and anytime. If requested, the downtime information which was above the defined periods was transmitted to the phone or e-mails of the people who were identified in the system.

4. Analysis and Findings OEE calculation was performed with 2 different methods at the same time for the same machine and the results were examined. The reasons and durations of the downtime in the downtime follow-up forms filled by the operator are shown in Table 2 together with the calculated OEE value. According to the calculation on the table, there were 3585 minutes planned production time for the machine P64 in a week time. Besides of this, according to the records there were 1185 minutes of downtime happened. The main subjects leading to machine downtime, which was showed at Graphic 2, were as follows; the first main 3 causes had 24% which were adjustment time, scheduled maintenance time and supply&filling time, the second main cause had 18% which was setup time and the third main cause had 10% which was unplanned downtime. For this reason, the availability value was calculated as 68% and according to the daily production quantities and cycle time data it was determined that the machine performance was used at a rate of 48% over the remaining operating time. The quality ratio was calculated as 100% because during the week there were 48571 parts produced and only 21 parts scrapped. The OEE value of 68% of availability, 48% of performance and 100% of quality was 33%.

25

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

26

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

When the data obtained from Prowmes were analyzed, it was seen that the OEE value calculated as 33%, same with the first calculation, which was actually the result of cycle time and production quantities (See Table - 3). However, there were serious differences when the main factors of OEE were examined. The first of these differences and the most important one, the availability value was calculated as 38%, which actually showed that the downtime data provided by the operator did not fully reflect the truth. According to the data obtained from Prowmes, the total downtime was 2215 minutes and in this case the downtime analysis became totally different. In 2 different simultaneously collected data group, Prowmes collected 1030 minutes of downtime more, and when Graph 3 was examined, it was seen that chokotes, the performance decreaser which were less than 3 minutes, became together and had 41% of the downtime. So it had the biggest downtime ratio and this was changed all the thoughts which we had at the previous data analysis. Setup and supply&filling time had different percentage in this analysis but when the time period checked at both two data groups, actually they had the same time value. According to the information received from Prowmes, the performance value was 86% and the quality value was seen as 100% again because of the production and scrap quantities were same. Graph 3– Machine Time Information (Prowmes Graph)

Table 3 – Online OEE Rate Of P64 In Daily, Weekly, Monthly Bases (Prowmes Graph)

27

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Graph 4 – Weekly OEE (Prowmes Graph)

Graph 5 – Downtime Rates (According to the Data from Prowmes)

28

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

5. Conclusion and Recommendations In the light of the analysis and the resulting data, the firm requested selecting a project issue to be increased the OEE value. The demand of the firm was selecting a project topic that could reduce the current duration of one of the downtime cause by 50% and provided maximum OEE increasement. According to the information obtained from the downtime follow-up forms, it was concluded that firstly it was necessary to make improvements in adjustment time, supply& filling time and planned maintenance. Because of supply&filling time could be improved easily, 50% of improvement target calculated. There was 285 minutes of supply&filling time at Table 2. With this improvement study 142 minutes regained and supply&filling time decreased to 143 minutes. After this improvement study, the total downtime decreased from 1185 minutes to 1043 minutes and availability value increased from 68% to 72%. When the performance and quality value were same, the OEE value increased from 33% to 35%. So totally 2% of OEE value would be increased.

According to the downtime obtained from Prowmes, it was concluded that firstly it was necessary to make improvements in 41% of downtime cause, chokotes. There were 2215 minutes of downtime and 908 minutes of these downtime were chokotes. When 50% of improvement target calculated, chokotes decreased from 908 minutes to 454 minutes. So the total downtime decreased from 2215 minutes to 1761 minutes and availability value increased from 38% to 51%. When the performance and quality value were same, the OEE value increased from 33% to 44%. So totally 11% of OEE value increased. In addition to this, there were no need to fill downtime follow-up forms, collect the forms and enter the data to computer. All these operations take time and time is priceless.

29

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

As a result of this study, it was decided to focus on chokotes, determine reasons and remove the reasons one by one. The target of project was to increase the OEE value at least 11%. The company organized a team with 5 personels from production, toolshop, maintenance, production management and quality departments to achieve this target in one month. After one month, when the OEE calculated again it was seen that they achieved the target by increasing OEE 13%. Even every improvement study contributes more efficiency, the companies which have a low value of OEE, should focus on projects which will be achieved easily with a high increasement of OEE. In order to make the right decisions, it is very important to reach the most accurate data. In today's world, with Industry 4.0 applications, cloud computing systems which provides the accurate data and away from personal interventions have a critical role to obtain the right data. Online data processing, minimum workforce and maximum benefit should be the basic working area of all companies. Otherwise, the companies will lose their competitiveness gradually and with an inevitable ending they will lose their customers.

6. Future Research At the present time, many programs offer data flow and data reporting as in Prowmes example. However, determining whether the product produced is scrap or good without being produced is the basis of the projects that need to be focused and studied after the improvement studies. These cloud computing programs can be integrated with ERP programs and without any data input, the production quantity can be seen automatically. The companies should aim to set up and keep alive an automatic tracking system which is away from personal interventions.

References

1. Alkan, E. (2015). Mikronize maden öğütme tesisinde toplam ekipman etkinliği nin araştırılması (Unpublished postgraduate thesis). Eskişehir Osmangazi University, Turkey. 2. Ayane, N., Gudadhe, M. (2015). Review study on improvement of overall equipment effectiveness in construction equipments, International Journal of Engineering Development and Research, 3(2), 2321- 9939. 3. Görener, A. (2012). Toplam verimli bakım ve ekipman etkinliği, Electronic Journal of Vocational Colleges, 2(1). 4. Görener, A., Yenen, V. Z. (2007). İşletmelerde toplam verimli bakım çalışmaları kapsamında yapılan faaliyetler ve verimliliğe katkıları, İstanbul Ticaret Üniversitesi Fen Bilimleri Dergisi, 6(11), 47-63. 5. Iannone, R., Nenni, M. E. (2013). Managing OEE to optimize factory performance, Operations management, 31-50. InTech. 6. Lahri, V., Pathak, P. (2015). A case study of implementation of overall equipment effectiveness on cnc table type boring & milling machine of a heavy machinery manufacturing industry, IOSR Journal of Mechanical and Civil Engineering, 12(5), 63-70. 7. Maraşlı, H., Kemahlı, H. (2013). Yalın üretim bazlı üretim izleme ve iyileştirme, Kahramanmaraş Sütçü İmam Üniversitesi İktisadi ve İdari Bilimler Fakültesi Dergisi, 3(2), 45-64. 8. Naik, G.R., Raikar, V.A., Naik, G.P. (2015). A simulation model for overall equipment effectiveness of a generic production line, IOSR Journal of Mechanical and Civil Engineering, 12(5), 52-63. 9. Polat, İ. (2014). İşletmelerde toplam ekipman etkinliği (oee)kullanımı ile elektrik enerji tasarrufu (Unpublished postgraduate thesis). Marmara University, Turkey. 10. Sahu, S., Patidar, L., Soni, P.K. (2015). 5S tranfusion to overall equipment effectiveness for enhancing manufacturing productivity, International Research Journal of Engineering and Technology, 2(7), 1211- 1216.

30

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

11. Singh, R., Gohil, A. M., Shah, D. B., Desai, S. (2013). Total productive maintenance (TPM) implementation in a machine shop: a case study, Procedia Engineering, 51, 592-599. 12. Singh, R., Shah, D. B., Gohil, A. M., Shah, M. H. (2013). Overall equipment effectiveness (OEE) calculation -automation through hardware & software development, Procedia Engineering, 51, 579-584. 13. Sohal, A., Olhager, J., O’Neill, P., Prajogo, D. (2010). Implementation of OEE – issues and challenges, Competitive and Sustainable Manufacturing Products and Services, Milano: Poliscript, 1-8. 14. Sönmez, V., Testik, M.C. (2015). Sürekli üretim hatlarında kullanılan makinelerin performanslarının izlenmesi amacıyla geliştirilen bir bilgi sistemi çerçevesi, Savunma Bilimleri Dergisi, 14(2), 111-140. 15. Temiz, İ., Atasoy, E., Sucu, A. (2010). Toplam ekipman etkinliği ve bir uygulama, Dokuz Eylül Üniversitesi Sosyal Bilimler Enstitüsü Dergisi, 12(4), 49-60. 16. Yaşin, M.F. (2014). Küçük ve orta ölçekli işletmelerde toplam ekipman etkinliğinin belirlenmesinde yeni bir yaklaşım: bir ahşap işleme tesisinde uygulama (Unpublished postgraduate thesis). University of Turkish Aeronautical Association, Turkey. 17. Yazgan, H. R., Sarı, Ö., Seri, V. (1998). Toyota üretim sisteminin özellikleri, Sakarya Üniversitesi Fen Bilimleri Enstitüsü Dergisi, 2(2), 129-134. 18. Soysal, M., Pamuk, N.S. (2018) Yeni sanayi devrimi endüstri 4.0 üzerine bir inceleme, Verimlilik Dergisi, (1), 41-66.

31

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Application of Tree-Seed Algorithm to the P-Median Problem İbrahim Miraç Eligüzel1, Eren Özceylan2, Cihan Çetinkaya3

1, 2 Department of Industrial Engineering, Gaziantep University, Turkey [email protected] [email protected] 3Department of Management Information Systems, Adana Science and Technology University, Turkey [email protected]

Abstract

This paper presents an application of tree-seed algorithm (TSA) which is based on the relation between trees and their seeds for the p-median problem. To the best knowledge of the authors, this is the first study which applies TSA to the p-median problem. Different p-median problem instances are generated to show the applicability of TSA. The experimental results are compared with the optimal results obtained by GAMS-CPLEX. The comparisons demonstrate that TSA can find optimal and near optimal values for small and medium size problems, respectively.

Keywords: Tree-seed algorithm, meta-heuristic, p-median problem.

1) Introduction

The P-median problem is one of the well-known problems in facility location field. Numerous versions of the p-median problem are defined in literature and it has been shown that it belongs to the class of NP-hard problems (Lorena and Senne, 2004). To find acceptable solutions for the p-median problem in a reasonable time, genetic algorithm (Herda, 2017); simulated annealing (Chiyoshi and Galvão, 2000); ant colony optimization (De França et al. 2005); tabu search algorithm (Rolland et al. 1997); artificial bee colony (Jayalakshmi and Singh, 2017) and particle swarm optimization (Lin and Guan, 2018) are applied by the researchers. Besides aforementioned well-known meta-heuristic approaches, there have been also other approaches which can be applied or tested for facility location problems. TSA is one of the recent meta-heuristic approaches which is a new nature based meta-heuristic algorithm that has been proposed by Kiran (2015). There have been only few papers which uses TSA in the literature such as structural damage identification problem (Zhao et al. 2017), li-ion battery parameter problem (Chen et al. 2017), power system problem (El-Fergany and Hasanien, 2018) and uncapacitated facility location problem (Cinar and Kiran, 2018).

In this paper, TSA is applied to the p-median problem for the first time in the literature. The rest of the paper is structured as follows. In Section II mathematical model of the p-median problem is given. In Section III the steps of TSA is presented. Computational results are shown in Section IV and finally, in Section V conclusion and suggestion for further research are given.

2) P-Median Problem

The p‐median problem is a network problem that was originally designed for, and has been extensively applied to, facility location. The p-median problem seeks to define the number of p candidate facility (source nodes) to be opened, and which customers (demand nodes) will be assigned to each facility under the minimization of total distance/cost (Reese, 2006). The problem formulation is given as follows (Teixeira and Antunes, 2008):

32

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Decision variables: 1, 푖푓 푠표푢푟푐푒 푛표푑푒 푘 푖푠 푠푒푙푒푐푡푒푑 (∀푘 ∈ 퐾) 푦 = { 푘 0, 표푡ℎ푒푟푤푖푠푒 1, 푖푓 푑푒푚푎푛푑 푛표푑푒 푖 푖푠 푎푠푠푖푔푛푒푑 푡표 푠표푢푟푐푒 푛표푑푒 푘 (∀푖 ∈ 퐾, ∀푘 ∈ 퐾) 푥 = { 푖푘 0, 표푡ℎ푒푟푤푖푠푒

Parameters: 푑푖푘 = Distance between source node k and demand node i 푝 = Number of source node to be opened

Objective function: 푀푖푛 푍 = ∑푖∈퐼 ∑푘∈퐾 푑푖푘푥푖푘 (Eq. 1)

Subject to ∑푘∈퐾 푥푖푘 = 1 ∀푖 ∈ 퐾 (Eq. 2) 푥푖푘 ≤ 푦푘 ∀푖 ∈ 퐾, ∀푘 ∈ 퐾 (Eq. 3) ∑푘∈퐾 푦푘 = 푝 (Eq. 4) 푦푘, 푥푖푘 ∈ {0, 1} ∀푖 ∈ 퐾, ∀푘 ∈ 퐾 (Eq. 5)

The objective function (1) is to minimize total distance. Constraint (2) provides the assignment of each demand node to a source node, while Constraint (3) provides the assignment of demand nodes to the opened source nodes. Constraint (4) determines the number of source nodes which should be opened. Constraint (5) is the constraint of the decision variables.

3) The Tree-Seed Algorithm

TSA is the population based meta-heuristic algorithm which is inspired from trees and their seeds. Initialization, seed production mechanism and replacement procedure are the main parts that algorithm consists of. In the initialization part, trees are placed in the search space and their fitness value is calculated. After the initialization part, seeds number for each tree is decided and comparison between seeds and tree is done with respect to their fitness value. In last part, according to fitness value, comparison is made between trees and their seeds. After comparison, if the seed has better fitness value, seed takes the place of parent tree. Otherwise, parent tree preserve its place. Another aspect is the parameters that required to be decided according to problem type, parameters can be expressed as ST (control parameter between [0-1] interval), NS (number of seeds produced for tree), D (dimensionality of problem) and max_FEs (referred as iteration) (Cinar and Kiran, 2018). NS and ST used for convergence and exploitation purpose. It can be summarized as “when ST is close to 1, the algorithm provides a fast convergence and this is useful for lower dimensional optimization problems. If NS is a higher number, the effective local search is provided” (Cinar and Kiran, 2018). Following flow chart (Figure 1) demonstrates the working principle of the TSA.

33

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 1 – Flow chart of TSA (Cinar and Kiran, 2018)

First two parts in the flow chart belongs the initialization phase. In that process trees are generated randomly in solution space and parameters according to type of problem is settled. Following fife parts belongs the seed production mechanism phase which includes working procedure of seed production and selecting best seeds to be tree in next generation. Last part is consists of replacing the seed that better value than its parent, with parent tree.

4) Computational Analysis In order to investigate the performance and accuracy of applied TSA, 10 p-median problems with different node sizes are generated and used. All test problems are coded and solved in MATLAB 2018b based on TSA. To make a comparison, test problems are also run using GAMS-CPLEX optimization package program. The parameters of TSA …. We took all of runs on a server with 2.7 GHz Intel Core processor and 2 GB of RAM. The results of two applications are given in Table 1. First three columns

34

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY in Table 1 show the problem name, size and the number of p. While the other four columns show the objective function values and computation times required for GAMS-CPLEX and TSA, respectively; the gap between two objective functions values is given in the last column. The gap is the percentage of the deviation from the optimum solution for the obtained best solution from the algorithm and it is calculated as follows:

푓(푠표푙)−푓(표푝푡) 퐺푎푝 = 푥100 (Eq. 6) 푓(표푝푡) where f(opt) is the optimal solution obtained by GAMS-CPLEX, f(sol) is the best solution obtained by the TSA at the end of the run.

Table 1 – Results of the TSA and GAMS-CPLEX GAMS-CPLEX Tree-seed algorithm Problem # of nodes p Gap Obj. Func. Value CPU Time (s) Obj. Func. Value CPU Time (s) Pm1 100 5 9121 0.62 9121 1.61 0.00 Pm2 100 10 6186 0.88 6186 162.16 0.00 Pm3 100 15 4964 0.93 5544 224.01 11.68 Pm4 100 20 4456 1.00 5047 170.98 13.26 Pm5 100 25 2571 1.00 3080 106.12 19.80 Pm6 200 5 12809 6.24 12809 33.65 0.00 Pm7 200 10 10163 7.75 10324 605.17 1.58 Pm8 200 15 7830 2.30 8537 147.56 9.03 Pm9 200 20 6389 4.19 7388 598.67 15.64 Pm10 200 25 5182 2.63 6403 201.60 23.56

In Table 1, objective function values obtained by GAMS-CPLEX are optimal. According to the Table 1, TSA can find the optimal solutions in 3 problems. However, optimal solutions cannot be obtained in rest of the 7 problems. The gap values between two solutions are shown in the last column of Table 1. As can be seen, 23.56% gap value which is found in last test problem is the maximum value.

5) Conclusion

In this paper, one of the novel meta-heuristic approaches, namely TSA, is applied to the p-median problem for the first time in the literature. The main objective of this study is to test the applicability of TSA on p-median problem. To do so, small and medium sized test problems are generated. The obtained results are compared with the optimal results obtained by GAMS-CPLEX. While the optimal solutions are achieved for a few test problems, near optimal solutions are obtained for the medium sized test problems. Further work can include modifications of the TSA that can improve for the large sized test problems.

References Chen, W. J., Tan, X. J., Cai, M. (2017). Parameter identification of equivalent circuit models for li-ion batteries based on tree-seed algorithm. IOP Conference Series: Earth and Environmental Science, 73(1), 12–24. Chiyoshi, F., Galvão, R. D. (2000). A statistical analysis of simulated annealing applied to the p-median problem. Annals of Operations Research, 96(1-4), 61–74. Cinar, A. C., Kiran, M. S. (2018). Similarity and logic gate-based tree-seed algorithms for binary optimization. Computers and Industrial Engineering, 115, 631–646. De França, F. O., Von Zuben, F. J., De Castro, L. N. (2005). Max min ant system and capacitated p-medians: Extensions and improved solutions. Informatica, 29(2), 163–171. El-Fergany, A. A., Hasanien, H. M. (2018). Tree-seed algorithm for solving optimal power flow problem in large- scale power systems incorporating validations and comparisons. Applied Soft Computing, 64, 307–316. Herda, M. (2017). Parallel genetic algorithm for capacitated p-median problem. Procedia Engineering, 192, 313– 317. Jayalakshmi, B., Singh, A. (2017). A hybrid artificial bee colony algorithm for the p-median problem with positive/negative weights. Opsearch, 54(1), 67–93.

35

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Kiran, M. S. (2015). TSA: Tree-seed algorithm for continuous optimization. Expert Systems with Applications, 42(19), 6686–6698. Lin, G., Guan, J. (2018). A hybrid binary particle swarm optimization for the obnoxious p-median problem. Information Sciences, 425, 1–17. Lorena, L. A., Senne, E. L. (2004). A column generation approach to capacitated p-median problems. Computers & Operations Research, 31(6), 863–876. Reese, J. (2006). Solution methods for the p-median problem: An annotated bibliography. Networks, 48(3), 125– 142. Rolland, E., Schilling, D. A., Current, J. R. (1997). An efficient tabu search procedure for the p-median problem. European Journal of Operational Research, 96(2), 329–342. Teixeira, J. C., Antunes, A. P. (2008). A hierarchical location model for public facility planning. European Journal of Operational Research, 185(1), 92–104. Zhao, Y., Liu, J., Lyu, Z., Ding, Z. (2017). Structural damage identification based on residual vectors and tree- seed algorithm. Zhongshan Daxue Xuebao, 56(4), 46–50.

36

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Artificial Intelligence And Social Transformation Invited Speaker Prof. Dr. Ercan Öztemel Marmara University EXTENDED ABSTRACT The aim of the presentation is to provide some information regarding industrial and social transformation as well as artificial intelligence with its respective effects on the society. The presentation will also provide some foresights about the progress towards the future. Everything is prone to change. There is one thing that ca not change. That is the CHANGE itself. The first industrial revolution appeared towards the end of 17th century. This was the invention of the first machine which can work using steam power. This became the cause of a remarkable social changes. This machining process resulted with industrial society which became the main source of development especially after the invention of electricity and mass production capabilities. With the introduction of computers and progress on information technology made it possible for the enterprises to device automatic machines and software operations. After 1980’s, artificial intelligence started become popular. With this technology, it was and is possible to generate machines which can self behave and autonomy started to play a main role in industrial applications. Along this line, sensor technologies and information networking for communication has increasingly progressed and made it possible to generate cyber-physical systems. Additionally, internet of things (IoT), virtualization, and big data applications surpasses other systems yielded unmanned factories and systems. This changed manufacturing vision from “energy-machine-money-material” to “product- intelligence- information- communication network”. It seems that the progress along this line, the society will seem to converge “knowledge society” and from there on to “wisdom society”. There are various factors driving the social and technological change over years. Some of them are listed below. ➢ Innovation ➢ Customer expectations ➢ Rapid response ➢ Enhanced security ➢ Digital marketing & consumption ➢ New business models ➢ Global competitiveness ➢ Digital assistants ➢ Expensive labour costs ➢ Cheap labor in the east

Artificial Intelligence (AI) studies are heavily effects the change and become the main source of new systems and products. Sin AI makes it possible to learn from experience, apply knowledge from experience, handle complex situations, solve problems even when important information is missing, determine what is important, react quickly and correctly to a new situation, understand visual images, process and manipulate symbols, be creative and imaginative, use heuristics to solve problems, focus the attention and prioritization, make decisions under uncertainty and perform reasoning about the events based on the perceived knowledge from the environment; it will be possible to generate more and more intelligent, autonomy and self-behaving systems. This naturally makes it easy to generate machines that can analyze, establish relations, make inferences and solve problems with the help of

37

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Expert Systems of AI, learn events using Artificial Neural Network technology of AI, solve complex combinatorial problems with Genetic Algorithms of AI, understand words, and perform computation using words under the philosophy of Fuzzy Logic of AI, walk over stairs, play football, answer questions, communicate each other through utilizing Intelligent Agents of AI, read text, understand and teach Natural Language as well as observe, perceive, priorities, focus attention over the range of events occurring surroundings. With understanding the progress along this line, it seems that the machine will soon be able to speak to each other over exchanging information over computers network (Knowledge Protocols), work for the same goals by modelling sensor and goals, get socialized, cooperate, and help each other by embedding emotional Intelligence techniques, teach each other and let them convert the learning ability by employing Intelligent Tutoring Systems and perform R&D and do innovation Modelling Scientific Discoveries. Social life is also heavily affected by the progress on AI. Some new technologies such as cloud, IoT, Big data systems, 3D printing and blockchain etc. are emerged. Those generated completely new systems which are more effective and more functional in every area of life from transportation to health, from agriculture to manufacturing, from public administration to trade, from energy to education etc. By using these technologies in those areas of life make it possible to generate AI products implanted to human body, AI robots which can perform the role of decision board members in meetings, wearable internets, smart city applications making the life easy for human, vision system utilizing augmented reality for real time decision making, 3D printed products even organs and driverless vehicles on the streets. There are numerous examples which can be listed here. AI is driving the society to emerge towards knowledge-based society through focusing the attention on competition of imagination, learning, independent thinking, creativity and invention as well as totally effective competition of innovation. This progress also encourages the futurist to make prediction about the future societies. Some of the interesting claims along this line are listed below. ➢ Innovative brands experience brand is considered to have a value appreciation 9 times more of today brands. ➢ Self-driving vehicles is thought to save millions of lives each year. ➢ Over 100 million consumers is expected to shop in augmented reality by 2020. ➢ Artificial intelligence could be used predict where epidemics will happen (%87) ➢ 75% of workers is going to be employing intelligent personal assistants, by the 2020. ➢ Of the children entering primary school today, 65% is estimated to end up working in completely new job types that do not yet exist. ➢ İt is forecasted the blockchains could store as much as 10% of global GDP, by 2027. ➢ Some of the human organs is expected to be biologically 3D-printed on demand ➢ ..

The change is continuing to exist even in more rapid speed. It should be well-understood that digital world can upend business models and enterprises must be open to radical reinvention to find new, significant and sustainable sources of revenue and must be prepared to tear themselves away from routine thinking and behavior. This will definitely change business orientation which can bring some risk together. The processes should be redesigned and thoughtfully to be de-risked. Establishing innovation and competitiveness strategy should be encouraged and implemented with utmost attention and the employees should be trained for better understanding the change (Long life learning). The final remark on utilizing AI to foster the social change, the transformation as whole society should to be the case and all should do every effort to cope with it.

38

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Brain Computer Interface Chess Game Platform Gizem Karaca, Yakup Kutlu

Department of Computer Engineering, Iskenderun Technical University, Turkey [email protected] [email protected]

Abstract

ALS patients who have a neural system-related disease, people without limbs, people who have lost their ability to move because of a particular accident or disease may not be treated but this study is aimed to be foundation for other studies to help them to continue their social lives. BMI systems are being developed as a tool / hardware to help individuals with reduced mobility capabilities. In this study, BBA based chess platform is designed. With this study it is desired both to develop a program to be able to support patients with disabilities so that they can play chess without any difficulty and to help people who do not have any disabilities or people who just want psychological treatment. The system designed in this context consists of three main structures: Chess module, BMI module and Robot arm module. These structures were evaluated separately. Later they will be used as a single system. The first structure consists of the GUI part, the interface that will be used to play chess. In this interface Chess Engine is also used, an interface is designed in which two people can play mutual or play with the Engine. The second structure consists of taking the EEG (Electroencephalography) signals of the people with the help of EMOTIV Epoc + device and separating the received signals by certain operations and classifications. Finally the robot arm module after specifying the desired motion according to the received information the joint angles are going to be calculated by inverse kinematic calculation techniques and it will be activated in a controlled way on the platform. With the platform that is going to be created, it is thought that disabled individuals can play chess with EEG signals without any obstacles and this can be implemented with the help of robot arm. Unlike previous studies, it is thought to examine the effect of EEG signals on performance in both silent and noisy environments. In a social environment like a chess tournament, it can not be expected to people to be quiet so it is desired to create a system that can work in such environments as well.

Keywords: Brain Computer Interface (BCI), Chess, BCI Chess, Brain Machine Interface (BMI), Robot Arm, Electroencephalography (EEG).

1.Introduction

ALS disease is a disease caused by the neural system. Since ALS disease is a disease related to the neural system, the communication between the muscles and the brain cannot be achieved smoothly, so people with this disease lose their ability to move slowly after a certain period by slowly losing their muscle movements. Some people with disabilities are born without limbs. And this is due to a hereditary disease can not have a chance to be treated like. Such people, although they have requested, have the need for the help of third parties to play games such as backgammon, chess, etc. due to their physical obstacles. For this reason, they do not make such requests with the concern of not being a burden to anyone. Chess is a strategy and intelligence game that requires individual skill from very old times to the present day. With this study, it is tried to make a study which can be social activity for individuals who have restricted movement abilities due to a disease, traumatic event or an accident, and can help them to be rehabilitated by playing chess, to look at life from a positive point of view as well as their quality of life.

Brain Machine Interface (BMI) is a system developed to improve the quality of life of individuals who have a certain disease, because of an accident or a congenital mobility limit, and to better communicate with their social environment. In these systems, signals received from the brain can be sent to the

39

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY appropriate devices so that the robot can be moved, the machine is operated or a process in the smart home system can be done by disabled individuals.

Robot arm works are designed to enable faster production in factories, but nowadays, prosthetic hand, foot or a device is made as a device to facilitate lives for disabled individuals. The humanoid robots produced by using Artificial Intelligence or Deep Learning are also among the most recent robot studies.

2. Summary of the literature

The aim of the study by Maruthappan et al. (2011) is to give people who have lost their physical ability a chance to play chess with computer systems that support Berlin Brain Computer interface. In the literature, Maruthappan et al. (2011) argued that many EEG reception methods could be performed. Visual Evoked Potential (VEP), Slow Cortical Potentials (SCP), P300 Evoked Potentials, Spontaneous Rhythmic Activity, Cortical Neuron Activity, Sensory Motor Rhythm Signals (SMR), Thought- Translation Device (TTD), Cross-Correlation Template are the methods that have been tried and proven.

P300 row / column speller, Java chess game, GNU WinBoard, XBoard protocol and by Enginee the study of the interface is tested. Brain computer interface chess based 8x8 P300 printer designed in the study of Toersche et al. (2007) P300 speller, P300 cell speller (P300 cell printer), BCI 2000, BCI input paradigm, P300 row / column speller, Java chess game, GNU WinBoard, XBoard protocol, by using Enginee the study of the interface is created and tested (Toersche, 2007).

The main focus of another study is on structural and functional neuroimaging techniques divided into two main groups. BCI and Brain Imaging Techniques were studied in the interface. Emotiv EPOC Neuroheadset with interface created in this application is explained with the application that allows to play chess with the help of (Vokorokos, 2014).

Marshall et al. (2013) evaluated how the most appropriate playing style and type of play were affected by the BCI paradigm with the interaction of noninvasive BCI-based computer games and BCI control strategies. No separate coding or signal processing was specifically performed in the study. Good and bad aspects of the methods used in the previous studies were evaluated. Games such as strategy games, puzzle games and attack games are evaluated within the scope of BCI usage.

Kaplan et al. (2013) analyzes the deficiencies of BCI in gaming applications. These shortcomings and techniques show the solutions that exist on several different games. In addition, new approaches are proposed to improve P300 BCI accuracy and flexibility in the more general P300 BCI study. In this study, BCI Chess is discussed and the laboratory developed robotic arm study is mentioned. Since chess will be used as interface, studies about the development of chess game in the literature are investigated (Janko, 2016; Newell, 1988; Nabiyev, 2015). As an application of Artificial Intelligence, Hajari et al. (2014) developed a chessboard with a sensor to play first. Then, chess is displayed on the screen and divided into two main sections: the detection of extra-ordinary movements and the best possible actions to be taken. In literature, there are different speller models used for P300 or alternative BCI techniques. Grouping as shown in Figure 1 (Corley, 2012), blocking (Capati, 2016) are some examples of these.

40

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

a)

b) Figure 1. Speller application models a) Grouping Method (Corley, 2012) b) Blocking Method (Capati, 2016)

In the study of Yoon et al. (2018) a different time analysis of ERP spell was performed and the best was determined. The studies that is made about robotic arms, ensuring the communication and control of devices like wheelchairs by BCI have been examined. (Bhat, 2013; Asensio- Cubero, 2016; Bell, 2008; Onose, 2012). BMI logic, basic principles, devices (robotic arms, wheelchairs, etc.) to communicate and control the articles on the research (Lobel, 2014; Ushiba,

41

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

2016; Menon, 2009; Tonet, 2008; Dura-Bernal, 2014; Hortal, 2015; Lebedev, 2006). In the study of Luqman et al. (2016) 4-degree freedom (DOF) chess recognition system based on robotic manipulator and computerized vision was created. The robot system consists of a system which can play chess games independently of human or other robotic system.

In this study, the interface will be formed based on the studies done in advance. After the creation of the interface, a structure will be created to provide communication between the robot arm and the interface. With the signals transmitted, the robot arm will be performed on the platform where the desired movement is created. These individually constructed structures will be turned into a single system that works together. After this system is created, unlike previous studies, the effect of silent and noisy EEG signals on performance is considered. People cannot be expected to be silent in chess tournaments, games rooms or in a social environment. For this reason, this study tries to create the basis of a study which can help the individuals with restricted mobility to interfere in society without recognizing obstacles when they are with their family or in a crowded social environment or in a chess tournament in which they apply with confidence. Considering this situation, it is desired to create a system that can work in such environments.

3. Method The proposed brain computer interface based chess game platform consists of three main structures interacting with one as shown in Figure 2: Chess Module, BMI Module and Robot Arm Module.

Figure. 2. Recommended brain computer interface based chess game platform structure

3.1. Chess Module

42

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

This module is created from the interface to play chess. The interface, which includes chess rules, consists of an interface in which two people can play mutual or a against the computer. The game will be a user-friendly interface that will show the moves played.

3.2. BMI Module

In the first stage of this module creation, data collection and careful labeling of these records will be ensured. Then the necessary pretreatment and analysis will be performed and the attributes will be extracted. In order to determine the most significant of these attributes, attribute selection algorithms will be applied. The classification / recognition system will be established with the determination of the most significant attributes. At this stage, different models such as multilayer networks, decision support machines, and the nearest neighbor algorithm will be used to determine the best performance. The following figure shows the flow chart of the Module.

Figure.3. Pattern Recognition Process Steps

3.2.1. Collection of Data

In this study, first EEG signals will be taken from the individuals. The signals shall be taken in different ways in the form of the selection of the coordinates of the moves after consideration of the position on the chessboard or the display of the images of the stone. Thus, different methods such as ERP, P300 Speller (Maruthappan,2011; Toersche,2007; Vokorokos,2014; Marshall,2013) will be tried to find the method to achieve more accurate and faster results. A database consisting of EEG signals will be used to detect stones during the chess game and to determine possible moves that can be made depending on the stone. The EEG signals in the generated data set will be examined in detail and various signal processing techniques will be used.

A 14-channel EMOTIV EPOC device will be used to receive signals to generate the database to be used in the study. The device designed as a hood shall be placed at the head of the volunteer. First of all, it is determined that the device receives a proper signal from 14 channels on the control panel and the recording process is started. During recording, the person will be recorded in a quiet environment in such a way that his hands rest on his / her feet while sitting properly. It will be taken into account that the person is not disturbed at the time of registration and that he / she does not experience any problems related to his / her lack of sleep. Because if the person blinks or moves too much or if he/she is sleepless than the signals will be more noisy and distorted. Therefore a questionnaire will be made before the

43

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY registration to make sure that the person has no discomfort and then a record will be taken. The person's registration will be taken at intervals of 2-3 seconds. During this time period, the person will rest for 2 seconds by resting his / her eyes and then the record will be recorded as appropriate. This process will end after receiving 10 entries of different records from the person. The same procedure will be carried out in a crowded environment or in a noisy environment without any movement of the person, and will be applied differently from the previous works. Thus, the effects of the noisy environment will be analyzed.

3.2.2. Removing an Attribute

With the device attached to the volunteer, it is necessary to determine the chess pieces and to differentiate the brain signals formed for each movement of the stone by subtracting the attribute from the signals generated during the thought of the person. The changes and movements of the EEG signals will be tried to determined and the signals that will be used are determined according to the location of the stone and to the place stone will be sent. It is tried to obtain the best result first by examining the stone which is chosen and the movement they want to do with the stone on the system.

EEG signals consist mainly of frequency ranges. For this reason, raw signals are Delta (0.1Hz- 3Hz), Theta (4Hz- 7Hz), Alpha (8Hz- 12Hz), Low Beta (12Hz- 15Hz), Mid-range Beta (0-12 Hz) before EEG data analysis and classification. 16Hz- 20Hz) and High Beta (21Hz to 30Hz). The band-pass is filtered through a zero-phase filter to allow for the next step. The data obtained in some studies are directly used as raw data without any signal processing method.

Here, the data will first be divided into frequency ranges such as delta, theta, alpha and beta. Then, depending on the divided intervals, each interval will be separated separately by applying the method of extraction, and will be passed to the next process by taking the necessary intervals.

3.2.3. Classification

In the study, according to the brain signals received from 14 channels, the necessary signals will be separated and the signal indicated by each signal will be perceived correctly and will be realized on the platform with the robot arm. Once the brain signals from 14 channels have been subjected to various processes, the parameters with the characteristic feature will be extracted. Those who will maximize the performance of the system will be selected. The system will be modeled with the selected attributes. Artificial Neural Networks (ANN), Support Vector machines (SVM), K-NN (k-nearest neighbor) methods such as the system by making the most appropriate, fast and optimum results will be tried to find the algorithm.

The signals received from the same person may be different from the signal received for the same stone or location as this study examines the signals generated during thinking. In order to minimize the confusion in the data and to distinguish the signals correctly, firstly the classification of the signals connected to the person will be followed, followed by the classification of the signals occurring while considering the same location or stone from different persons. With this classification process made from person to general, the signals generated during thinking will be classified more accurately. The signals are taught by the robot arm and the robot is supposed to perceive and display the signals with the same purpose.

3.3. Robot Arm Module

Finally, the robot arm module will determine the desired movement according to the transmitted signal and the joint angles will be calculated and controlled in a controlled manner. When the signals transmitted to the robot module are classified by the robot arm, then the joint angles are calculated (Iscimen, 2015, Kutlu, 2017a; Kutlu, 2017b) to make the movement smooth and fluid. In this module, angles will be determined by using forward system kinematics, numerical approach in inverse kinematics and closed form approximation methods to move with 6 axis robotic arm using embedded system equipment (Alanoğlu, 2017).

44

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

5. Conclusion

In this study, unlike previous studies, it is thought that the system will be able to work and control properly on the platform even though there is noise in a social environment. Since similar studies in the literature are insufficient to present the results as numerical data, it can be a source for subsequent studies and it is thought to be a resource study that can be used when evaluating.

In the literature, the works on the chess game as well as the work done within the scope of the computer human interface were examined. Based on the methods used and the paths used in previous studies, it is desired to create a system that provides faster and better results.

Once the three main structures have been created separately and their work has been tested, it will become a single system operating in three different environments.

The duration of the EEG signals, the duration of each of the other signals, the frequency values, the interval lengths between the signals will be examined in detail. The similarities and differences between the different EEG signals will be determined. An environment will be created in order to be able to play games depending on the environmental effects, signals and other noises caused by the movement and emotional state of the person. The signals obtained after various operations will be collected in a database.

Then, it will be used as a hardware to perform physical activity in social areas by using Robot Arm Module and BMI Module. In this section, the data received from the BMI Module will be transmitted to the Robot Arm Module and will be performed on the platform with the Robot Arm in accordance with the data received.

Finally, the robot will be calibrated and made to work in the auto-robot environment where Auto-Robot chess is played. The angles and controls are tested to ensure that the movements of the robot can be moved in the correct position and that the movements of the robot move smoothly before being used in the platform. In this section, it is thought to create a Telechess structure similar to the Telesurgery structure. In this structure, the person will command the moves he wants to do with the computer and the robot will move on the platform according to this command he takes.

As a result, it is thought that the disabled individuals can play chess with EEG signals and this can be realized with the help of robot arms. With the methods used, it is aimed to get results faster and to realize these results in real time with the help of robot. This study, which was developed to play only chess, is thought to have the potential to form the basis of many studies that can be developed later.

References

Asensio-Cubero, J., Gan, J. Q., & Palaniappan, R. (2016). Multiresolution analysis over graphs for a motor imagery based online BCI game. Computers in biology and medicine, 68, 21-26.

Bell, C. J., Shenoy, P., Chalodhorn, R., & Rao, R. P. (2008). Control of a humanoid robot by a noninvasive brain– computer interface in humans. Journal of neural engineering, 5(2), 214.

Bhat, V., Hegde, N. M., Vikas, N., & Siddarth, T. C. (2013). Manojavitvam-LabVIEW based Thought Controlled Wheel-Chair for Disabled People. International Journal of Engineering, 6(1), 153-160.

Capati, F. A., Bechelli, R. P., & Castro, M. C. F. (2016, February). Hybrid SSVEP/P300 BCI Keyboard. In Proceedings of the International Joint Conference on Biomedical Engineering Systems and Technologies (pp. 214- 218). SCITEPRESS-Science and Technology Publications, Lda.

Corley, J., Heaton, D., Gray, J., Carver, J. C., & Smith, R. (2012). Brain-Computer Interface virtual keyboard for accessiblity. Imaging and Signal Processing in Health Care and Technology. Human-Computer, 14-16.

Dura-Bernal, S., Chadderdon, G. L., Neymotin, S. A., Francis, J. T., & Lytton, W. W. (2014). Towards a real-time interface between a biomimetic model of sensorimotor cortex and a robotic arm. Pattern recognition letters, 36, 204-212.

45

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Hajari, P., Iyer, R., & Patil, A. (2014). Implementation of Artificial Intelligence for Best Moves using Chessboard.

Hortal, E., Planelles, D., Costa, A., Iáñez, E., Úbeda, A., Azorín, J. M., & Fernández, E. (2015). SVM-based Brain– Machine Interface for controlling a robot arm through four mental tasks. Neurocomputing, 151, 116-121.

Iscimen, B., Atasoy, H., Kutlu, Y., Yildirim, S., & Yildirim, E. (2015). Smart Robot Arm Motion Using Computer Vision. Elektronika Ir Elektrotechnika, 21(6), 3-7. doi: 10.5755/j01.eie.21.6.13749

Janko, V., & Guid, M. (2016). A program for Progressive chess. Theoretical Computer Science, 644, 76-91.

Kaplan, A. Y., Shishkin, S. L., Ganin, I. P., Basyul, I. A., & Zhigalov, A. Y. (2013). Adapting the P300-based brain–computer interface for gaming: a review. IEEE Transactions on Computational Intelligence and AI in Games, 5(2), 141-149.

Kutlu, Y., Alanoğlu, Z., Yeniad, M, (2017). Raspberry Pi Based Intelligent Robot: Recognize and Place Puzzle Objects. International Advanced Researches Engineering Congress-2017

Kutlu, Y., Alanoğlu, Z, Gökçen, A, Yeniad M. (2017). Raspberry Pi Kullanarak Robot Kol ile Bilgisayar Görme Uygulaması. Akıllı Sistemlerde Yenilikler ve Uygulamaları (ASYU2017) Sempozyumu

Lebedev, M. A., & Nicolelis, M. A. (2006). Brain–machine interfaces: past, present and future. TRENDS in Neurosciences, 29(9), 536-546.

Lobel, D. A., & Lee, K. H. (2014, May). Brain machine interface and limb reanimation technologies: restoring function after spinal cord injury through development of a bypass system. In Mayo Clinic Proceedings (Vol. 89, No. 5, pp. 708-714). Elsevier.

Luqman, H. M., & Zaffar, M. (2016, May). Chess Brain and Autonomous Chess Playing Robotic System. In Autonomous Robot Systems and Competitions (ICARSC), 2016 International Conference on (pp. 211-216). IEEE.

Marshall, D., Coyle, D., Wilson, S., & Callaghan, M. (2013). Games, gameplay, and BCI: the state of the art. IEEE Transactions on Computational Intelligence and AI in Games, 5(2), 82-99.

Maruthappan, N., Iyengar, N., & Patel, P. S. (2011). Brain chess—Playing chess using brain computer interface. In ICBMG 2011 Workshop, IPCSIT (Vol. 20, pp. 183-191).

Menon, C., de Negueruela, C., Millán, J. D. R., Tonet, O., Carpi, F., Broschart, M., ... & Citi, L. (2009). Prospects of brain–machine interfaces for space system control. Acta Astronautica, 64(4), 448-456.

Nabiyev, V. V. Satranç Metinlerinin İncelenmesiyle Çeşitli Notasyonlar Arasında Uyumun Sağlanması Providing Harmonization Among Different Notations in Chess Readings.

Newell, A., Shaw, J. C., & Simon, H. A. (1988). Chess-playing programs and the problem of complexity. In Computer games I(pp. 89-115). Springer, New York, NY.

Onose, G., Grozea, C., Anghelescu, A., Daia, C., Sinescu, C. J., Ciurea, A. V., ... & Popescu, C. (2012). On the feasibility of using motor imagery EEG-based brain–computer interface in chronic tetraplegics for assistive robotic arm control: a clinical test and long-term post-trial follow-up. Spinal cord, 50(8), 599.

Toersche, H. (2007). Designing a brain-computer interface to chess. In 7th Twente Student Conference on IT (pp. 1-9).

Tonet, O., Marinelli, M., Citi, L., Rossini, P. M., Rossini, L., Megali, G., & Dario, P. (2008). Defining brain– machine interface applications by matching interface performance with device requirements. Journal of neuroscience methods, 167(1), 91-104.

Ushiba, J., & Soekadar, S. R. (2016). Brain–machine interfaces for rehabilitation of poststroke hemiplegia. In Progress in brain research (Vol. 228, pp. 163-183). Elsevier.

Vokorokos, L., Adam, N., & Madoš, B. (2014). Non-Invasive Brain Imaging Technique for Playing Chess with Brain Computer Interface. International Journal of Computer and Information Technology, 3(5), 877-882.

46

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Yoon, J., Lee, J., & Whang, M. (2018). Spatial and Time Domain Feature of ERP Speller System Extracted via Convolutional Neural Network. Computational intelligence and neuroscience, 2018.

47

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Competition in Statistical Software Market for ‘AI’ Studies Vahit Calisir, Domenico Belli

Department of Barbaros Hayrettin Faculty of Naval Architecture And Maritime Iskenderun Technical University, Turkey

[email protected], [email protected]

Abstract Open source programs are now having more attraction than license (with high price) required statistical package programs. However, the question is that do package programs lose their prestige because of economic reasons or because of being technically inadequate? This study is focusing on the advantages and disadvantages of both open source programs and licensed package programs. Because, Artificial Intelligence (AI) fully depends on critical thinking and reasoning for qualified “Quantitative Decision Making” which makes such computing tools valuable. This assertion’s main aim is to give a perspective over the future of statistical software tools in terms of AI studies.

Keywords: Open Source Programs, Statistical Packages, Artificial Intelligence, Quantitative Decision Making Introduction While world is running to an unknown future, humankind is still trying to keep searching unknown and/or undiscovered truths/objects…etc. Up to the invention of computers and internet, every study was being done with human intelligence. Now, a new concept, called “Artificial Intelligence” is the partner of Human Intelligence, may be only for now. As its name indicates, AI is an intelligence form, which is a type of simulation of human intelligence. Human intelligence has abilities such as learning, reasoning, self-correction …etc. AI’s difference is that machines (computer systems) are assisting these abilities. Expert systems, object recognition, machine vision, speech recognition are examples of such abilities of AI. In problem solving what humans do is that they put intelligence to the center and apply what they learnt during their life. It is somehow similar in AI applications. AI use agents that try to change the its internal state to handle a problem with a sequence of actions (Ekbia, 2010). Agent is the key that finds the most appropriate solution for the problem. Agents move through problems, they first analyze the situation and then decide best choice. AI is beyond the statistical description and inferences but statistics is a part of AI. This study will focus on that part of AI and the competition among tools. The Role of Statistical Tools in AI Applications Imagine that a robot or machine is taking risk or deciding with its own initiative to anything that it has nothing about it. Is it possible? For now, no! ‘AI’ needs to calculate the failure probability for risk assessment. If you do not code it, they can behave without risk assessment but it will not be risk taking or using initiative, it will be a procedure for ‘AI’ application not to take into account risks. However, this is not the subject of that study. Does artificial intelligence in effort to implement pre-defined information on computers while human intelligence can discover? According to Pearl et al. (1991), it is true for ‘AI’ that they do it through statistical methods, Bayesian methods in particular(Pearl et al., 1991).

48

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

How machine learns? Machine Learning is a new concept but it stands in the middle of almost all ‘AI’ applications. Learning is a process. A human being can use his/her common sense to separate meaningless conclusions in learning process but for machines it is up to the crisp principles that we provide (code) to protect the program from reaching senseless or useless conclusions. (Shalev-Shwartz & Ben-David, 2013). What machines do for learning is to use the above-mentioned crisp principles as protocols for the cases defined by human beings (coders). These principles are mostly based on probability and statistics. Quantitative Decision Making with Statistical Tools and Applications in ‘AI’ Intelligence agents’ main mission is to handle the complex real world problems under uncertainty. Statistical AI focuses on uncertainty while logical AI focuses on complexity (Domingos et al., 2016). What statistics does in AI is to mitigate uncertainty as much as possible. There are several examples of statistics usage in AI. To include all is impossible but mainly we can consider two targets for AI applications. These are; to describe the situation and to predict the probability of the case. For example, a hydroelectricity power plant is located in a basin and for scheduling and management of the reservoir needs to predict future water inflows and outflows. Wang et all (2009), developed a prediction model for hydrological forecasting in their recent studies (Wang, Chau, Cheng, & Qiu, 2009). Another example is the study called “Predictive Statistics and Artificial Intelligence in the U.S. National Cancer Institute’s Drug Discovery Program for Cancer and AIDS” of Weinstein et al (1994). This study is about to introduce a package program called “DISCOVERY” developed by the authors. This program integrates the matrix data for the recognition (Weinstein et al., 1994). Gören et al., (2011) studied on the quantity of municipal solid waste (MSW) generated in Istanbul between 1996 and 2008 and they used statistical models such as linear regression, nonlinear regression, and time series and radial basis network, one of artificial intelligence techniques (Gören et al., 2011) In 2004, Hamzaçebi & Kutay tried to predict the energy consumption of Turkey with Artificial Neural Networks. In their study, the results come by artificial neural networks are compared to those obtained by Box-Jenkins models and regression technique. According to their studies results, artificial neural network is an effective tool for forecasting of electric energy consumption (Hamzaçebi & Kutay, 2004). All these examples can be augmented. However, in this study, the effectiveness of the statistical computing tools in AI will be discussed rather than the application examples of statistics in AI. Licensed Package Programs There are many license required package programs in the market. MS EXCEL, SPSS, STATA, SAS…etc are the examples which are known well by the users. These programs common point is that they do not require coding but they have internal procedures to initiate the analysis processes. For example, if you want to make a regression analysis in SPSS you should use the toolbar, find the regression menu and insert the target variables as required in the pop up screen, which appear on the screen, and then you can see the results in another screen (log page).

49

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 1. A Regression Example from SPSS Menu

Figure 2. A log page screenshot after Regression Command from Toolbar It is fast and accurate. All logs can be saved in a file for future use. STATA and SAS is similar. Their main powerful side is to implement analysis without any coding knowledge and they are practical. MS EXCEL is different. It needs some part of coding which is specifically defined just for EXCEL cells. Importing and exporting data is easily possible. On the other side, after a few usage, these packages become so easy and useful for analysis. Learning possibilities are quite high. In order to disseminate the learning possibilities of such programs, researchers created different methods. For example, Chu et al., (2010) developed an SPSS simulator for teachers and students on web (Chu et al., 2010). The most important weakness of such package program in AI Applications is that to be embedded in another system is hard. If you want to develop a specific AI application and require data analysis inside the application, licensed programs will not be useful. Their data or log files can be imported by some other programs but this will make the system discrete. In brief, licensed packages’ advantages; I- They have simple operations II- They can tell what to do without telling it how to do (Li & Cao, 2010) III- Their procedures are short and fast

50

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Disadvantages; I- They are not free II- They have less possibility to create tailor made analysis rather than menu options III- Import and Export of Data is limited IV- Usage in AI systems as embedded object in the system is not possible V- Big data analysis is not easy due to limitations of the programs Open Source Programs (Languages) Open source programs can be defined as programming languages rather than programs. We can consider R and Python programming languages as the examples. Their first advantage is that anyone who wants to use these languages can easily download the platforms for their operating system without any cost. Learning processes of such languages is long in comparison with licensed programs. They have their own coding syntax. However, their learning possibilities are not less than licensed programs, without computer programming background it will be difficult to use such programs. During analysis processes open source programs have some powerful advantages against licensed programs. For example, if you want to make a descriptive statistical analysis of a variable through licensed programs you should follow the procedures as defined in the program, such as using toolbar, finding related command, determining the variable on pop-up screen and then observe from the log file. If you want to do that in R program, it is enough to write “summary(variable_Name)” on the command line and press enter.

Figure 3. A screenshot from R Program. Summary of the variable “mpg” in “mtcars” dataset.

51

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

In brief, open source programs’ advantages; I- They are free, II- Elastic to create tailor made codes for analysis III- Big data analysis is possible, depends on the abilities of the computer that is used for it. IV- They can be embedded in AI applications V- Importing and exporting data is possible and easier Disadvantages; I- Need to know how to code II- Need to tell how to do to computer to get the analysis III- Slow on large datasets but it depends on the memory of the computer that is used for analysis. Conclusions One can say that it is not fair to compare programs such as SPSS, SAS… with open source programs that they are built for specific aims like social sciences or management…etc. However, the issue is that this study is just for to give an insight for the future of the statistical computing tools, in particular in AI applications. Enterprises and society are not separate from AI in terms of statistical analysis. These packages can remain for individual purposes for fast and flexible analyses. However, if AI is the subject, open source programs will be their main competitors.

Figure 4. Career opportunities’ comparison in Analytics Jobs. Source: https://www.learnunbound.com/articles/sas-vs-r-vs-spss-which-statistical-software-to-learn

52

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 5. Job trends of R and SPSS Source: https://www.learnunbound.com/articles/sas-vs-r-vs-spss-which-statistical-software-to-learn

Figure 6. Comparison table for Statistical Tools’ Advantages and Disadvantages Source: https://www.r-craft.org/r-news/python-r-vs-spss-sas/ As can be seen from the above table, to say that open source programs are about to dominate the analysis environment due to high prices of the package programs.

References

Chu, Y. S., Tseng, S. S., Weng, J. F., Lin, H. Y., Wang, N. C., Liao, A. Y. H., & Su, J. M. (2010). Adaptively learning and assessing SPSS operating skills using online SPSS simulator. Proceedings - International Conference on Technologies and Applications of Artificial Intelligence, TAAI 2010, 398–404. https://doi.org/10.1109/TAAI.2010.70

Domingos, P., Lowd, D., Kok, S., Nath, A., Poon, H., Richardson, M., & Singla, P. (2016). Unifying Logical and Statistical AI. Proceedings of the 31st Annual ACM/IEEE Symposium on Logic in Computer Science - LICS ’16, 1–11. https://doi.org/10.1145/2933575.2935321

Ekbia, H. R. (2010). Fifty years of research in artificial intelligence. Annual Review of Information Science and Technology, 44, 201–242. https://doi.org/10.1002/aris.2010.1440440112

53

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Gören, S., Anil, İ., Camci, F., Engül, B. Ş., Hatice, N., & Mehan, A. (2011). MODELING THE QUANTITY OF MUNICIPAL SOLID WASTE IN ISTANBUL BY USING ARTIFICIAL INTELLIGENCE AND STATISTICAL TECHNIQUES, (212), 165– 175.

Hamzaçebi, C., & Kutay, F. (2004). Yapay sinir aǧlari ile Türkiye elektrik enerjisi Tüketiminin 2010 yilina kadar tahmini. Journal of the Faculty of Engineering and Architecture of Gazi University, 19(3), 227–233. https://doi.org/10.17341/GUMMFD.19292

Li, Y., & Cao, W. (2010). Application of SPSS in the ideological and political education. Proceedings - 2010 International Conference on Artificial Intelligence and Education, ICAIE 2010, 187–190. https://doi.org/10.1109/ICAIE.2010.5641424

Pearl, J., Morgan, M. B., Jowell, S., Genderen, R. Van, Pearl, D., & Medoff, L. (1991). Probabilistic Reasoning in Intelligent Systems. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. https://doi.org/10.1016/0167-9236(92)90038-Q

Shalev-Shwartz, S., & Ben-David, S. (2013). Understanding machine learning: From theory to algorithms. Understanding Machine Learning: From Theory to Algorithms (Vol. 9781107057). https://doi.org/10.1017/CBO9781107298019

Wang, W. C., Chau, K. W., Cheng, C. T., & Qiu, L. (2009). A comparison of performance of several artificial intelligence methods for forecasting monthly discharge time series. Journal of Hydrology, 374(3–4), 294– 306. https://doi.org/10.1016/j.jhydrol.2009.06.019

Weinstein, J. N., Myers, T., Buolamwini, J., Raghavan, K., Van Osdol, W., Licht, J., … Paull, K. D. (1994). Predictive statistics and artificial intelligence in the U.S. National Cancer Institute’s drug discovery program for cancer and AIDS. Stem Cells, 12(1), 13–22. https://doi.org/10.1002/stem.5530120106

54

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Comparison of Investment Options and an Application in Industry 4.0 Adnan Aktepe1,Damla Tunçbilek2,Süleyman Ersöz3,Ali Fırat İnal4,Ahmet Kürşad Türker5

Department of Industrial Engineering, Kırıkkale University, Turkey [email protected] [email protected] [email protected] [email protected]

Abstract

Together with Industry 4.0, also known as the fourth industrial revolution, a change and revolution is developing that will affect all sectors and companies of all sizes. Intelligent technologies and factories stand at the center of this development It is known that all of the forward-looking strategies and policies of the institutions include innovations along with the previous technologies. Faster, higher quality with intelligent and self-managing factories and while an efficient system is taking its place in enterprises, the negative effects on employment are the most discussed issues. These changes and innovations have important changes on the investment and operating costs of the enterprise.

To have enough knowledge about digital technologies in order to make decisions about starting a digital transformation journey define their usage areas and make cost-benefit analysis.In this study, a cost-benefit analysis of a multi-functional positioner robot investment was made to accelerate the welding process in a cauldron production plant. The results of the analysis are indicated in the study.

Keywords: Industry 4.0, Cost-Benefit Analysis, Smart Systems

1. Introduction

Constantly evolving technology, Industrial productivity and efficiency made it possible to significantly increase. Working machines with steam power, to enter the electricity production and robotic automation technologies are becoming widespread after the revolution was triggered by these three. Today, with the development of digital technologies, Industry 4.0 revolution has emerged. İntelligent robots entered our lives with Industry 4.0, big data, the internet of objects, We observe that technologies such as cloud is a very important role in the triggering of this revolution.

The use of new technologies has enabled the transition from labor and capital-oriented production systems to knowledge-oriented production systems. With this change, new standards and forms of employment have emerged. Changes in Technology, professional requirements at the organizational level, the workplace environment, workplace health and safety, and also led to the emergence of qualified elaman needs. Implementation of new systems emerging with technology will have the benefits as well as the costs.

55

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

1.1. Industry 4.0

Technological advances, since the beginning of the industrial revolution, it made it possible to cover three main stages, which indicated a large increase in industrial productivity.

Steam powered machines were used in factories in the late 18th century, mass production possible with electrical energy at the beginning of 20th century, Since the 1970s, automation has become widespread in the industry with electronic and information technologies. Today, we are witnessing the fourth phase of the industrial revolution in which cyber- physical systems and dynamic data processing and value chains are end-to-end.

Figure 1 – The steps of industrial revolutions [12]

In order to increase the efficiency of systems in your future smart factories, many subsystems needs to be integrated with each another. For this integration, machines that are independent of each other must communicate intelligently. Process monitoring can be done by online data which can be obtained from machine and other units. Thus, machines can signal to correct problems with the ability to stop production.

1.2. Cost-Benefit Analysis

Cost-Benefit Analysis (CBA) can be defined as the systematic comparison of all the costs and all the benefits required by the investments in order to determine the highest economic return among various investment alternatives and choosing the most optimal among them.

A theoretical background was created by determining the concept of CBA in 1844 for the first time in J. Dupuit. The contributions of the French economists in 19th century cannot be overlooked in the future. In addition, the contribution of Italian social scientists especially V. Pareto and N. Kaldor and Sir John Hicks in the 1940s cannot be denied in the more advanced dimension, especially in the social development dimension.

56

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

The cost-benefit analysis was initially limited to natural resource projects, but it became available for all public and private sector investments in which costs and benefits in other areas could be monetized. Cost-benefit analysis, It is an analysis method used to compare alternative investment projects. Both the private sector and public institutions both varieties, usually labor in their hands, capital, of limited resources such as machinery-equipment and raw materials are faced with the choice to use the alternative investment field. Resources to be in a certain area, that it would mean giving up investments in other areas, Choosing and ranking among investment projects is a necessity to use scarce resources in the most efficient way. Moreover, while the investment in a project can be measured quantitatively, In cases where the return on this investment is qualitative, the benefit-cost analysis is a useful tool to measure these qualitative returns.

Cost-benefit analysis as a project evaluation method, income expected to be provided during the economic life of the investment, the comparison with the expenditures incurred by the project. In other words, the expected revenues during the economic life span of the investment, is reduced to the present value of the investment (present value) with the help of a certain interest rate and this value, compared with the new product expenditure.

2. Literature Review

When we look at studies on the cost-benefit analysis in the literature, we found a wide variety of examples with different investment projects. Kang et al. [1] have made a comprehensive literature review and designed a system for production scheduling in a factory that manufactures automobiles. Also they developed an sequence-error proofing system to avoid accidental line stops due to incorrect part sequencing and they used cost-benefit analysis to decide that this system should be installed or not. Kim [2] used object tracking technology to improve productivity in a cargo company in Korea. In addition, they made a cost-benefit analysis study by determining the investment cost and the benefits it will provide after installation. Roper et al. [3] proposed a model for easy patient tracking technology into a hospital. In this model, an automated monitoring system can be used to schedule surgery operations and etc. also they have tested this investment by cost-benefit analysis. Erminio and Pilloni [4] conducted a cost-benefit analysis in a blood bank, suggesting that a high technology labeling system should be used as an alternative to current system so that the blood bags are transported more safely and that the blood in different specifications does not mix with each other. Tege and Verma [5] did a cost- benefit analysis of the idea of integrating a control system into the shelves of a store. With this system, empty shelves can be detected immediately or the amount of products stocked on the shelves can be seen. Uckelmann [6] conducted a comprehensive performance appraisal and cost-benefit analysis of the integration of radio frequency identification technology and the Internet of Things in the field of logistics. Khoo and Cheng [7] worked in a company that manufactures electronic parts, to set up a track monitoring system in order to prevent the stealing of some precious parts continuously. They have demonstrated that this system will be profitable for the company in the long run by making cost-benefit analysis of this system. Ustundag and Satoglu [8] used simulation modeling to evaluate the effect of the radio frequency identification technology on a pallet pooling system. They used the simulation output of this application on a pallet supplier and a supply chain consisting of three customers in cost-benefit analysis. Wu et al. [9] focused on the shortcomings of the cost-benefit analysis method by examining many studies on investment projects. They have devised a new methodology based on the idea that the cost-benefit analysis method is not flexible enough because future uncertainties do not add up. They have presented with proof that this method they have designed works reasonably.

57

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

3. Application

This study was carried out in a factory that manufacturing cauldrons in Kırıkkale province. 3 new production lines are planned to be installed in the factory. 2 options are available at this stage. According to the first option, one Semi-Automatic welding machine will be taken on each line. In order to be able to turn the cauldron easily during the welding process, it is necessary to install 1 crane which can be used jointly by 3 lines. In addition, one unskilled labor will be required in each of the production lines.

The most important process for cauldron manufacturing is welding. Fast welding of this welding process is of great importance for the factory. According to the first option, it takes an average of 8 hours to complete the welding process of 1 cauldron. When welding with these welding machines, the worker must also interfere with this process.

At some stages during the welding process, the worker creates dangerous positions while interfering with the conversion of the cauldron. In such cases, various occupational accidents occur. Eye burns caused by work accidents, skin burns, limb losses, lung cancer and related psychological traumas occur. It is the case that the worker cannot come to work as a result of these unwanted events. Due to workers who cannot come to work, there is a loss in the labor force, which is reflected to the factory as a cost.

In the second option we discuss 3 Full-Automatic welding machines with positioners. In this option, 3 unqualified workers do the job done by 1 qualified person is also the subject of the average time to 1 cauldron will be reduced by 1 hour according to the Semi-Automatic system.

The most important feature of positioner welding machine, during the welding process without turning the cauldron. This method has many advantages. During welding, the cauldron can be turned with extreme precision. As a result, the Full-Automatic welding machine can be said to be defect-free if it is working properly. One of the advantages of dialing the cauldron without touching it is to greatly reduce the risk of work accidents. With this new technology, there will not be 3 workers who use their hands during welding. This reduces the risk of loss of limb and burns to zero. In addition, the qualified element will control the source pool remotely and will be a little further away from the harmful gases and the risk of lung cancer will decrease.

Table 1 – Comparison of the two options Semi-Automatic Full-Automatic Labor 3 unqualified worker 1 qualified Production Time per Cauldron 8 hour 7 hour Economic Life Span 5 year 10 year

3.1. Determination of Cost Factors

The first step of Cost-Benefit Analysis is the determination of cost factors. The main identified costs are shown in Table 2.

58

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Table – 2 Cost Factors C1) Initial Investment Costs C1.1) Machine Cost C1.2) Computer Cost C1.3) Software Cost C2) Trainer Costs C3) Maintenance Costs C4) Labor Costs

C1. Initial Investment Costs:

This cost is the initial investment cost. In this context, the amount to be invested in the factory was determined for the establishment of new lines. Machine cost, computer cost and software cost will arise.

C2. Trainer Costs:

It is the money paid to the trainers for the necessary trainings to understand how to use the positioner machine and to continue the process systematically.

C3) Maintenance Costs:

Cost of labor and spare parts paid for maintenance of the positioner machine. It is the cost that occurs in periodic maintenance of the positioner and in unexpected failure situations.

C4) Labor Costs:

1 qualified personnel will work in positioner welding machine. In the Semi-Automatic welding machine, 3 unskilled workers will work. The costs associated with these two situations are labor costs. The costs of qualified personnel and unskilled workers are calculated in more detail in the following sections.

3.2. Determination of Benefit Factors

Another step of Cost-Benefit Analysis is the determination of benefit factors. In this step, the benefits of the investment project will be taken into consideration. The identified benefits are given in Table 3.

59

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Table – 3 Benefit Factors B1) Reduced Welding Time B1.1) Quick response to orders B1.2) Increase in Sales B2) Defect-free and Quality Production B3) Minimizing Industrial Accidents B4) Scrap Value

B1) Reduced Welding Time:

It is known that the welding machine with positioner will make the welding process faster than the Semi-Automatic welding machine. The processing time is 8 hours with the Semi-Automatic machine, it is reduced to 7 hours with Full-Automatic machine and saves 1 hour. With the acceleration of the welding process, the order from the customer can be responded to in a shorter time, thus increasing customer satisfaction. In addition, the amount of sales will increase due to more production.

B2) Defect-free and Quality Production:

With the Full-Automatic welding machine, the errors due to the worker will be minimized and a smooth weld line will be obtained. Depending on quality production, demand and brand value will increase.

B3) Minimizing Industrial Accidents:

When welding with a Semi-Automatic welding machine, the worker must manually intervene in the cauldron. In the case of Full-Automatic welding machine, welding process will be done without touching hands. Therefore, the risk of limb loss and burn will decrease to zero and the risk of lung cancer will be minimized.

B4) Scrap Value:

Both investments will have a scrap value after economic life. A scrap value will be generated for the Full-Automatic machine at the end of 10 years and at the end of 5 years for the Semi-Automatic machine, depending on the machine's wear rates due to the intensive use of the machines.

3.3. Determinition of Break-even Point

The analysis that should be made while making investment decisions is the Break-even Analysis. With Break-even Analysis, it is possible to find out which of the two investment projects will be more efficient in which period. There is a Break-even Point as a result of Break-even Analysis.

The first step of Break-even Analysis is to determine Fixed Expenses (퐹푒). Equation 1 is used to determine Fixed Expenses.

퐹푒 = 퐼 + 푀 − 푆 (Eq. 1)

퐹푒: Fixed expenses

60

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

퐼: Equal Installments of Investment Amount

푀: Maintenance Costs

푆: Net Present Value of Scrap Value

When performing Break-even Analysis, the monetary amount is always calculated by calculating the Net Present Value. The Net Present Value formula used in calculating “퐼” is given in Equation 2.

(1+푖)푛−1 퐴 = 퐼 [ ] (Eq. 2) 1 푖(1+푖)푛

퐴1: Investment Amount

퐼: Equal Installments of Investment Amount

푖: Interest Rate

푛: Economic Life Span

The formula of the “푆” value that we reached using the total scrap value is given in Equation 3.

(1+푖)푛−1 퐴 [ ] (Eq. 3) 2 = 푆 푖

퐴2: Total Scrap Value

푆: Net Present Value of Scrap Value

푖: Interest Rate

푛: Economic Life Span

Table 4 – Available Data Semi-Automatic Full-Automatic Investment Amount (푨ퟏ) 770,000 ₺ 5,040,000 ₺ Economic Life Span (풏) 5 years 10 years Total Scrap Value (푨ퟐ) 110,000 714,000 Maintenance Costs (푴) 115,500 756,000 Interest Rate (풊) 0.22 0.22

While calculating the investment costs, the average price in the market was taken into consideration. Maintenance costs correspond to 15% of the investment costs. The scrap value is the sales price of the machine when the economic life span of the machine is over. Interest rates were taken as 0.22 for both investments. Also for both investment options, Fixed expenses (퐹푒) calculated using together with the data in Table 4, Equation 1 and Equation 2. The results are shown in Table 5.

61

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Table 5 – Fixed expenses Semi-Automatic Full-Automatic Fixed expenses (푭풆) 370,180 ₺ 2,015,756 ₺

In order to be able to calculate the Break-even Point (퐵푝), it is necessary to calculate the Variable Expenses and Fixed Expenses.

푇푐 = 퐹푒 + 푉푒 ∗ 퐵푝 (Eq. 4)

푇푐: Total Cost

퐹푒: Fixed expenses

푉푒: Variable expenses

퐵푝: Break-even Point

Table 6 – Determination of variable expenses Semi-Automatic Full-Automatic Electric Consumption 0.83 ₺/h 1.33 ₺/h Electrode Consumption 30 ₺/h 30 ₺/h Labor Costs 37.5 ₺/h 27.5 ₺/h Variable Expenses (푽풆) 68.33 ₺/h 58.83 ₺/h

The hourly electric consumption for the Semi-Automatic welding machine is calculated as 2kW and the electric consumption for the Full-Automatic welding machine is 3.2kW. The price of 1 kW of electricity is set at 0.4151 ₺/h The hourly electric consumption is shown in Table 6.

Employees work 25 days in a month and 8 hours a day. Semi-Automatic welding machine needs 3 unqualified workers and Full-Automatic welding machine needs the 1 qualified worker. The monthly cost of a qualified worker is 5500 ₺, and the unqualified worker is 2500 ₺. The labor cost is calculated from Table 6.

Along with calculating variable expenses, total cost has become computable. Break-even Point (퐵푝) for both investments using X is calculated below.

2,015,756 + 58.83 * 퐵푝 = 370,180 + 68.33 * 퐵푝 (Eq. 4)

In the above Equation 4, Break-even Point (퐵푝) = 173,218 is calculated.

62

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Graphic 1 – Graphical representation of Break-even Point

“퐵푝” is calculated for comparison of the two investments and shown in Graphic 1. The “퐵푝” point shows which investment will be more efficient after which period. After the sale of 173,218 products, the profit will be started. However, if the Full-Automatic machinery investment is made, the costs after “퐵푝” will increase more slowly. So in summary, applying the Semi-Automatic investment option to the “퐵푝” point and switching to the Full-Automatic investment option after the “퐵푝” point will provide the maximum profit.

4. Results

Accepted as a project analysis and evaluation technique, the CBA is now accepted as a method that many countries use in their decision to make decisions about their investments in developing technology. In this study, CBA analysis was performed and two investment options comparison were made.

With the introduction of Industry 4.0, the fundamental changes that take place in the enterprises are evolving day by day. Intelligent robots taking on human duties means pointing to dangers on employment. The fact that businesses switch to Full-Automatic investment means that they face this danger for unqualified workers as they turn into an opportunity for qualified personnel.

As it is understood from the work done above, because the investment amount of the Full- Automatic welding machine is very high, it cannot pay for itself for many years. If we take this into account, the Semi-Automated machinery can do the same job and also save on the expensive investments without allocating the budget and removing the workers. However, intensive and rigorous work must be carried out to obtain the benefits of the Full-Automatic welding machine from the Semi-Automatic machine. Efforts should be made with OHS in order to eliminate all these situations that will reduce the work accidents to zero and include ergonomic studies and affect the life of the employees. But this does not mean that smart technologies should not be implemented in enterprises. Although smart technologies are very expensive, we have a Computer Integrated Manufacturing (CIM) laboratory in Kırıkkale

63

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

University, Industrial Engineering Department and there are 3 smart robots in our laboratory. CIM laboratory in Kırıkkale, there are very few like and is very useful for educational purposes.

References 1. Kang Y.S., Kim H. & Lee Y.H., Implementation Of An RFID-Based Sequencing-Error-Proofing System For Automotive Manufacturing Logistics, Applied Sciences, 8, 109, 2018. 2. Kim M.S., A Cost-Benefit Analysis Of Case-Level Radio Frequency Identification Tagging-Based System On Logistics Service, Advanced Science Letters, 21, 3324-3328, 2015. 3. Roper K., Sedehi A. & Ashuri B., A Cost-Benefit Case For RFID Implementation In Hospitals: Adapting To Industry Reform, Facilities, 33, 367-388, 2015. 4. Erminio P. & Pilloni M.T., RFID Systems For Risk Reduction In Blood Bags: A Cost-Benefit Analysis, International Journal of Mechanics and Control, 16, 39-58, 2015. 5. Tege S. & Verma D.S., Cost Benefit Analysis Of Shelf Replenishment Of RFID In Retail Outlet, 2012. 6. Uckelmann D., Performance Measurement And Cost Benefit Analysis For RFID And Internet Of Things Implementations In Logistics, Quantifying the Value of RFID and the EPCglobal Architecture Framework in Logistics, 71-100, 2012. 7. Khoo B. & Cheng Y., Using RFID For Anti-Theft In A Chinese Electrical Supply Company: A Cost- Benefit Analysis, IEEE Proceedings of the Wireless Telecommunications Symposium, 1-6, 2011. 8. Ustundag A. & Satoglu Ş., Cost-Benefit Analysis For RFID Based Pallet Pooling Systems, 2010. 9. Wu X., Yue D., Bai J. & Chu C., A Real Options Approach To Strategic RFID Investment Decision, IEEE International Conference on RFID, 314-321, 2009. 10. Isik, A., Organ, G. ve Karayilmazlar, E. (2005). Kamu Maliyesi, Ekin Kitabevi Yayınları, Bursa. 11. AB&R®, Advantages of RFID vs Barcodes, (2018), http://www.atlasrfid.com/jovix-education/auto-id- basics/rfid-vs-barcode. 12. Ozel, M. A., (2016), Endustri 4.0 Nedir, https://www.muhendisbeyinler.net/endustri-4-0-nedir/ 13. Smiley, S., (2014), Monostatic vs Bistatic RFID Systems https://blog.atlasrfidstore.com/ monostatic-vs- bistatic-rfid. 14. Gramlich, E. M. (1981). Cost-Benefit Analysis of Government Programs, Prentice-Hall Int, USA. 15. Dyson, N., (n.d.)., Working in a smart world - Industry 4.0, https://www.tuv-sud.co.uk. 16. Ar Apps for Industry. (n.d.)., http://www.igs.com.ar/ar-apps-for-industry.

64

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Computer-Aided Detection of Pneumonia Disease in Chest X-Rays Using Deep Convolutional Neural Networks Murat Uçar1, Emine Uçar2

1Ministry of National Education, Ataturk Boulevard, Number 98, Ankara, Turkey [email protected] 2Turk Telekom Vocational High School, 2432th Street, Number 44, Ankara, Turkey emineucar@ meb.gov.tr

Abstract Chest X-Rays are most accessible medical imaging technique for diagnosing abnormalities in the heart and lung area. Automatically detecting these abnormalities with high accuracy could greatly enhance real world diagnosis processes. In this work we propose a computer aided diagnosis (CAD) system to identify pneumonia in chest radiography. We have created deep convolutional neural network architecture to detect pneumonia and we have used an advanced version of AlexNet model to compare our results. In this study we have used publicly available Mendeley dataset including of 5,232 chest X- ray images from children. The results indicated that, convolutional neural network model produced the best results with 93.80% accuracy. Convolutional neural network model is followed by AlexNet model with an accuracy of 93.48%. The results obtained here demonstrate that our CNN model and AlexNet model gave almost the same results.

Keywords: Chest radiographs, pneumonia, deep convolutional neural network

1. Introduction The pneumonia is a form of acute respiratory infection that affects the lungs and can be caused by bacteria, virus or fungi [1]. According to the report of World Health Organization, pneumonia killed 920,136 children under 5 years old in 2015, accounting for 16% of all the pediatric deaths (World Health Organization (WHO), 2015). Chest X-rays are recently the best available method for diagnosing pneumonia (World Health Organization (WHO), 2001), playing a critical role in clinical care (Franquet, 2001) and epidemiological studies (Cherian et al., 2005). Computer Aided Diagnosis (CAD) is system that assists doctors in the interpretation of medical images. Development of a CAD system for the evaluation of medical images would increase the productivity of physicians and accessibility of better healthcare services in remote areas. In recent time, deep convolutional neural network (DCNN) has gained popularity given its excellent performance in different image recognition challenges, such as image classification (Krizhevsky, Sutskever, & Hinton, 2012; Simonyan, & Zisserman 2014; Szegedy et al., 2015; He, Zhang, Ren, & Sun, 2016) and semantic segmentation(Mostajabi, Yadollahpour, & Shakhnarovich, 2015; Noh, Hong, & Han,2015; Chen, Papanndreou, & Kokkinos, 2016 ). DCNN is also applied in many medical images processing tasks (Van Ginneken, Setio, Jacobs, & Ciompi, 2015; Li et al., 2014, Roth, Lu, Liu, Yao, & Seff, 2015; Bar et al., 2015; Shin et al., 2016) (Bar et al., 2018) recently.

In this work, we report DCNN based detection of pneumonia disease in chest X-Rays on the publicly available Mendeley dataset. The paper is organized as follows. In section 2 we overview of the related work. In section 3, we describe the dataset and analysis method. Then in section 4, we present our results and compare the result obtained by this network with AlexNet model. Finally, we discuss about the lacking and future possibilities of the presented network.

65

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

2. Related Work Over the past decades, the volume of clinical data in machine-readable form has increased, especially in medical imaging. While previous generations of algorithms have sought to use of this high- dimensional data effectively, modern neural networks have been successful at such tasks. Most successful applications of neural networks to medical images are largely based on convolutional neural networks (ConvNets), which were first proposed in (LeCun et al., 1998). This is no surprise that ConvNets are the basis of the top performing models for natural image understanding. For abnormality detection and segmentation, the most popular variants are UNets from (Ronneberger et al., 2015) and VNets from (Milletari et al., 2016), both built on the idea of fully convolutional neural networks introduced in (Long et al., 2015). Representative examples of neural network-based models from the medical literature for classification are: (Esteva et al., 2017) for skin cancer classification, (Gulshan et al., 2016) for diabetic retinopathy, (Lakhani, & Sundaram, 2017) for pulmonary tuberculosis detection in x-rays, (Huang et al., 2017b) for lung cancer diagnosis with chest CTs. (Rajpurkar et al., 2017) used CheXNet model to detect automatically many lung diseases and (Gordienko et al., 2018) demonstrated efficiency of dimensionality reduction performed by lung segmentation, bone shadow exclusion, and t- distributed stochastic neighbor embedding (t-SNE) techniques for analysis of 2D CXR of lung cancer patients. All of the examples above used 2D or 3D ConvNets and all of them probably achieved near- human level performance in their particular installation.

3. Materials and Methods

a. Dataset

In this study we used publicly available Mendeley dataset including of 5232 chest X-ray images from children. 3883 images contain a pneumonia, and the remaining 1349 are negative for pneumonia.

b. Methods

Convolutional neural network (CNN) is a type of feed forward neural network in machine learning and a CNN usually combines the following five types of layers.

Convolution layers are the main components of a CNN. The layer consists of several filters (aka kernels) that we want to learn during the training face. We usually add an activation layer after the convolution layer, increasing the non-linearity of the network. The ReLu (rectified linear unit) function f(x) = max(0, x) is used as a common practice by researchers (Jarrett, Kavukcuoglu, Ranzato, & LeCun, 2009). The pooling layer is another important concept of CNN, and it performs a non-linear down-sampling operation. Max-pooling is the most commonly used pooling operation, and Average-pooling is also used according to the tasks (Dong, Pan, Zhang, & Xu, 2017).

The last layer of the network is usually a fully-connected layer. After several convolution and pooling layers, the network is ended by one or more fully-connected layers. Neurons in this layer have full connections with the previous layer. There is no spatial information after a fully-connected layer. The loss layer is used to train the neural network. Various loss functions are used for different tasks. For example, softmax loss function is used for classification problem, and sigmoid cross entropy loss is used for predicting some independent probabilities (Dong et al., 2017). Figure 1 shows the basic convolution operation of a CNN.

66

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 1 – Basic CNN architecture showing alternating convolution and pooling layers. The highlighted small boxes are the receptive regions. The connections show implicit hierarchical learning of features (Aloysius & Geetha, 2018).

Results

In this section, the performance of the used methods is evaluated via its effect on the pneumonia detection in chest radiographs. The DCNN was trained in MATLAB machine learning framework on the dataset to predict presence (3883 images) or absence (1349 images) of pneumonia. Several training and validation runs for the DCNN on CXR images from Mendeley database were performed. The results of our runs are shown in Figure 2.

a) b)

c) d) Figure 2 – A comparison of pneumonia detection performance with different models a) Validation accuracy b) Training accuracy c) Validation loss d) Training loss.

67

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

The evident over-training can be observed after comparison of training and validation results, where the averaged and smoothed value of training accuracy is going with epochs to the theoretical maximum of 1 and training loss is going to 0. As the results indicate, DCNN model produced the best results with 93.80% accuracy. DCNN model is followed by AlexNet model with an accuracy of 93.48%. The results obtained here demonstrate that our CNN model and AlexNet model gave almost the same results.

5. Conclusion

In this paper, the detection of pneumonia is illustrated by using deep learning neural networks. We proposed a computer aided diagnosis (CAD) system to identify pneumonia in chest radiography. We have created deep convolutional neural network architecture to detect pneumonia and we have used an advanced version of AlexNet model to compare our results.

In our future work, we want to extent our detection to other types of lung infections like lung nodules, tuberculosis, and cardiomegaly.

References 1. Mcluckie, A. (2009). Respiratory disease and its management. Springer. 57(5):469. 2. World Health Organization (WHO). (2015).Pneumonia. 3. World Health Organization (WHO). (2001). Standardization of interpretation of chest radiographs for the diagnosis of pneumonia in children. 4. Franquet, T. (2001). Imaging of pneumonia:Trends and algorithms. European Respiratory Journal, 18(1):196–208. 5. Cherian, T., Mulholland, E. K., Carlin, J. B., Ostensen, H., Amin, R., De Campo, M., … Steinhoff, M. C. (2005). Standardized interpretation of paediatric chest radiographs for the diagnosis of pneumonia in epidemiological studies. Bulletin of the World Health Organization. 83(5):353–359. 6. Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). 1 ImageNet Classification with Deep Convolutional Neural Networks. Advances In Neural Information Processing Systems (NIPS). pp. 1106-1114. 7. Simonyan, K., & Zisserman A. (2014). Very deep convolutional networks for large-scale image recognition, Computer Science. 8. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., … Rabinovich, A. (2015). Going deeper with convolutions. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR.) 1-9. 9. He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep Residual Learning for Image Recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). ), pp. 770–778. 10. Mostajabi, M., Yadollahpour, P., & Shakhnarovich, G. (2015). Feedforward semantic segmentation with zoom-out features. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition. pp. 3376–3385. 11. Noh, H., Hong, S., & Han, B. (2015). Learning deconvolution network for semantic segmentation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2015), pp. 1520–1528. 12. Chen, L., Papandreou, G., & Kokkinos, I. (2014). [F] Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs. Computer Science , 4 357–361. 13. Van Ginneken, B., Setio, A. A. A., Jacobs, C., & Ciompi, F. (2015). Off-the-shelf convolutional neural network features for pulmonary nodule detection in computed tomography scans. In Proceedings International Symposium on Biomedical Imaging. (ISBI) (2015), 286–289. 14. Li, R., Zhang, W., Suk, H. Il, Wang, L., Li, J., Shen, D., & Ji, S. (2014). Deep learning based imaging data completion for improved brain disease diagnosis. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). 305–312. 15. Roth, H., Lu, L., Liu, J., Yao, J., & Seff, A. (2016). Improving Computer-aided Detection using Convolutional Neural Networks and Random View Aggregation. In IEEE Transactions on Medical Imaging 35(5):1170–1181. 16. Bar, Y., Diamant, I., Wolf, L., Lieberman, S., Konen, E., & Greenspan, H. (2015). Chest pathology identification using deep feature selection with non-medical training. In IEEE International Symposium on Biomedical Imaging (ISBI). 294–297.

68

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

17. Shin, H. C., Roberts, K., Lu, L., Demnerfushman, D., Yao J., & Summers, R.M., (2016). Learning to read chest x-rays: Recurrent neural cascade model for automated image annotation, in IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2497–2506. 18. LeCun, Y., Bottou, L., Bengio, Y., & Haffner P., (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11): 2278–2324. 19. Ronneberger, O., Fischer, P., & Brox, T.,(2015). U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention, 234–241. 20. Milletari, F., Navab, N., & Ahmadi, S. A., (2016). V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: 2016 Fourth International Conference on 3D Vision (3DV). 565–571. 21. Long, J., Shelhamer, E., & Darrell, T., (2015). Fully convolutional networks for semantic segmentation. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 3431–3440. 22. Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M., Blau, H. M., & Thrun, S., (2017). Dermatologist-level classification of skin cancer with deep neural networks. Nature 542(7639):115–118. 23. Gulshan, V., Peng, L., Coram, M., Stumpe, M. C., Wu, D., Narayanaswamy, A., Venugopalan, S., Widner, K., Madams, T., Cuadros, J., et al. (2016). Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. Jama. 316(22):2402–2410. 24. Lakhani, P., & Baskaran, S., (2017). Deep learning at chest radiography: Automated classification of pulmonary tuberculosis by using convolutional neural networks. Radiology 284(2):574–582. 25. Huang, P., Park, S., Yan, R., Lee, J., Chu, L. C., Lin, C. T., Hussien, A., Rathmell, J., Thomas, B., Chen, C., et al. (2017). Added value of computer-aided CT image features for early lung cancer diagnosis with small pulmonary nodules: A matched case-control study. Radiology 286(1):286-295. 26. Rajpurkar, P., Irvin, J., Zhu, K., Yang, B., Mehta, H., Duan, T., Ding, D., Bagul, A., Langlotz, C., Shpanskaya, K., Lungren, M. P., Andrew, Y. Ng., (2017). CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 27. Gordienko, Y., Kochura, Y., Alienin, O., Rokovyi, O., Stirenko, S., Gang, P., Hui, J., Zeng, W., (2018). Dimensionality Reduction in Deep Learning for Chest X-Ray Analysis of Lung Cancer. In: International Conference on Advanced Computational Intelligence (ICACI). 28. Jarrett, K., Kavukcuoglu, K., Ranzato, M., & Lecun, Y., (2009). What is the best multi-stage architecture for object recognition?. In Proc. International Conference on Computer Vision. 2146 – 2153. 29. Dong, Y., Pan, Y., Zhang, J., & Xu, W. (2017). Learning to Read Chest X-Ray Images from 16000+ Examples Using CNN. In Proceedings - 2017 IEEE 2nd International Conference on Connected Health: Applications, Systems and Engineering Technologies, CHASE 2017. 51–57. 30. Aloysius, N., & Geetha, M. (2018). A review on deep convolutional neural networks. In Proceedings of the 2017 IEEE International Conference on Communication and Signal Processing, ICCSP 2017. 588–592.

69

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Convolutional Neural Network for Environmental Sound Event Classification İlyas Özer1, Muharrem Düğenci2, Oğuz Findik1, Turgut Özseven3

1Department of Computer Engineering, Karabuk University, Turkey [email protected] [email protected] 2Department of Industrial Engineering, Karabuk University, Turkey [email protected] 3Department of Computer Engineering, Tokat Gaziosmanpasa University, Turkey [email protected]

Abstract

Automatic sound recognition (ASR) is a remarkable research area in recent years. It has a wide variety of subcategories. Environmental sound classification (ESC) is one of these categories. ESC is a challenging task because of environmental sounds have noise-like structure and also its created by so many different sources. In this study, we propose a model that uses spectrogram image features and convolutional neural networks for ESC task. In the proposed model, environmental sound events are converted to fixed-size spectrograms. After this step spectrograms are converted to linear greyscale images, and also reduced dimensions with image resizing. Three grayscale image are directly combined, without quantization to create RGB image. Finally feature extraction process and classification are performed via convolutional neural networks (CNN), which have very powerful performance in image classification. Proposed model are tested on the two publicly available dataset which is namely ESC- 50. As a result %74.85 classification success are obtained.

Keywords: Environmental Sound Classification, Convolutional Neural Network

1. Introduction In the literature, sound recognition is divided two basic category, which are speech and non- speech sound recognition. Although these two main categories are closely related, there are relatively more studies on speech recognition in the literature. On the other hand, non-speech sound recognition studies are gaining speed in the recent years. Non-speech sound recognition has several variety of fields such as bioacoustic monitoring[1], multimedia [2] and intruder detection in wildlife areas [3]. Also ESC is an another important field of non-speech sound recognition and its widely used many areas such as machine hearing [4], home automation [5] and surveillance [6].

When we compare the speech recognition, sound events in the ESC tasks very diverse and also the frequency ranges are very wide. In addition to this, in recent years the number of classes to be defined are increased. However the initial studies on ESC were generally based on speech recognition tasks. For this reason, studies are conducted to develop new technics for non-speech sound recognition and ESC tasks. In this direction, mel-frequency cepstral coefficients (MFCC) [7], stabilized auditory image [8], SIF [9] and perceptual linear prediction coefficients [10] are started to be widely used in this days. In addition, the methods used for the classification of these features have started to change in recent years. Gaussian mixture models [11], support vector machine (SVM) [10] and CNN [12] are some of the most widely used classification methods in recent epoch. In recent years, the task of automatic recognition of sound events has become an attractive research area. Classification of environmental sound events is one of these study areas. In this respect, ESC-50 data set consisting of environmental sounds was created to make a standard evaluation task [13]. In addition, classification performance was evaluated in the same study. The zero crossing rate and MFCC features were classified by k- nearest neighbour, SVM and random forest, as a result 32.2%, 39.6% and %44.3 classification accuracy are obtained, respectively. In another study using the same

70

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY data set [14], the features generated by the short-time Fourier transform were classified by a two different CNN architecture which are consisting of 3 and 5 layers. In the best case %56.37 classification accuracy are obtained. CNNs have recently started to be used in the classification of sound events as well as the tasks related to the image. On the other hand, CNNs have very high computational costs. Therefore, a number of studies are being carried out to provide a more compact CNN architecture [15]. The proposed CNN architecture was tested in the same evaluation task, and a success of 66.25% was achieved. On the other hand, in another study using a more complex CNN architecture [16], spectrograms were classified which is produced with the same data set. %73.2 classification score was obtained in the conditions 16 kHz sampling rate and GoogLeNet architecture.

In this study, we propose a model that uses spectrogram image features and also CNN for classify the wide variety of environmental sounds. Due to their robustness to noise, SIFs are widely used in automatic sound recognition tasks. A spectrogram image is obtained by the fast Fourier transform and its provides a visual representation of the frequency spectrum of a sound signal. In addition to this, CNNs are widely used for computer vision problems and image related applications also have a serious popularity in this fields. Moreover, the CNN is insensitive to the pattern position on the image. In this case, it makes a CNN suitable tool for classifying SIFs.

In the proposed model, the sound signal is divided into fixed length pieces and the FFT operation is performed. The obtained time-frequency matrix is normalized between the range [0, 1] and the result of this process grayscale image is created. All values 0.5 and above are taken as 1 to obtain a more sharper image. Because the dimensions of the image obtained at this stage is too large and the image size is reduced by the resizing method to decrease the computational cost. After this step three grayscale image are directly combined, without quantization to create RGB image. Finally created RGB images are classified by the CNN.

The remaining parts of this article are organized as the following. SIF properties are described in Section 2; CNN is given in Section 3; The proposed approach is described in Section 4; Section 5 gives Experimental Evaluation, system performance is given Section 6; Section 7 contains the results.

2. Spectrogram Image Features

The spectrogram image is a visual representation of the frequency spectrum of a signal and is generated by a stack of magnitude spectrum obtained via fast Fourier transform (FFT). The FFT is performed by dividing a given sound vector into parts that are highly overlapping with the lengths of the 푤푠, and the spectral vector f is obtained. In this case, the F spectral magnitude vector 푓퐹of the current frame is obtained as follows:

sF ()()() n=+ s F n w n for nw=−0...(s 1) (Eq. 1)

ws −1 − j2 nkw s for (Eq. 2) fFF()() k=  s n e kw=−1...(s 2 1) n=0

Where w (n) is the hamming window function, and cons is the increase between two consecutive frames. The linear spectrogram values are obtained as follows:

SLinear( k , F )= | f F ( k ) | (Eq. 3)

The time-frequency matrix is normalized between the range [0, 1] so grayscale image is created.

71

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

SkFS(,)min()− GkF(,) = (Eq. 4) max()min()SS−

3. Convolutional Neural Network

CNN is a ordinary multi-layer neural network [17], and its firstly used for computer vision applications [18] also its quite appropriate for image-related problems in machine learning [19], [20]. As generally CNN consists of 3 basic layers which are namely convolution, pooling and fully connected. On the other hand CNN's layers have an 3-dimensional neurons on the contrary to the traditional neural networks.

Convolutional layer is a most important component of CNN and its defined by the two variables, which are filters size and number of produced maps. By acting on the idea that an image of an object is independent on the picture, each neuron binds to only a small portion of the input and extends over the entire depth of the input. 2-dimensional activation map is produced via dot product process on the input and the filters.

CNN operations are generally linear process such as matrix multiplication, addition and convolution. On the other hand most of the data on real world problems are non-linear. ReLU layers provide the non-linearity with a non-saturating activation function. Which is defined as follows:

f( x )= max(0, x ) (Eq. 5)

Another important process on CNN is the non-linear sub-sampling which is called pooling. Pooling is a feature reduction operation in the traditional computer vision problem. Several types of pooling layers are used in CNN models, but most widely knowns are average pooling and max-pooling. In this study max-pooling methods are used. The maximum pooling method divides the input into non- overlapping parts and selects the biggest one of all subparts.

In addition to the this layers fully connected layer are widely used in CNN architectures. Fully connected layer neurons are connected to the earlier layers and the following layers completely, like a traditional neural networks. Finally in the output layer a classifier is used such as softmax or support vector machine. The multilayer structure built in this way allows the network to identify more abstract features and discover more complex structures.

4. Proposed approach

Different approaches and parameters are used in the literature to create spectrogram images such as overlapping rates, log or linear scale, RGB or greyscale format, frame lengths and there is no exact certainty in this item. In our previous work [12] we used linear quantized spectrogram images for sound event classification. On the other hand in this study we used spectrogram images without quantization. There are several reasons for this approach. In an ESC-50 dataset consist of a small number of samples combined with a relatively large number of classes. Therefore when we made quantization some information is loss and classification performance are decrease. Another reason some audio recordings contain data that overlaps with background noise in ESC-50. On the other hand, the data set in the other study [12] was created in a sterile environment.

72

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 1 – Block diagram of the environmental sound classification system.

Spectrograms are a picture that occurs from the visual representation of the frequency spectrum. It is therefore very suitable for classification by CNNs. But there are two important problems at this point. The first is that the size of the spectrogram images is very large compared to features like the MFCC.

For this direction our approach consist of 6 basic step. Firstly down sampled the sound recordings in amplitude-time domain. After this step we create a fixed dimensional spectrogram. Thereafter normalization process is performed and greyscale image obtained. Equation 4 is used for this step but all values 0.5 and above are taken as 1 to obtain a more sharper image. At this stage, the dimensions of the spectrogram images are quite large and are reduced by the resizing process. Three grayscale images which are dimensions reduced are directly combined, without quantization. Finally classification is performed by the CNN. Detailed steps of proposed approach are depicted in Figure 1.

5. Experimental evaluation

This section primarily describes the ESC-50 data set and then it provides information about the spectrogram parameters and CNN architecture.

5.1 Dataset

The ESC-50 data set contains 50 different class of audio files, each consisting of 40 environmental sound recordings. In total, the records of 2000 labeled environmental sound event data which are each 5s long. In addition, each file was recorded as mono with a sampling rate of 44.1 Khz. Also, the data set is grouped under 5 basic groups of 10 different classes. The names of these 5 basic groups are: • animal sounds, • natural soundscapes and water sounds, • human (non-speech) sounds,

73

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

• interior/domestic sounds, • exterior/urban noises. A small number of samples combined with a relatively large number of classes have made this data set extremely challenging for conventional classification methods. In addition, while creating the data set, background noise was kept as low as possible, while audio events were intended to be foregrounded. However, field records are far from being sterile. Therefore, some audio recordings contain data that overlaps with background noise. It is an additional source of difficulty.

5.2 Experimental setup

Original sound records have an 44.1 kHz sampling rate and its directly down sampled to the 22.05 kHz. In all cases the spectrogram produced with highly overlapped by 1378 point and 1411 point hamming window. As a result, a spectrogram is created which is dimensions 3299x706. For the resizing process Lanczos kernel methods are applied and each image fixed 513x200 dimensions. In addition, all tests were performed with 5 fold cross validation.

5.3 CNN architecture

We use similar CNN architecture in a study [12]. It consist of 7 learnable layer and 5 of them are convolutional layer and 2 of them are fully connected layer. On the contrast to [5], in this study we used two additional dropout layer. They were placed before each fully connected layers.

We trained our network with a batch size of 50, initial learning rate 1푥10−4, learning rate drop factor 0.2, learning rate drop period 150 iteration and training process after stop 960 iterations. All other parameters are same with our previous study [12].

5.4 Result and Discussion

The results obtained with the proposed model in Table 1 and the results of some studies with this data set can be seen. When the results are evaluated, it is seen that the proposed model gives better results than many approaches. On the other hand, classification performance is 11.65% below the best value. On the other hand, compared to the worst score, a better result is obtained with a rate of 42.65%. The results obtained as such are promising for future studies. It is possible to move the classification performance upwards by making changes to the spectrogram image parameters and CNN architecture.

Table 1 - Classification results for several state-of-the-art methods ( proposed approach show with standard deviation)

Model Mean accuracy(%)

FBM ⊕ ConvRBM-BANK 86.5 CNN pretrained on AudioSet 83.5 GoogLeNet on spectrograms 73.2 SoundNet 8-layer CNN 66.25 3-layer CNN with vertical filters 56.37 MFCC & ZCR + random forest 44.3 MFCC & ZCR + SVM 39.6 MFCC & ZCR + k-NN 32.2 Proposed approach 74.85 (0.99)

74

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Table 2 shows a number of results obtained with different parameters. Tests were repeated by changing only one parameter at a time and the results in Table 2 were reached. When we use quantized images classification performance dramatically decreased to %68.3. Which is score under the proposed approach and also standard deviation score is very higher when compare the other results. Also, when we remove the 2 dropout layers we added, the classification performance is reduced by 3.15%. Similarly, when we apply a fixed learning rate, the classification performance decreases.

Table 2 - The effect of the several parameters (with standard deviation)

Model Mean accuracy(%)

Quantized Image 68.3 (2.8) Without Droupout Layer 71.7 (1.7) Fixed Learning Rate 72.5(1.1) Proposed approach 74.85 (0.99)

6. Conclusion

In this article, CNN and spectrogram image based environmental sound event classification approach are proposed and the classification performance was compared with standard evaluation task. In the proposed model, highly overlapping spectrogram images are combined without quantization. As a result, generated images were classified by CNN and a classification score of 74.85% was obtained with ESC-50 data set. However, it is possible to obtain a higher classification performance by applying the proposed model with different spectrogram image parameters and CNN architectures.

References

1. Weninger, F., & Schuller, B. (2011, May). Audio recognition in the wild: Static and dynamic classification on a real-world database of animal vocalizations. In acoustics, speech and signal processing (ICASSP), 2011 IEEE international conference on(pp. 337-340). IEEE. 2. Wold, E., Blum, T., Keislar, D., & Wheaten, J. (1996). Content-based classification, search, and retrieval of audio. IEEE multimedia, 3(3), 27-36. 3. Ghiurcau, M. V., Rusu, C., Bilcu, R. C., & Astola, J. (2012). Audio based solutions for detecting intruders in wild areas. Signal Processing, 92(3), 829-840. 4. Lyon, R. F. (2010). Machine hearing: An emerging field [exploratory dsp]. Ieee signal processing magazine, 27(5), 131-139. 5. Vacher, M., Serignat, J. F., & Chaillol, S. (2007, May). Sound classification in a smart room environment: an approach using GMM and HMM methods. In The 4th IEEE Conference on Speech Technology and Human-Computer Dialogue (SpeD 2007), Publishing House of the Romanian Academy (Bucharest)(Vol. 1, pp. 135-146). 6. Radhakrishnan, R., Divakaran, A., & Smaragdis, A. (2005, October). Audio analysis for surveillance applications. In Applications of Signal Processing to Audio and Acoustics, 2005. IEEE Workshop on (pp. 158-161). IEEE. 7. Guo, G., & Li, S. Z. (2003). Content-based audio classification and retrieval by support vector machines. IEEE transactions on Neural Networks, 14(1), 209-215. 8. Walters, T. C. (2011). Auditory-based processing of communication sounds (Doctoral dissertation, University of Cambridge).

75

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

9. Dennis, J., Tran, H. D., & Li, H. (2011). Spectrogram image feature for sound event classification in mismatched conditions. IEEE signal processing letters, 18(2), 130-133. 10. Portelo, J., Bugalho, M., Trancoso, I., Neto, J., Abad, A., & Serralheiro, A. (2009, April). Non-speech audio event detection. In Acoustics, Speech and Signal Processing, 2009. ICASSP 2009. IEEE International Conference on (pp. 1973-1976). IEEE. 11. Atrey, P. K., Maddage, N. C., & Kankanhalli, M. S. (2006, May). Audio based event detection for multimedia surveillance. In Acoustics, Speech and Signal Processing, 2006. ICASSP 2006 Proceedings. 2006 IEEE International Conference on (Vol. 5, pp. V-V). IEEE. 12. Ozer, I., Ozer, Z., & Findik, O. (2018). Noise robust sound event classification with convolutional neural network. Neurocomputing, 272, 505-512. 13. Piczak, K. J. (2015, October). ESC: Dataset for environmental sound classification. In Proceedings of the 23rd ACM international conference on Multimedia (pp. 1015-1018). ACM. 14. Huzaifah, Muhammad. "Comparison of Time-Frequency Representations for Environmental Sound Classification using Convolutional Neural Networks." arXiv preprint arXiv:1706.07156(2017). 15. Jin, X., Yang, Y., Xu, N., Yang, J., Feng, J., & Yan, S. (2017). WSNet: Compact and Efficient Networks with Weight Sampling. arXiv preprint arXiv:1711.10067. 16. Boddapati, V., Petef, A., Rasmusson, J., & Lundberg, L. (2017). Classifying environmental sounds using image recognition networks. Procedia Computer Science, 112, 2048-2056. 17. Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems (pp. 1097-1105). 18. LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278-2324. 19. Cireşan, D., Meier, U., & Schmidhuber, J. (2012). Multi-column deep neural networks for image classification. arXiv preprint arXiv:1202.2745. 20. Ma, J., Wu, F., Zhu, J., Xu, D., & Kong, D. (2017). A pre-trained convolutional neural network based method for thyroid nodule diagnosis. Ultrasonics, 73, 221-230.

76

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Current Situation of Gaziantep Industry in the Perspective of Industry 4.0: Development of a Maturity Index Kerem Elibal1, Eren Özceylan2, Mehmet Kabak3, Metin Dağdeviren3

1 BCS Metal Industry, Gaziantep, Turkey [email protected] 2 Department of Industrial Enginnering, Gaziantep University, Turkey [email protected] 3 Department of Industrial Enginnering, Gazi University, Turkey [email protected] [email protected]

Abstract

This paper presents maturity measurement model for determining the readiness level of Gaziantep/Turkey Industry, in the concept of Industry 4.0. Neccessity of implementing Industry 4.0 philosophy, to have a sustainable, competitive and leading economy, forces corporates and governments to assess themselves in the terms of Industry 4.0 principles, in order to prepare accurate road maps. Although there are many studies about measuring methods and organizational and/or national readiness studies for Industry 4.0 in all over the world, especially in Europe, our study motivated from the insufficients research for Turkey industry. Our study aims to complete this gap by investigating different maturity models and as a result, new model draft had been proposed. Results encouraged us to develop this draft into its final form and to conduct the model for evaluation of Gaziantep/Turkey Industry. Keywords: Industry 4.0, Maturity Index, Maturity Modelling

1. Introduction The term “Industry 4.0” is a new phenomenon, presented first at Hannover Fair 2011. Despite the fact that the term came on the scene from Germany, because of its promising benefits, it spreads out to many countries with a significant speed, which has developed economies. In general context, all industrial revolutions aim operational and productional efficiency for a profitable and sustainable economy, so as Industry 4.0, but Industry 4.0 also provides gains to satisfy developed economies’ sensivities about environment, work-life balance and demographic changes.

To stay alive in today’s competitive economy, many countries are aware of that, this challenge will be possible to define and improve Industry 4.0 concepts. Governments, strategy companies, academic researchers are trying to clarify the expectations, necessities and road maps both globally and domestically. Germany; the frontrunner of Industry 4.0, expects increase of %23 (78.77 billion Euros) in Gross Domestic Product (GDP) from 2013 to 2025 (Gökalp et al., 2017). According to the survey conducted by Pricewater House Coopers (2016) with over 2000 senior executives from industrial product companies in 26 countries across Europe, the Americas, Asia Pasific, Middle East and Africa; %56 of respondents expect over %20 efficiency gains in the next five years, and in the same research it is stated that total 907 billion US dollars investment (%5 of annual revenue) will be done for Industry 4.0 implementation, globally. RolandBerger Strategy consultants report (2014) indicates that Europe must invest 90 billion Euros per year for Industry 4.0 implementation until 2030 to reach its economic targets. Sung (2017) aimed to make suggestions about what kind of policies Korean Government and companies should imply for Industry 4.0 revolution and states the competitiveness rank of Korea for the 4th industrial revolution. National Academy of Science and Engineering (2013) served a detailed report about recommendations for implementing Industry 4.0, especially for the German Industry. Also,

77

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

VDMA (The Mechanical Enginnering Industry Association) published an Industry 4.0 guideline for small and medium sized German mechanical engineering industry companies. Horvat et al. (2018) presented a readiness method and applied for Kazakhstan’s manufacturing sector. Veza et al. (2015) analyzed the current situation of Croatian manufacturing industry with regard to industry 4.0. Rajni and Kocsis (2018) give information about Hungarian Industry 4.0 National Technology Platform (NTP) who aims to clarify national readiness and Hungarian companies’ readiness about Industry 4.0. Basl (2017) investigated the readiness level of Czech Republic companies for the Industry 4.0. Hamzeh et al. (2018) conducted a survey study on Industry 4.0 for local New Zealand manufacturing and gave some suggestions.

Mentioned literatures above show us there are many studies about expectations, readiness levels and suggestions for the fourth industrial revolution, all over the world, especially in Europe. It is obvious that there should be some studies about the Industry 4.0 for Turkey to construct suggestions and road maps for companies and policy makers on behalf of reaching economical targets. Bulut and Akçacı (2017) tried to examine Turkish R&D and communication indicators from the perspective of Industry 4.0 and defined some basic concepts of Industry 4.0 but this study is lack of understanding Turkish companies readiness for Industry 4.0, generally focused on R&D investments and internet usages and claimed that improvement of these parameters also will improve Industry 4.0 revolution in Turkey. MÜSİAD (2017) (Independent Industrialist’s and Businessmen’s Association) published a report about Industry 4.0 and logistics, but this report investigates the Industry 4.0 in the field of logistics only, defines concepts of Industry 4.0 and make suggestions about the application of logistics sector, so it does not deal with the readiness degree of Turkey for Industry 4.0. TÜBİTAK (2016) (Scientific and Technological Research Council of Turkey) published a national report about Industry 4.0 road map for Turkey. This report includes the current level (maturity) of Turkey for the Industry 4.0, states that Turkey’s digital maturity level is between Industry 2.0 and Industry 3.0, also indicates the most valuable technologies which are necessary for the Industry 4.0 and which sectors will make the highest added value with the implementation of Industry 4.0, also declares that which sectors are aware of Industry 4.0. This report gives some information and opinions about Turkey readiness for Industry 4.0 at a national level, but gives highly specialized recommendations and road maps for spesific sectors, and it is lack of make suggestions for SMEDs. Özkurt (2016) investigated the status of Industry 4.0 in Sakarya/Turkey region with a questionarrie which includes 73 questions. This study gives quite information about in which concepts of Industry 4.0 the firms are aware of, but does not include a conceptual approach and from this study it is hard to define the readiness level of the region and also does not make spesific suggestions.

Reviewed literature until here gives us the motivation about our study to make a conceptual approach for Industry 4.0 readiness and awareness in Turkey. In this study, it is aimed to prepare a maturity measuring approach which will show the current status of companies according the determined dimensions. To the best of our knowledge there is no study which assesses companies according to the organizational levels (dimensions) by numerical values in Turkey region with the concepts of Industry 4.0, so it is aimed to make this assessment at Gaziantep/Turkey Industry. This study structured as follows; in section 2 basic concepts of Industry 4.0 will be described briefly. In section 3, current maturity models will be investigated. In section 4, our methodogy will be described. In section 5 conclusions will be done.

78

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

2. Industry 4.0 Industry 4.0 can be birefly defined as digitization of production. Until the term Industry 4.0 has been declared, previous industrial revolutions did not defined as versions according to their apperance in the time zone of production revolutions. Previous revolutions also defined as 1.0, 2.0 and 3.0 after the term 4.0 came on the scene.

Today, emerging technologies especially about IT systems and electronics allow digital integration of all of the components in an organization, starting from the development of the product till to the customer. Internet of things(IoT), big data analysis, additive manufacturing(3D printing), manufacturing execution systems (MES) applied with artificial intelligence(AI), cloud data, autonomus manufacturing assets , agumented reality systems, smart sensors, smart products are some of the key concepts of Industry 4.0 and provide not only the integration and communication of the organization in a cyber physical system (CPS), also delivers the customized production in low bathces in affordable and profitable levels.

It should also considered that, the term industry 4.0 is not only about connecting machines and/or products with each other via internet or it is not only about fully automated systems, but it is a technological philosophy which the final goal is making fast and right decisions by using digitalized data. Digitization of manufacturing and/or other departmens of a company, and giving fast and agile decisions by using these data is the main concept of Industry 4.0. Companies should consider automated decision systems with automated phsyical systems, the peak point of a “Dark Factory” should not be fully automated physical resources, and also there should be autonomous and learning decision systems, in a decentrealized manner.

There are many researhes which defines key technologies of Industry 4.0. Lu (2017) provides a significant amount of literature survey about Industry 4.0 technologies. Also Bibby and Dehe (2018) give information about attributes of “Factory of Future”.

2.1 Benefits of Industry 4.0 Benefits of the 4th industrial revolution can be summarized as;

• Efficiency in production and operation areas by digitization and integration of all aspects of the systems. • Ability to produce in low batches according to the spesific customer demands. • Ability to control and reduce energy consumption to meet the environment protection policy of the societies. • Tracebility of the product and ability to invervene the product while it is at the customer facility (Example: Some machine producers can trace and communicate machine faults online while the machine is at the customer’s plant). • By decreasing human contrubition to the system, work-life balance will be improved and also organizations will make agile actions for the demographic changes.(Increasing of old population in Europe forces organizations creating of less human independent manufacturing systems, and Industry 4.0 is able to manage this). • Agility should be consumed as the most important benefit of Industry 4.0. Speed of analysis and action for the changes in an organization, no matter what kind of changes will occur, is the most important aspect for efficiency and customer satisfaction. Industry 4.0 applications provides organizations to create fast strategies against unexpected changes, like machine failure, customer order changes, lack of material, product failure, etc. Figure-1 and Figure-2 shows the difference

79

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

between the traditional systems and digitalized systems considering the response time for any kind of changes.

Figure 1 – Corporation adaptation process (Acatech, 2017)

Figure 2 - Corporation adaptation process with Industry 4.0 concepts (Acatech, 2017)

3. Maturity Models for Industry 4.0-Literature Review As we mentioned in the introduction, aim of this paper is create a maturity model to measure the readiness status of Gaziantep/Turkey Industry. Kohlegger et al. (2009) defines maturity models as rating capabilities of maturing elements and select appropriate actions to take the elements to a higher level of maturity. Becker et al. (2009) states that a maturity model consists of a sequence of maturity levels for a class of object and these objects are organizations or processes. Also in the same study it is declared that maturity models can be supported by questionarries and more than a hundred different maturity

80

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY models have been proposed according to the researches. But in our study, instead of studies about creating a maturity model, literature about construction of maturity models for Industry 4.0 has been investigated. The aim of this investigation is to clarify what kind of structures and what kind of aspects about these structures had been determined as maturity measurement topics.

Acatech Maturity Model (Acatech-2017) is one of the important studies about Industry 4.0 maturity modeling. This model describes maturity levels as computerisation, connectivity, visibility, transparency, predictive capacity and adaptibility. Computerisation is the basic digital level and includes different information technologies which are used seperatly in the organization. Connectivity is the integration of these information technologies which are used in all assets. Visibility stage provides captured data monitoring via systems like MES and ERP. Transparency stage is the stage where root cause analyses are done to answer “why is it happening?” and big data concept is the key factor of this level. Predictive capacity stage includes simulation and prediction activities by using collected data. Last stage adaptability aims to give autonomus decisions with minimum human assistance.

Figure 3-Levels of Industry 4.0 (Acatech, 2017)

The Acatech Industrie 4.0 maturity model aims to assign levels (mentioned in Figure 3) to the organization’s functional areas. These functional areas of an organization are development, production, logistics, services and marketing & sales. And the model breaks down these functional areas to 4 structural areas; resources, information systems, organizational structure and culture.

81

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Table 1-Design of Acatech Maturity Modeling

Structural Areas Resources: Human Information Systems: Organizational Structure: Rules of Culture: Willingles force, machinery, Systems like CRM, ERP, data usage/restriction of the and awareness of equipment, parts MES which are used for employees, determining new systems and products visibility and analysis of centralized and decentralized development. collected data decision making hierarchies

Development

Production

Logistics

Functional Areas Functional

Services

Marketing&Sales

Levels mentioned in Table 1 for the functional areas of an organization are measured by a questionarrie prepared by Acatech and also expert visits to the companies.

VDMA(2015) constructed a maturity model named IMPULS, which differs from the Acatech Maturity Model by defining dimensions as smart factory, smart products, smart operations, data-driven services, employees and strategy&organization instead of functional areas. These dimensions are detailed in 18 fields and maturity levels are stated as outsider, beginner, intermediate, experienced, expert and top performer. Smart factory concept defines sensor equipped and integrated assets of a manufacturing organization. Smart products are the products which collect data for itself and contact with the manufacturer even after it is the customer’s usage. Smart operations include smart data usage for planning and control actions. Data-driven services are the integration of manufacturer, product and customer to add value to the product after sold. IMPULS measures the maturity level of a company with an online questionarrie which includes 26 questions.

82

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 4- Dimensions and Associated Fields of Industrie 4.0 (Source VDMA-2015)

Schumacher et al. (2016) proposed a maturity model which includes 62 maturity items grouped into nine company dimensions. In this study it is stated that each maturity item’s importance may be different in the sense of Industry 4.0 maturity, so a weighting factor has been assigned to each item by 23 experts. An e-mail based questionarrie which consists of closed-ended questions with Likert-scale answers reaching from 1 to 5 has been constructed.

Table 2- Dimensions and maturity items of Industry 4.0 (Schumacher et al., 2016)

Dimension Exemplary maturity item

Strategy Implementation I40 roadmap, Available resources for realization, Adaption of business models

Leadership Willingless of leaders, Management competences and methods, Existance of central coordination for I40,

Customers Utilization of customer data, Digitalization of sales&services, Customer’s Digital media competence

Products Individulization of products, Digitalization of products, Product integration into other systems

Operations Decentralization of processes, Modeling and simulation, Interdisciplinary, interdepartmental collaboration

Culture Knowledge sharing, Open-innovation and cross company collaboration, Value of ICT in a company

People ICT competences of employees, openness of employees to new technology, autonomy of employees

Governance Labour regulations for I40, Suitability of technological standarts, Protection of intellectual property

Technology Existence of modern ICT, Utilization of mobile devices, Utilization of machine-to-machine communication

83

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Gökalp et al. (2017) investigated different maturity models and stated that none of the investigated ones fully satisfy the assessment criterias which are defined as fitness for purpose, completenes of aspects, granularity of dimensions, definition of measurement attributes, description of assessment method and objectivity of the assessment method. Besides none of them fully satisfy these criterias, it can be seen two of the investigated studies have a more achievement degree than the others so, Gökalp et al. (2017) made us to highligt the VDMA (2015) and Schumacher et al. (2016) . Also Gökalp et al. (2017) proposed a new approach to measure the maturity level of Industry 4.0 in 5 dimensions; those are asset management, data governance, application management, and organization allignment and process transformation.

Berger et al. (2018) also studied different maturity models, mostly inspired from Acatech Maturity Model of Acatech (2017) and proposed a 5 dimensional model which includes governance, technology, connectivity, value creations and competence. According to these dimensions a multiple answer questionnarie composed by 24 questions has been conducted.

Canetta et al. (2018) investigated several studies for Industry 4.0 maturity modelling which are done by both academia and practitioners. They presented that majority (%78) of the reviewed studies include a general approach with a self-assessment instrument and only few of them are constructed spesifically for the companies’ context. They proposed a questionarrie with 36 questions that can measure the maturity of the aspects which are strategy, processes, technologies, products&services and people.

Capgemini Group Consulting (2017) evaluates the maturity of an organization in sense of Industry 4.0 with 5 dimensions which are asset information management&analytics, process management, reliability&performance, governance&standarts, people&culture management, tools&technologies.

In our study, the aim of the literature review about Industry 4.0 maturity measurement methods is not to find gaps and compare them with each other, but to understand what kind of organization levels and assets are evaluated in the sense of clarifying Industry 4.0 readiness. By this review it was seen that, asset technology, information technology, employee skills, corporate culture&strategy are the major levels which should be assessed in the terms of Industry 4.0 concepts. Almost at all of the studies, these levels are evaluated seperately for each functions of an organization like product&production development, manufacturing, logistics, sales&services, marketing, human resources, etc.

4. Methodoloy This study will aim to construct a maturity model for Industry 4.0 and try to apply this method to investigate Gaziantep/Turkey readiness for industry 4.0 concepts.

Gaziantep has the 6th order in Turkey, according to the exporting values which is over 6 billion US dollars with %4 share of Turkey (TIM, 2017). The city has 950 manufacturing corporations (GAOSB) which are commonly grouped in textile, agro-food, machinery-metal, chemical-plastics and construction materials sectors (GTO, 2013).

Proposed method will assess the maturity of Industry 4.0 for 5 functional departmens of a corporation; those are manufacturing, logistics, sales, human resources and management; determined by considering the major sectors are not producing high-tech products; so the functions like product&production development, after sale activities are ignored to investigate.

84

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Determined functions will be assessed according to 2 dimensions which are; asset technology and information technology.

Maturity items mentioned in Table 3 will be assessed with a questionarrie which will be conducted by visits and/or online self-assessment. The questionarrie will include 2 sections; first section will consist of general questions about the company structure which will not taken into account for maturity calculation; like the size of the company or employee amount of the company. This section will also try to identify the Industry 4.0 awareness of the corporations.

In the 2nd section of the questionarrie, there will be closed-ended questions which have Likert- scaled answers from 1 to 5 to asses the functions mentioned in Table 3, where 1 means not implemented and 5 means fully implemented.

Table 3 - Proposed maturity model

Explanations of Some Maturity Items

Functions Manufacturing Logistics Sales Human Resources Management

Dimensions

Asset Sensors,RFID, Tracking Smart Employee ability to Identified Key Technology/Property barcode devices, GPS products (that Internet Performance technologies,M2M Technologies, can Communication Indicators(KPIs), communications(for automated in- communicate Technologies(ICT), documentated products, parts, land even after ability to monitor processes, open machines, tools or transportation, sold), and analyze data, communications, equipments),…. real time intervenable data security flexible and location products,…. awareness, willing multi- systems,… to adapt disciplanary technological hierarchies, changes,… procedures of data monitoring and sharing,…

Information MES systems, Logistics CRM HR management ERP systems, IT Technology predictive tracking systems, systems (include services, Cloud maintenance systems (inner customer education and usage,data methods,simulation or outer), portals, data career planning security methods, decision logistics ex-change management systems,… support methods,…. decision and modules) systems,… monitoring systems, customer behaviour and/or sales prediction,…

85

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Table 4- Example questions of the assessment for manufacturing function within asset/technology dimension

Can machines in your facility communicate with each other via sensors or some other 1 2 3 4 5 technologies?

How can you describe the rate of your sensor usage for data collection or 1 2 3 4 5 communication?

Table 5-Example questions of the assessment for management function within information technology dimension

Do you have an ERP system in an efficient usage? 1 2 3 4 5

Do you use Cloud Technology in your corporation? 1 2 3 4 5

Schumacher et al. (2016) stated that maturity items may have different contributions to Industry 4.0 concept, every item may not have same importance level, so they identified weighting factors for all maturity items by calculating the average importance rates from 23 experts. For example, “willing to adapt technological changes” may be more important than “employee ability to Internet Communication Technologies” for the human resources function of the organization. In our study we will try to assign these weighting factors by multiple criteria decision-making techniques (MCDM), with in its related dimension.

Each question will try to identify and score a maturity item by multiplying answer level (1 to 5) with weighting factor of the related item. There may be more than one question to score a maturity item and in this case average scores will be taken into account for that item. After determining all the scores, a graphical representation as shown in Figure 5 can be obtained, which will show all scores of functions for their dimensions, and overall score of the corporation.

86

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Asset Technology Information Technology Manufacturing

2

1,2

Management Logistics 2 2 1,4

1 1,5

2,3 2 Overall Maturity Human Resources Sales Score:2.9

Figure 5 - An example of graphical representation of maturity scores

5. Conclusion In this paper we have tried to propose a maturity measurement method to determine the Industry 4.0 readiness of Gaziantep/Turkey Industry. First step of our study is reviewing the current maturity models and decide what kind of aspects of the organization functions will be assessed. We have determined 5 organization functions and decided to evaluate these functions within two Industry 4.0 dimension. For each organization and dimension, we have suggested different maturity items and MCDM methodology to assign weights to these items. Furthermore, we have proposed a maturity calculation method and showed how we can represent our results in graphical ways.

This method will not only show us the level of a corporation, also will give ideas which organizational functions should be developed in what kind of aspects; asset or information systems. Besides, results will be a road map for each company, also by the cumulative results of Gaziantep Industry, it will be a road map for policy makers of government, regional Industrial Chamber, non- govermental organizations (like TMMOB) or universities.

As well as determining corporal or regional readiness for Industry 4.0, we also expect from this study to make an analysis about which sectors are suitable for Industry 4.0 or which sectors will have the minimum effort to implement the 4th industrial revolution concept.

Further research activities will mainly aim to develop the maturity items, the questionarrie and conduct the surveys with this questionarrie by visits.

References

Acatech (National Academy of Science and Engineering) (2013). Recommendations for implementing the strategic initiative Industrie 4.0. Final report of the industrie 4.0 working group. Acatech (National Academy of Science and Engineering) (2017). Industry 4.0 maturity index. Managing the digital transformation of companies.

87

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Basl, J. (2017). Pilot study of readiness of Czech companies to implement the principles of industry 4.0. Management and Production Engineering Review. 8, 3-8. Becker, J., Knackstedt., R., Pöppelbuss, J. (2009). Developing maturity models for IT management- a procedure model and its application. Business and Information Systems Engineering. 1, 213-222. Bibby, L.and Dehe, B. (2018). Defining and assessing industry 4.0 maturity levels-case of the defence sector. Production Planning and Control. doi: 10.1080/09537287.2018.1503355. Bulut, E., Akçacı, T. (2017). Endüstri 4.0 ve inovasyon göstergeleri kapsamında Türkiye analizi. Assam International Referred Journal. 7, 50-72. Canetta, L., Barni, A., Montini, E. (2018). Development of a digitalization maturity model for the manufacturing sector. IEEE International Conference of Engineering, Technology and Innovation. Capgemini Group Consulting, Technology, Outsourcing. (2017). Asset performance management maturity model. Retrieved from https://www.capgemini.com/wp- content/uploads/2017/08/asset_performance_management_maturity_model_paper_web_version.pdf. Colli, M., Madsen, O., Berger, U., Moller, C., Waehrens, B.J., Bockholt, M. (2018). Contextualizing the outcome of a maturity assessment for industry 4.0. 16th IFAC Symposium on Information Control Problems in Manufacturing. GAOSB, http://www.gaosb.org/index.php/firmalar/. GTO. (2013). Invest in Gaziantep, Gaziantep Chamber of Commerce, Gaziantep, Turkey. Retrieved from: https://www.investingaziantep.gov.tr/upload/yazilar/Gaziantep-Invest-Report-636975.pdf. Gökalp, E., Sener, U., Eren E. (2017). Development of an assessment model for industry 4.0: Industry 4.0-MM. International Conference on Software Process Improvement and Capability Determination, 128-142. Hamzeh, R., Zhong, R., Xu, X.W. (2018). A survey study on Industry 4.0 for New Zealand Manufacturing. 46th SME North American Manufacturing Research Conference. 26, 49-57. Horvat, D., Stahlecker, T., Zenker, A., Lerch, C., Mladineo, M. (2018). A conceptual approach to analysing manufacturing companies’ profiles concerning industry 4.0 in emerging countries. 28th International Conference on Flexible Automation and Intelligent Manufacturing. Kohlegger, M., Maier, R., Thalmann, S. (2009). Understanding maturity models results of a structured content analysis. Proceedings of I-KNOW ’09 and I-SEMANTICS ’09. Lu,Y. (2017). Industry 4.0: A survey on Technologies, applications and open research issues. Journal of Industrial Information Integration. 6, 1-10. MÜSİAD (2017). Endüstri 4.0 ve geleceğin lojistiği. 2017 Lojistik sektör raporu. İstanbul, Müsiad. Özkurt, C. (2016). Endüstri 4.0 perspektifinden Türkiye’de imalat sanayinin durumu: Sakarya imalat sanayi üzerine bir anket çalışması. Sakarya Üniversitesi, Sakarya. Pricewater House Coopers (2016). Industry 4.0: Building the digital enterprise. Retrieved from https://www.pwc.com/gx/en/industries/industries-4.0/landing-page/industry-4.0-building-your-digital- enterprise-april-2016.pdf. Rajnai, Z., Kocsis, I. (2018). Assessing industry 4.0 readiness of enterprises. IEEE 16th World Symposium on Applied Machine Intelligence and Informatics. doi: 10.1109/SAMI.2018.8324844 Roland Berger Strategy Consultants (2014). Industry 4.0: The new industrial revolution. How Europe will succeed. Schumacher A., Erol, S., Sihn, W. (2016). A maturity model for assessing industry 4.0 readiness and maturity of manufacturing enterprises. Procedia CIRP, 52,161-166. Sung, T.K., (2017). Industry 4.0: A Korea perspective. Technological Forecasting and Social Science,132, 40-45. TIM.(2017),http://www.tim.org.tr/files/downloads/rakamlar/2017/12/iller_bazinda_ihracat_rakamlari_aral%C4 %B1k_2017.xlsx. TÜBİTAK (2016). Yeni sanayi devrimi, akıllı üretim sistemleri yol haritası. Ankara, Tübitak Bilim, Teknoloji ve Yenilik Politikaları Daire Başkanlığı. VDMA. Guideline Industrie 4.0 Guiding principles for the implementation of Industrie 4.0 in small and medium sized businesses. Retrieved from https://industrie40.vdma.org/documents/4214230/0/Guideline%20Industrie%204.0.pdf/70abd403-cb04-418a- b20f-76d6d3490c05. VDMA. (2015). Impuls-Industrie 4.0 readiness. Retrieved from https://industrie40.vdma.org/documents/4214230/26342484/Industrie_40_Readiness_Study_1529498007918 .pdf/0b5fd521-9ee2-2de0-f377-93bdd01ed1c8. Veza, I., Mladineo, M., Peko, I. (2015). Analysis of the current state of Croatian manufacturing industry with regard to industry 4.0. 15th International Scientific Conference on Production Engineering.

88

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Demands in Wireless Power Transfer of both Artificial Intelligence and Industry 4.0 for Greater Autonomy Sedat Tarakci1, 2, H. Gokay Bilic1, Aptulgalip Karabulut1, Serhan Ozdemir1

1Artificial Iytelligence & Design Laboratory, Department of Mechanical Engineering, [email protected] [email protected] [email protected] Izmir Institute of Technology, Urla, Izmir, Turkey [email protected] 2Tirsan Kardan A.Ş., Manisa, Turkey [email protected]

Abstract

There are many autonomous applications in daily life and they are limited only by our engineering. It is critical barrier for Industry 4.0 to make efficient communication between all physical objects to transfer of information and power in today’s technology. This paper presents the efficient wireless energy transfer methods for different applications which is already used and will be widely used for Artificial Intelligence technology. After the review of historical background, mostly used inductive coupling and capacitive coupling methods in Artificial Intelligence and their importance related with Industry 4.0 applications are demonstrated. Energy transfer demands for radio frequency identification (RFID) application are discussed with the definition of backscatter coupling. Finally, using wireless communication and power transfer methods for greater autonomy is investigated.

Keywords: Capacitive Coupling, Energy, Inductive Coupling, Industry 4.0, Power Transfer, RFID, Wireless

1) Introduction

New technology era, named as Industry 4.0 is characterized by the collaboration between computer based elements and physical entities. Researchers are developing new intelligent applications for the field which will act as the eyes and ears of IT [1]. There are many automatic identification methods which play key role for Artificial Intelligence communication such as barcodes, optical characters, voice, fingerprint, radio frequency and etc. Power demands and transfer methodologies with different coupling methods will be the key points in the near future for more intelligent developments.

Today, one of the most important controllers is the electric power to detect the quality of life. Usually, this quality factor is transferred with wires. However, the existence of wires could be constituted a problem for transmitting electric power. Some of these problems are; heavy losses of power and high costs. According to these undesirable conditions have emerged wireless power transfer methods (WPT) [2]. The ubiquitous WPT technology is integrated inside of many variety charging systems application. Wireless power transfer method provides many advantages for embedded circuitry. For instance, the galvanic isolation prevents transferred electric energy on foil surfaces from electro shocking to human beings [3]. While the wireless power transmission techniques are inquired, they can be diversified according their types. These are respectively; inductive coupling, capacitive coupling, magnetic resonance, laser, microwave [4].

Inductive wireless power transfer is the transmission of the electrical energy from a power source in order to drive an electrical load without any physical connections [5].Wireless power transfer system was asserted by Nikola Tesla in the 1890’s using the patented ‘Tesla coil’ which are acted as resonant transformers [5], [6]. His main objective was to supply mains power wirelessly. However, at

89

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY those days, wireless power transfer methods could not became popular regarding their low efficiency which is highly dependent to distance respectively [7]. With the recent advances, WPT technology has gained importance thanks to their advantages over conventional methods [8], [9]. Especially it is considered for the cases where the connection of wires may cause hazardous and dangerous incidents [5].

Medical implants had started to become objective of WPT during 1960s [10]. The researchers from MIT carried out a WPT system configuration in order to powered a light bulb of 60W with the distance from source over 2 meters in 2007 [6]. Today, demand of WPT technology for both consumer and industrial electronics arise rapidly [8]. In market, wide range of applications are reported such as charging devices for cell phones, laptops, toothbrushes and shavers as well as sensors for rotating components, medical implants and subsea devices with the power transfer efficiencies up to 90% [6],[9].

CPT method is based on electric line that it is used electric field lines in order to transmit power and data. Electric field line do not interference of stability of imbedded. Meanwhile, the design criteria of parallel capacitive plate are very suitable for power and data transfer in systems. The electric shock is obstructed by galvanic isolation where on the capacitive plates. However CPT is effective less than 1 mm distance between two capacitor plates. Nevertheless, the efficiency of capacitive coupling is less than IPT. Although the efficiency of CPT is lower than IPT, these lower efficiencies were reached over 90% at 1 kW [11].

Mostly used wireless identification method is RFID. It uses backscatter coupling technology for long range communication which will be discussed in chapter 4. The history RFID dates back to end of 1940s with the famous paper which was written by Harry Stockman on reflected power communication. After the development of radar and Friend of Foe (IFF) identification systems, technology became reality. With the help of theoretical developments and also laboratory experiments, RFID widely deployed and became a part of everyday life [12]. At first glance, there are many potential hurdles for adoption like high cost values of IC chips or integration of the technology in to industrial applications [13]. But in reality product or application specific energy requirement is the most critical design issue for the engineers, especially for zero power consumption targets with Industry 4.0 applications.

2) Inductive Coupling Case

The type of WPT can be understood with help of wavelength of the antenna (Eq.1. Short range applications can be classified in the range of one antenna diameter whereas midrange applications are in the range of 10 times of antenna diameter. c λ = (Eq.1) f

Inductive power transfer (IPT) is realized the power transfer with help of magnetic coupling between coils. Power is transferred by means of induced current on the receiver coil by the transmitter

90

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY coil which is driven by source. IPT system can be short or midrange application depending on resonance configuration which can be on

Table 3 - Classification of IPT.

Table 3 - Classification of IPT[14]

Energy-carrying Technology Power Range Efficiency medium

Electromagnetic Traditional IPT High Low High

Field Coupled magnetic High Medium High resonance

IPT system architecture can be seen in Figure 2 - Fundamentals of IPT. It contains several modules. The system is powered via DC power supply which is also trigger oscillator unit. Both units come into power amplifier which is responsible to transform weak high frequency alternating signal to high frequency alternating current in order to excite transmitting coil. Power is transferred wirelessly to receiver coil that is located some distance away from transmitting coil with help of magnetic coupling. Transmitted power is rectified and regulated in order to drive load circuit.

Figure 2 - Fundamentals of IPT

Power transfer can be reached kilowatts range from miliwatts. Power transfer efficiency and reasonable operating range depends on the IPT configuration. One of the phenomenons is coupled magnetic resonance. In contrast to typical IPT configuration, RLC resonance circuits are integrated to both transmitter and receiver sides to enrich power transfer efficiency and transfer range. The difference between traditional IPT and coupled magnetic resonance IPT can be seen in Figure 3. [14]Figure 3 - Topologies of (a) traditional IPT, (b) coupled magnetic resonance,

(c) Strongly coupled magnetic resonance

91

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 3 - Topologies of (a) traditional IPT, (b) coupled magnetic resonance,

(c) Strongly coupled magnetic resonance

There are four combinations for resonant circuits and the choice of configuration depends on the application. Objective is to vanish reactance part of both load and source impedances and to reduce VA rating [6].

Resonant frequency of the system can be calculated as Eq.(Eq.2. [15]

1 1 (Eq.2) ω0 = = √LTxCTx √LRxCRx

Class-E power amplifiers are used to achieve high switching speed in case of high operation frequencies as seen in figure 3.

Figure 4 - Class-E power amplifier circuit[16]

Basically, class E amplifier consists of a MOSFET whose gate is driven by an oscillator. RF choke is connected serial to DC power supply in order to ensure only constant DC voltage is supplied.

92

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

A shunt capacitor which is connected parallel LC resonator circuit is the components left to complete the circuitry. MOSFET acts as an on off switch and theoretically, do not appear at the same time and product of voltage and current is always zero [17]. Efficiency is dependent to power dissipation which is caused by switching losses and reflected signals.

Since efficiency levels rise rapidly, there is growing demand for wireless technology for Industry 4.0 applications such as wireless sensors, wireless charging, devices of measurement and monitoring for rotating components.

3) Capacitive Coupling Case

The basic concept of wireless power transmission (WPT) with capacitively, relies on polarized capacitors. In order to understand the operating logic of capacitors, the fundamental formulas should be investigated which have been demonstrated since past. However, this logic is provided by certain materials which are respectively, one coupling capacitor plates and dielectric materials. There are two different types capacitive power transmission method exists. The general demonstration of these designs was shown in figure 4.

(a) (b)

Figure - 4 a) unipolar demonstration coupling b) Bipolar demonstration of capacitive coupling [18]

However in order to achieve high transferred energy the distance between two coupling capacitors should be minimized as much as possible as aforementioned by Dai, J. et al. [19]. If the starting point is tried to research, the basic formulas for capacitors should be considered as given Equation 3.

QVC= (Eq.3)

V, C and Q are respectively represented the voltage, capacitance and charge. The capacitance of capacitor is calculated with respect to Equation 4.

A C =  d (Eq.4)

In equation, d shows that the distance between two capacitor plates. A represents the cross- sectional area of capacitor plates.  is the permittivity of free space. Generally, equation 4 is not enough in order to evaluate the equivalence capacitance of circuit. Equation 5 and equation 6 are used for evaluating equivalent capacitance of circuit parallel and series connection respectively. Current that can be stored by capacitor is calculated with formula 7.

93

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

CCCCC=++++ ... eqn 123 (Eq.5)

11111 =++++ ... CCCCC eqn 123 (Eq.6)

dv iC= dt (Eq.7)

In order to provide the energy and data transfer, the resonance of tank circuit should be used. The specifications of tank circuit were given on table 2.

Table 2 - General demonstration of components [20]

Component Resistance Reactance Impedance

R 0 Resistor ZRR = 0 =R 0

0 2 fL Inductor ZjLR =  0 = +L 90

Capacitor 0 1 1 − Zc = 2 fC jL 1 = − 900 L

According the last decade development, capacitive coupling method is widely used. Today, ubiquitous capacitive power transfer method was evolved beyond bio-signal, biomedical implementation, imbedded system, wireless charger circuitry, amplifiers, LED, electric vehicles, robotic application [11]. Capacitive coupling technique was used variously applications with respect to these formulas. According to these, capacitive coupling method could be very useful for industry 4.0. For example, the power supply of fully automated systems could be provided by capacitive coupling in order to get rid of wire system which can be easily deformed. The most important advantage of capacitive coupling method is that the electric field lines are used when energy is transferred another capacitor plate. Therefore as mentioned before, unlike magnetic field, electric field lines do not interference through the stability of fully automated imbedded system. Besides, data can be transferred while power is transferred at the same frequency. Therefore, there would not communication interruption in industry 4.0 at information transfer according to executed study from past to present. In the absence of wires, the industry 4.0 becomes quite compact. The cable connection suffers from axial rotation or movement of robotic parts. Especially, capacitive coupling method could be very useful for robotic arm sections. As mentioned before, the data and power can be transferred at the same frequency by capacitive plates. Therefore if capacitive power transfer (CPT) technique will be applied for the data and power requirement of systems, these systems will be cost-effective, compact, and confident which has been shown in below.

94

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure - 5 The capcitive coupling application with robotic arm [30]

With the usage of capacitive coupling technique in robotic arm architecture, the business or task will not interrupted by the damaged wire connections. Additionally, electric field lines will not interference to stability of systems when data and power is transmitted.

4) Backscatter Coupling Case

RFID systems are classified related with the gap between the reader and the transponder. If this gap is greater than 1 meter, they are called long range systems. They are operated at the Ultra-High and microwave frequencies with far field coupling methods and require different amount of energy transfer. If the communication gap is smaller than 20cm, systems are operated with near field coupling methods and generally operated with 13, 56 MHz frequencies [21].

There are many coupling methods for the radio communication such as backscatter coupling, close coupling, electrical coupling and etc. It is critical to choose optimized method related with the suitable application considering the communication range, energy requirements and the property of tracked targets. In our research energy requirement of backscatter coupling and the use of this method are investigated for healthy RFID communication specialized for Industry 4.0 applications.

In physical meaning, backscattering means the reflection of electro-magnetic waves, particles, or signals back to the direction they came from. Reflected weights travel not only in one direction, but also in many other directions [22]. Passive tags and the readers have coil which can cause slight fluctuations. The RF link acts actually as a transformer; as the tag coil is momentarily shunted, the reader winding experiences a momentary voltage drop. This amplitude-modulation loading of the reader’s transmitted field provides a back scatter path to the reader. The data can be modulated in a number of methods [23].

Minimum energy requirements to track a transponder are generally no more than 5 µW levels. This situation changes related with the communication distance and also application specific properties. As an example energy requirement of the RF communication related with the communication distance is seen on table 3 which uses Piezo inside.

95

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Table 3 - Required Energy with Distances [24]

Transmission 30 40 50 60 Distance L(mm)

Required min. 98 169 264 510

Energy Ep (µW)

An electromagnetic field is the combination of an electric field and a magnetic field which oscillates in phase perpendicular to each other and orthogonally to the energy propagation. Electromagnetic waves will propagate an electrical current through any conductor element which it passes. This is where the passive RFID transponders scavenge their energy from. Application specific power demands are seen in figure 6. RFID tags require about 1 to 10 µW levels for the communication.

Figure 6 – Power demands in micro levels [25]

For this coupling method, free space loss is the key parameter for the determination of power demands because it has many reflected wave directions as discussed. Free space loss is a measure of the relationship between the RF power emitted by a reader and the RF power harvested by the transponder.

As a calculation, path loss can be classified as Eq.8 with dB levels [26].

푃푎푡ℎ 퐿표푠푠 (푑퐵) = 40 + [35 log(퐷)] (Eq.8)

D = Indoor distance between transmitter and reader, m

RFID technology with backscattering communication will be widely used in condition monitoring and fault detection applications for Industry 4.0 targets with the help of sensing elements such as vibration, temperature, position, force, flow and etc. For greater autonomy in a production line, all components can be identified with a tag which has less energy consumption for Artificial Intelligence products.

96

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

5) Conclusion

Demands in wireless power transfer is expanding and becoming more attractive especially with Industry 4.0 applications. Power management with circuitry allows energy harvesting from multiple sources such as Piezo elements [27][28] or engine vibration [29] for future researches.

As a result, all coupling methods will be widely used technologies to supply power for artificial intelligent systems. These wireless systems are very compact and reliable for industry 4.0 applications.

Acknowledgements

We are sincerely appreciated for the support provided by Tirsan Kardan A.S. during our research.

References

1. Siemens Global, Industry & Automation. (2015). Web:https://www.siemens.com/innovation/en/home/pictures-of-the-future/industry-and- automation/digitale-fabrik-rfid-in-industry.html 2. Mostafa, T. M., Muharam, A., & Hattori, R. (2017, May). Wireless battery charging system for drones via capacitive power transfer. In Emerging Technologies: Wireless Power Transfer (WoW), 2017 IEEE PELS Workshop on (pp. 1-6). IEEE, 3. Sepahvand, A., Kumar, A., Afridi, K., & Maksimović, D. (2015, July). High power transfer density and high efficiency 100 MHz capacitive wireless power transfer system. In Control and Modeling for Power Electronics (COMPEL), 2015 IEEE 16th Workshop on (pp. 1-4). IEEE, 4. Yi, K. H. (2015). 6.78 MHz capacitive coupling wireless power transfer system,”. Journal of Power Electronics, 15(4), 987-993, 5. Islam, A.B. (2011). Design of wireless power transfer and data telemetry system for biomedical applications, 6. Hassan, M.A. & A. Elzawawi. (2015). Wireless Power Transfer through Inductive Coupling. in Proc. of 19th International Conference on Circuits (part of CSCC'15), 7. Jiang, H., et al. (2012). Safety considerations of wireless charger for electric vehicles—A review paper. in Product Compliance Engineering (ISPCE), 2012 IEEE Symposium, 8. Ali, M. & H. Nugroho. (2016). Effective power amplifier of wireless power transfer system for consumer electronics. in Power System Technology (POWERCON), 2016 IEEE International Conference, 9. Pinuela, M., et al. (2013). Maximizing DC-to-load efficiency for inductive power transfer. IEEE transactions on power electronics. 28(5): p. 2437-2447, 10. Poon, A.S., O'Driscoll, S., & Meng, T.H. (2010). Optimal frequency for wireless power transmission into dispersive tissue. IEEE Transactions on Antennas and Propagation. 58(5): p. 1739-1750, 11. Dai, J., & Ludois, D. C. (2015). Single active switch power electronics for kilowatt scale capacitive power transfer. IEEE Journal of Emerging and Selected Topics in Power Electronics, 3(1), 315-323, 12. Landt, J (2005). The History of RFID. IEEE Potentials. Volume 0278-6648/05. Pages 8-11, 13. Chawla, V. & Sam Ha, D. (2007). An Overview of Passive RFID. IEEE Applications & Practice. Volume 0163-6804/07. Pages 11-17, 14. Qiu, C., et al. (2013). Overview of wireless power transfer for electric vehicle charging. in Electric Vehicle Symposium and Exhibition (EVS27), IEEE, 15. Xia, C., et al. (2012). Comparison of Power Transfer Characteristics between CPT and IPT System and Mutual Inductance Optimization for IPT System. JCP. 7(11): p. 2734-2741, 16. Casanova, J.J., Z.N. Low, & J. Lin. (2009). Design and optimization of a class-E amplifier for a loosely coupled planar wireless power system. IEEE Transactions on Circuits and Systems II: Express Briefs. 56(11): p. 830-834, 17. Sokal, N.O. (2001). Class-E RF power amplifiers. QEX, 2001. 204(1): p. 9-20,

97

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

18. Rozario, D. (2016). Design of contactless capacitive power transfer systems for battery charging applications (PhD. dissertation), 19. Dai, J., &Ludois, D. C. (2015, March). Wireless electric vehicle charging via capacitive power transfer through a conformal bumper. In Applied Power Electronics Conference and Exposition (APEC), 2015 IEEE (pp. 3307-3313). IEEE, 20. Nilsson, W. J., Riedel, N. 2011. Natural and step responses of RLC circuit, Sinusoidal steady-state Analysis. In: Electric Circuits (Gilfillan, A. and Kerman, F.), Pearson Education, pp.286-368, New Jersey, 21. Finkenzeller, K. (2010). RFID Handbook Fundamentals and Applications in Contactless Smart Cards, Radio Frequency Identification and Near Field Communication. West Sussex, United Kingdom: John Wiley & Sons Ltd., 22. Shen, W. (2010). Wireless Power in Passive RFID System. Mikkeli University of Applied Sciences, Bachelor’s Thesis Information Technology. Mikkeli, Finland, 23. Sorrells, P. (1998). Passive RFID Basics. Microchip Technology Inc. USA. DS00680B. Pages 1-5, 24. Fan, Z., Gao, X. R., & Kazmer O. D. (2010). Design of a Self Energized Wireless Sensors for Simultaneous Pressure and Temperature Measurement. IEEE Xplore. DOI:10.1109/AIM.2010.5695931, 25. Adair, N. (2005). RFID Power Budgets for Packaging Applications. Institute of Packaging Professionals, 26. Sayre, C. W. (2008). Complete Wireless Design. New York, NY: The McGraw-Hill Companies Inc., 27. Bilic, G. H., Ozdemir, S. (2018). The Use of Energy Harvesting for Condition Monitoring with RFID in Power Transmission Belts. UMTEB Conference Book. Pages 15-20, 28. Priya, S. & Inman, D. J. (2009). Energy Harvesting Technologies. New York, NY: Springer Science+Business Media, 29. Kinergizer BV. (2018). Delft, the Netherlands. Web: www.kinergizer.com 30. Oh Eunjeon. (2018). Robotic arm, Canada. Web : https://grabcad.com/library/robotic-arm-137

98

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Detection and 3D Modeling of Brain Tumors Using Image Segmentation Methods and Volume Rendering Devrim Kayali1, Ulus Çevik2

Department of Electrical & Electronics Engineering, Çukurova University, Turkey [email protected] [email protected]

Abstract

This paper is on detecting brain tumors using MRI images, and obtaining a 3D model of the detected tumor. With the developed software, image segmentation algorithms were applied to MRI images to separate tumor from healthy brain tissues. In the development phase, various image segmentation algorithms were tried, and high success rates were aimed. After obtaining an algorithm with a high success rate, a 3-dimensional image of the detected tumor will be generated using volume rendering. With this image, features of the tumor such as its location, shape and how it spreads in the brain can be observed.

Keywords: tumor, mri, image segmentation, volume rendering

1) Introduction

A brain tumor is an abnormal growth of cells inside the brain. Tumors are classified as benign or malignant. Tumors with cancer cells are called malignant. They may spread to other tissues, and sometimes to the other parts of the body. If there are no cancer cells the tumor is called benign, they usually do not spread to nearby tissues [1]. There are over 120 brain, and central nervous system (CNS) tumor types which are different for everyone. They form in different areas, develop from different cell types, and may have different treatment options [2].

There are different types of imaging modalities, such as Ultrasound Imaging, Computed Tomography (CT), Magnetic Resonance Imaging (MRI), and X-RAY. But when analyzing the internal structure of the body, like brain, MRI is the most efficient method for detection of the brain tumor [3]. Tumors can be detected from 2D MRI images. However, they will not give the exact information about the shape, and size of the tumor that has to be removed which makes the operation more complex [4].

It is possible to detect a brain tumor automatically using image segmentation methods. But, due to irregular noises, and complex structure of the brain, this task becomes difficult. This means using intensity alone, which is one of the most important features to discriminate between different types of tissues, is insufficient when making segmentation [5]. Because of this, a large variety of segmentation algorithms have been developed.

99

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 1 – 2D MRI Image without tumor.

Figure 2 – 2D MRI Image with tumor.

In the developed software some of these algorithms were used and modified to obtain an accurate segmentation. Then, volume rendering techniques were used to generate a 3D model. Volume rendering means displaying the 2D projection of a 3D dataset which is discretely sampled. In this study these are the MRI slices. After combining image segmentation and volume rendering techniques a 3D model of the brain tumor is obtained [6].

2) Related Work

Methods for tumor detection can be divided into two general groups: with and without ANN (Artificial Neural Network). In this paper image segmentation algorithms without ANN were used.

The morphological examination is one of the methods used for tumor detection. The tumor is detected by examining the structure and shape [7]. Another method used is the Robust Estimation method. With this method, both edema and tumor were detected with pixel density and geometric shape abnormalities [8].

Edge detection is an important technique used in image processing for detection and extraction of features. It is also used in image segmentation, which was used in some studies for marking the tumor area in the MRI image [9] [10] [11]. Later on, volume calculations were done besides detecting the

100

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY tumor. If slice thickness, and pixel spacing attributes are known it is possible to calculate the volume of the tumor after detecting it by using the whole set of the MRI images.

In one of the studies performed in 3D, region growing technique was applied for detecting the tumor. Then the tumor was imaged using volume rendering, and the surface area, and volume of the tumor were also calculated [2]. In [12], FLAIR, T1C, and T2, which are different types of MRI images, were used together instead of using just one type of MRI image. A 3D image of the tumor was obtained by assigning the color values of these 3 different types of images to R, G, B values and, using them at the same time.

In order to increase the success of tumor detection, besides the improvements made in the image processing phase, preprocessing can be done to the images as in [13], [14] and [15]. Algorithms such as median filter, histogram equalization, thresholding are very important because applying them at the preprocess stage greatly increase the success rate at the image processing phase.

In this paper, FLAIR MR images were used to develop the software with image segmentation algorithms without using artificial neural networks. Taken MRI images were processed automatically with the developed algorithms, and when the detection was completed for all of the slices, the tumor could be examined in 3D with the image obtained using volume rendering.

3) Material and Method

Datasets taken from the Cancer Imaging Archive website were used [16]. These are 2D FLAIR MRI slices of a couple of different patients. MRI slices of a patient are seen in Figure 3.

Figure 3 – MRI slices of a patient.

The software was designed using C#, and OpenGL codes. The software takes MRI images as input. First, the noises in the images are cleared using thresholding since noises reduce image quality that makes detection even more complex. Red colored pixels in Figure 4. shows salt and pepper noise. Getting rid of them highly increases the image quality.

101

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 4 – Salt and pepper noise.

To obtain a better image before tumor detection phase, histogram equalization and thresholding are also used. A histogram is a graph showing the numbers of color values in an image. Histogram equalization is also a method used to compensate for chromatic dispersion caused by the clustered color values of an image in a certain place. Thresholding is the display of only the desired color value range in the image.

Figure 5 – Image after histogram equalization.

Figure 6 – Image after thresholding.

102

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

After these methods are applied, additional filtering is done. Purpose of the additional filtering step is to clear more pixels which are not a part of the tumor. The decision for this is made by looking around every pixel to see if they are a part of a bigger object, and cleared if not. At the detection phase radius based detection is applied. A certain radius around every pixel is checked and marked if full. Then whole tumor is filled using the marked pixels. Once the image segmentation has been successfully accomplished for all of the slices, volume rendering is used to obtain the 3D model of the detected brain tumor. Volume rendering calculations are done according to the Regression Based Normal Estimation method. The 3D image is obtained by combining 2D MRI images that went through the image segmentation algorithms.

There must be a defined camera in the space to render the 2D projection of a 3D data set. Also, the opacity, and color of every voxel must be defined. Voxels are like pixels in a bitmap, but voxels represent values in three-dimensional space. RGBA (red, green, blue, alpha) transfer function is used to define the RGBA value for voxels which defines their opacity and color. In this transfer function R, G, B values define the color, and a value that defines the opacity. Figure 9 shows the block diagram of the process.

Figure 7 – Image after additional filtering.

Figure 8 – Detected tumor.

IMAGE VOLUME IMAGE INPUT 3D MODEL SEGMENTATION RENDERING

Figure 9 – Block diagram of the process.

103

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

4) Results After completing all the steps, an output image was obtained. In this image, the detected tumor was marked with red. In the developed software all slices can be inspected by a scrollbar both when the slices are first loaded to the program and when the output images for all of the slices are obtained. Output images can also be saved for using them later. By combining marked images and the volume rendering, the 3D model of the tumor was reconstructed as seen in Figure 11. Saved output images can be loaded any time to inspect, and to get the 3D model.

Figure 10 – Output image.

Figure 11 – 3D Model of the tumor.

5) Conclusion Nowadays, brain tumor cases are quite frequent. Brain tumors are graded from the first grade to the fourth grade, from low to high according to the risk they carry. Therefore, the sooner the tumor is detected and intervened, the better the treatment will be. But, if we have more information about the tumor location, size, and how it spreads inside the brain, the success rate of the operation will be greatly increased and the risk will be minimized. The developed software reconstructs the 3D shape of a detected tumor. It allows getting more information about the tumor in a short period of time. This research is still in progress. The software needs more improvements to detect tumors better. Tumor detection with a higher accuracy, and a better 3D model that allows us to determine the tumor location, shape and how it spreads in the brain is aimed.

104

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

References 1. Angulakshmi, M., & Priya, G. L. (2018). Brain tumour segmentation from MRI using superpixels based spectral clustering. Journal of King Saud University - Computer and Information Sciences. doi:10.1016/j.jksuci.2018.01.009 2. Sudharani, K., Rashmi, K., Sarma, T. C. and Prasad, K. S. (2016). 3D Multimodal MRI Brain Tumor Segmentation: A Volume Rendering Approach. International Journal of Current Trends in Engineering & Research (IJCTER). 3. Samriti and Paramveer Singh. (2016). Brain Tumor Detection Using Image Segmentation. International Journal Of Engineering Development And Research (IJEDR). 4. Narayanan, K. and Yogesh Karunakar. (2018). 3-D Reconstruction of Tumors in MRI Images. 5. Dubey, R. B., Hanmandlu, M., & Gupta, S. K. (2010). An Advanced technique for volumetric analysis. International Journal of Computer Applications,1(1), 91-98. doi:10.5120/13-117 6. Gordillo, N., Montseny, E., & Sobrevilla, P. (2013). State of the art survey on MRI brain tumor segmentation. Magnetic Resonance Imaging,31(8), 1426-1438. doi:10.1016/j.mri.2013.05.002 7. Gibbs, P., Buckley, D. L., Blackband, S. J., & Horsman, A. (1996). Tumour volume determination from MR images by morphological segmentation. Physics in Medicine and Biology,41(11), 2437-2446. doi:10.1088/0031-9155/41/11/014 8. Prastawa, M., Bullitt, E., Ho, S., & Gerig, G. (2003). Robust Estimation for Brain Tumor Segmentation. Lecture Notes in Computer Science Medical Image Computing and Computer-Assisted Intervention - MICCAI 2003,530-537. doi:10.1007/978-3-540-39903-2_65 9. Kumar, S. (2011). Detection of Brain Tumor-A Proposed Method. Journal of Global Research in Computer Science, 2(1), 55-63. 10. Ananda, R. S., & Thomas, T. (2012). Automatic segmentation framework for primary tumors from brain MRIs using morphological filtering techniques. 2012 5th International Conference on BioMedical Engineering and Informatics. doi:10.1109/bmei.2012.6512995 11. Banerjee, M., Chowdhury R. and Kumar S. (2015). Detection Of Brain Tumor From MRI Of Brain. International Journal of Information Research and Review, 2(12), 1555-1559. 12. Mitra, S., Banerjee, S., & Hayashi, Y. (2017). Volumetric brain tumour detection from MRI using visual saliency. Plos One,12(11). doi:10.1371/journal.pone.0187209 13. Isselmou, A. E., Zhang, S., & Xu, G. (2016). A Novel Approach for Brain Tumor Detection Using MRI Images. Journal of Biomedical Science and Engineering,09(10), 44-52. doi:10.4236/jbise.2016.910b006 14. Kanmani M. and Pushparani M. (2016). Brain Tumor Detection And Classification. International Journal of Current Research, 8(5), 31634-31637. 15. Toum, K. M., Mustafa, Z. A., Ibraheem, B. A., & Hamza, A. O. (2017). Brain Tumor Segmentation From Magnetic Resonance Imaging Scans. Journal of Clinical Engineering,42(3), 115-120. doi:10.1097/jce.0000000000000223 16. A growing archive of medical images of cancer. (n.d.). Retrieved from http://www.cancerimagingarchive.net/

105

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Determination of Groundwater Level Fluctuations by Artificial Neural Networks Fatih Üneş, Zeki Mertcan Bahadırlı, Mustafa Demirci, Bestami Taşar

Hakan Varçin, Yunus Ziya Kaya

Department of Civil Engineering, Iskenderun Technical University, Turkey [email protected]

Department of Civil Engineering, Osmaniye Korkut Ata University, Turkey

Abstract Groundwater level change is important in the determination of the efficient use of water resources and plant water needs. Groundwater level fluctuations were investigated using the variable of groundwater level, precipitation, temperature in the present study. The daily data of the precipitation, temperature and groundwater level are used which is taken from PI98-14 observation well station in Minnesota, United States of America. These data, which include information on rainfall, temperature and groundwater level of 2025 daily, were used as input in ANN method. The results were also compared with Multiple Linear Regression (MLR) method. According to this comparison, it was observed that the ANN and MLR method gave similar results for observation. The results show that ANN model will be useful for estimation of groundwater level to monitor possible changes in the future.

Keywords: Ground water level, Artificial neural networks, Multiple linear regression, Modeling

1) Introduction

Groundwater level determination is an important for determining water resources planning. Since the available data generally do not fully reflect the sum of the process, the process needs to be modeled in order to make more reliable decisions. Models can be used to generate data for planning and design, or to estimate the future value of processes.

If the values taken over time by a random variable are independent of each other, the time series of these values is called the stochastic process. In order to determine a stochastic process, it is necessary to indicate the intrinsic dependence between successive elements of the series, except for the probability distribution of the random variable (Bayazıt,1981). Multivariate stochastic analysis and multivariate modeling is an important issue, as planning, design and operation of water resources systems often involve meteorological and hydrological (precipitation, flow, evaporation, groundwater etc.) series (Pegram ve James, 1972).

Artificial neural networks (ANN), which is an artificial intelligence method, is a black box model that is frequently used in hydraulic and water structures planning in recent years. Artificial neural networks collect information about the samples, make generalizations and then make decisions about the samples by using the information they have learned compared to the samples they have never seen before. The artificial neural network model demonstrates the ability to successfully solve complex problems due to its learning and generalization features (Ergezer et all., 2003). Artificial intelligence approaches have been also widely used to in water resource management ( Unes et al (2015a, 2015b), Tasar et al (2017),Demirci et al (2017))

106

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

In this study, the data obtained from the observation well reservoir of PI98-14, near the island of Prairie in the state of Minnesota in the United States of America and The map and satellite images of the study area are shown in Figures 1 and 2, respectively.

Figure 1 - Selected Study area map image

Figure 2 - Selected Study area satellite image

2) Methodology

Multiple Linear Regression method, Artificial Neural Network Method and Data set are introduced in this part of the study. In the all models, Monthly Mean Precipitation (MP), Monthly Average Temperature (MT), Monthly Ground Water Level (GWL+1) were used for the Ground Water Level Estimates.

107

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

2.1. Attificial Neural Networks (ANN)

The superior characteristics of the human brain forced scientists to work on it and inspired by the neurophysical structure of the brain to try to extract the mathematical model. Various artificial cell and network models have been developed with the idea that the physical components should be modeled correctly in order to fully model all the behavior of the brain (Detienne, K. B et all., 2003). Thus, a new branch of science, which is different from the algorithmic calculation method of new and modern computers called Artificial Neural Networks, has emerged. In general, ANN can be defined as a system designed to model the method of performing a function of the brain. The ANN is composed of various forms of artificial nerve cells connected to each other and is usually arranged in layers (Koç, M. L et all., 2004). In accordance with the brain's information processing method, ANN is a parallel scattered processor capable of storing and generalizing information after a learning process (http://www.akademiyapayzeka.com). Artificial neural networks are basically a technology that has been completely exemplified by the human brain (Ergezer et all., 2003).

2.2. Multiple Linear Regression (MLR)

In engineering problems, we see that the values taken by two or more random variables during the same observation are not statistically independent from each other, and therefore, there is a relationship between these variables. For example, the relationship between flow and precipitation in a river basin arises from the effect of flow by precipitation. The relationship between the flows in the neighboring two basins depends on both of them being affected by rainfall in that region (Bayazıt, M., and Oğuz, B., 2005).

The multivariate regression relationship Y is assumed to be influenced by the independent variable of dependent variable X1, X2,…., Xm and if a linear equation is chosen for the relationship between them, the regression equation for Y can be written as follows:

(Eq.1)

a, b1, b2, …., bm regression coefficients are the sum of the squares of the distance from the plane indicated by the regression equation of the observation points, similar to that in the simple regression,

(Eq.2)

is calculated to make a minimum of the expression (Bayazıt, M., and Oğuz, B., 2005).

2.3. Data set used for the study

Figure 3., Figure 4. and Figure 5 are drawn to see the distribution of daily average air temperature (Tort), daily precipitation high and daily groundwater level changes, respectively.

108

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 3 - Daily average air temperature distribution

Figure 4 - Daily precipitation high distribution

Figure 5 - Daily groundwater level distribution

109

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

3) Model Results

Within the scope of the study conducted for the relationship between Precipitation, Temperature and Ground Water Level, a total of 2025 daily data from the station was used. 1419 daily data are used for training models and remaining 606 daily data are used for testing. Mean Absolute Error (MAE), Mean Square Error (MSE) and correlation coefficient (R) statistics are calculated for comparison of methods used. Artificial neural network results and MLR results are compared in Table 1.

Table 1 - Comparison statistics MODELS INPUTS MSE MAE R

MLR MP,MT, GWL+1 0.04 0.11 0,9959 ANN MP,MT, GWL+1 0.14 0.26 0,9852 MSE: Mean square error, MAE: Mean absolute error, R: Correlation coefficient

To see the relationship between created ANN model and observed values distribution graph are drawn in Figure 6 and scatter chart belong to this model was drawn in Figure 7.

Figure 6 - Distribution of ANN model results

Figure 6. shows that distribution of ANN model test results are quite close to observed values of groundwater level for the study area.

110

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 7 - Scatter chart of ANN model results

As it is seen in Figure 7. Correlation coefficient is calculated as 0.985 for test set of ANN method. In distribution and scatter charts, values are close to the actual values.

Distribution of MLR method results and scatter chart is given with Figure 8. and Figure 9., respectively.

Figure 8 - Distribution of MLR method results

111

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 9 - Scatter chart of MLR model results

Results of MLR model show that the correlation coefficient is high and the groundwater level estimate is closer to the actual values shown in figure 8. Correlation coefficient is calculated as 0.996 for MLR results as it is seen in Figure 9.

Conclusion

In this study, the relationship between the precipitation, temperature and groundwater level data of 2025 days of observation station 07040001 in the Goodhue County, Minnesota Reservoir was investigated by Artificial Neural Networks (ANN) method and the obtained values were compared with the Multiple Linear Regression (MLR) method. When the correlation coefficients and error calculations are evaluated it is understood that MLR results and ANN model gave similar results.

References Bayazıt, M. (1981). Hidrolojide İstatistik Yöntemler. İstanbul Teknik Üniversitesi, Yayın No : 1197, 223 s., İstanbul. Bayazıt, M., & Oğuz, B. (2005). Mühendisler için istatistik. Demirci, M., Üneş, F., Kaya, Y. Z., Mamak, M., Tasar, B., Ispir, E. (2017). Estimation of groundwater level using artificial neural networks: a case study of Hatay-Turkey. Detienne, K. B., Detienne, D. H., & Joshi, S. A. (2003). Neural Networks As Statistical Tools For Business Researchers. Organizational Research Methods, 6(2), 236- 265. Ergezer, H.; Dikmen M.,&Özdemir E. (2003). Yapay sinir ağları ve tanıma sistemleri. Pivolka, 2(6), 14-17. Koç, M. L., Balas C. E., & Arslan, A. (2004).Taş Dolgu Dalgakıranların Yapay Sinir Ağları İle Ön Tasarımı.İMO Teknik Dergi, 15(4), 3351-3375. Pegram, G.G.S., & James, W. (1972). Multilag Multivariate Autoregressive Model for the Generation of Operational Hydrology. Water Resources Research, Vol 8, No: 4, 1074- 1076. Tasar, B., Kaya, Y. Z., Varcin, H., Üneş, F., Demirci, M. (2017). Forecasting of Suspended Sediment in Rivers Using Artificial Neural Networks Approach, International Journal of Advanced Engineering Research and Science (IJAERS), 4(12), pp. 79-84. Üneş, F., Demirci, M., & Kişi, Ö. (2015a). Prediction of millers ferry dam reservoir level in usa using artificial neural network. Periodica Polytechnica Civil Engineering, 59, 309–318. Üneş, F., & Demirci, M. (2015b.) Generalized Regression Neural Networks For Reservoir Level Modeling. International Journal of Advanced Computational Engineering and Networking , 3, 81-84.

112

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Dynamic Scheduling in Flexible Manufacturing Processes and an Application Ahmet Kürşad Türker1, Ali Fırat İnal2, Süleyman Ersöz3, Adnan Aktepe4

Department of Industrial Engineering, Kırıkkale University, Turkey [email protected] [email protected] [email protected] [email protected]

Abstract

Flexible manufacturing systems need sub-systems that can be controlled, inspected and traced. Industry 4.0 (the Fourth Industrial Revolution) plays a major role in the creation of smart factory systems. With Industry 4.0 concept and technological improvements, collecting and analyzing each data in the production environment in a proper way enables rapidity in working processes and by this way the use of resources, raw materials, energy will decrease and productivity will increase. Manufacturing with machines communicating among each other and unmanned factories working with operational artificial intelligence that can make the right decisions for production efficiency. In this study, instant automatic data collection is realized by the objects in the system talking to each other in order to solve the daily operating difficulties encountered in the dynamic production processes efficiently. Thus, a system has been created with UHF-RFID (Ultra High Frequency-Radio Frequency Identification) technology to establish a system that can take operational decisions simultaneously on its own.

Keywords: Industry 4.0, Dynamic Scheduling, Simulation, RFID

1. Introduction

Most manufacturing systems operate in dynamic environments. In this dynamic environment, various disturbances can be encountered (such as machine breakdowns, urgent jobs). In response to these unexpected disturbances, we have to reschedule the production plan fast and accurate. In most cases, there is limited time to rescheduling and this is very complicated to do. In this paper, an intelligent system capable of dynamic scheduling is designed and operated under different monitoring technologies. First of all, we will examine the methods and terms used in this study.

1.1. Industry 4.0

The concept of Industry 4.0 emerged in Germany in 2011 and refers to the Fourth Industrial Revolution. In this revolution, with improvements in information and communication technologies digital conversion of production systems is essential. Development of Intelligent Manufacturing Systems by digitization of every stage of the production has generated great interest in the industry. In about 300 years, steam powered mechanical systems has turned into cyber-physical systems. This revolution has direct effects on firms. Cost reduce in maintenance, quality and stock keeping. Time reduce at machines do not work compulsorily and increase in technical staff productivity. Total productivity increase has been aimed because of these gains in the most important elements of production. Machines in most of today's factories are independent of each other and need to be setup for each change. Because of this structure process monitoring and control is rather difficult. Each

113

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY independent machine focuses on its own productivity. The whole system is negatively affected because of this structure, which objects does not talk to each other, and an uncertain structure emerges.

Figure 1 – The process of industrial revolutions [11]

In order to increase the efficiency of systems in your future smart factories, many subsystems needs to be integrated with each another. For this integration, machines that are independent of each other must communicate intelligently. Process monitoring can be done by online data which can be obtained from machine and other units. Thus, machines can signal to correct problems with the ability to stop production.

At job-shop production type workshop is a complex dynamic environment in which unexpected events occur and planned activities change constantly. Due to this dynamism and complexity, planning workflows within the workshop becomes a big problem. This problem is further complicated by the wide variety of products and the fact that each product has its own unique route. Due to complicated routes, production planning cannot be done efficiently and there are delays in delivery time because of unplanned works. Planning and control cannot be provided to increase the efficiency and effectiveness of the entire system. Radio Frequency Identification (RFID) and Barcode system can be used for efficient workshop control.

114

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 2 – Elements of Industry 4.0 [12]

The Elements of Industry 4.0 has given in Figure 2. The range of Industry 4.0 is very wide. Large topics such as Big Data and Internet of Things are subject matter of Industry 4.0.

1.2. Barcode Technology

If we look at the barcode, a barcode is a visual representation of the data that is scanned and interpreted for information. Each barcode contains a certain code, which works as a tracking technology for products; and is represented in a sequence of lines or other shapes. Initially this technology was symbolized by the width and spaces between parallel lines that were one-dimensional. Barcodes can be scanned manually by barcode readers.

115

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Table 1 – RFID and Barcode comparison [14] RFID Barcode High throughput. Multiple (>100) tags Very low throughput. Tags can only Read Rate can be read simultaneously. be read manually, one at a time.

Not required. Items can be oriented in Definitely required. Scanner must any direction, as long as it is in the read physically see each item directly to Line of Sight range, and direct line of sight is never scan, and items must be oriented in required. a very specific manner.

Virtually none. Once up and running, Large requirements. Laborers must Human Capital the system is completely automated. scan each tag. More than just reading. Ability to read, Read only. Ability to read items and Read/Write Capability write, modify, and update. nothing else.

High. Much better protected, and can Low. Easily damaged or removed; Durability even be internally attached, so it can be cannot be read if dirty or greasy. read in very harsh environments.

High. Difficult to replicate. Data can be encrypted, password protected, or Low. Much easier to reproduce or Security include a “kill” feature to remove data counterfeit. permanently, so information stored is much more secure.

Capable. Can be used to trigger certain Not capable. Cannot be used to Event Triggering events (like door openings, alarms, trigger events. etc.).

1.3. Radio Frequency Identification (RFID) Technology

In recent days we have often encountered RFID related work. RFID is a technology of automatically recognizing objects using radio frequency. Basically it comes from a tag and reader.

Figure 3 – RFID elements [13]

RFID tags can be programmed to receive, store and send object information such as the code, route, and coordinates of the part. By reading the tags placed on the product, the information can be automatically saved or changed. Dynamic scheduling techniques can also be used with the information.

116

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

1.4. Dynamic Scheduling Technique

Dynamic scheduling problems in job-shop type manufacturing; are problems related to queuing in a workbench, where the parts visit these workbenches according to their routes. Workbenches in the workshop are service providers in the network. For this reason, the search for solutions to dynamic scheduling problems based on the use of priority rules. Some of the priority rules used in dynamic scheduling can be based not on the unchanging properties of the jobs but on the changing characteristics of the process.

Some of the dynamic scheduling rules: • First in first out (FIFO), the first job waiting in the queue should be selected for processing. • Last in first out (LIFO), the last job waiting in the queue should be selected for processing. • Shortest processing time (SPT), the job which has shortest processing time in the queue should be selected for processing. • Earliest due date (EDD), the job which has closest delivery date in the queue should be selected for processing.

2. Literature Review

Most manufacturing systems work in dynamic environments; in this environment, it works in dynamic environments where inevitable real-time events can cause a change in planned plans and may not be possible when a feasible chart is released into the production area. MacCarthy and Liu (1993) discuss the nature of the difference between planning theory and programming practice as examples of real-time events such as the failure of scheduling theory to respond to machine breakdowns, the arrivals of urgent jobs and recent trends in the timing of the research that tries to make the practical environment more relevant. Shukla and Chen (1996) state that there is very little correspondence between the theory and scheduling practices, in their comprehensive research on intelligent real-time control in flexible manufacturing systems. Cowling and Johansson (2002) pointed to the important difference between scheduling theory and application and showed that scheduling models and algorithms cannot use real- time information.

In the presence of real-time events, the timing problem, called dynamic scheduling, is one of great importance for the success of real-world scheduling systems and RFID is the most effective way to solve this problem. Wang, Liu and Wang (2008) observed the simulated impact of the RFID-enabled supply chain on pull-based inventory replenishment in industry. Guojia and Dingwei (2014) studied a model design and simulation on RFID-enabled hybrid push/pull control strategy. Xiaoju and Dingwei (2016) made a simulation model on RFID-enable control strategy for a supply chain.

In this study, we compared the results numerically by comparing RFID and Barcode on a flexible manufacturing cell.

3. Production Cell

This study was carried out on the production cell at the Computer Integrated Manufacturing (CIM) laboratory, which was established within the Faculty of Engineering of Kırıkkale University. [5]

Production cell is completely autonomous and can work without a human if it is programmed properly. However, in order for the production to proceed in a formal way, the production cell must be constantly monitored and controlled. Both Barcode and RFID technologies are available in the cell. This study was done to observe how the system works better under which technology.

117

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 4 – Layout of the production cell under Barcode Technology

The layout of the production cell under Barcode technology is as shown in Figure 4. As you can see, CNC Lathe, CNC Milling, Assembly, Quality Control and AS/RS are available in the system. In addition, there is a conveyor and two robot arms for parts carriage. Racks and the Common Stock Area are available despite the fact that the CNC machines are busy. Buffers are available in front of robot arms to prevent parts from accumulating on the conveyor. Buffer capacity in front of Robot2 (R2) is 3 while buffer capacity in front of Robot1 (R1) is 2 parts. Also maximum Rack capacity and Common Stock Area capacity is 10 parts.

3.1. Functioning of the Production Cell

In order to compare RFID and Barcode technologies in the production cell, we first need to understand how the cell works. That is why the cell’s working technique is explained below. a. Order entry. b. Moving parts from AS/RS to Conveyor. -Barcode reading process. c. Moving parts to Station R1 with Conveyor. d. Moving parts from Conveyor to CNC with R1. -If CNC is busy, put the part in the queue. e. Milling/Lathe process. f. Moving parts to Conveyor with R1. -Barcode reading process. g. Moving parts to Station R2 with Conveyor. h. Moving parts from Conveyor to Assembly with R2. -If Assembly is busy, put the part in the queue. i. Assembly process. j. Moving parts to Quality Control with R2. -If Quality Control is busy, put the part in the queue. k. Quality Control process.

118

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

l. Moving parts to Conveyor with R2. m. Moving parts to Station AS/RS with Conveyor. n. Package the order.

As you can see, the barcode reading process should be done multiple times, making the system more complex. It coincides with the objectives of Industry 4.0 and makes it too slow for implementing dynamic scheduling techniques. In order to test this proposal, we disabled Barcode readers in the cell and built the RFID technology.

Figure 5 – Layout of the production cell under RFID Technology

If we examine the system in Figure 5 again in terms of workflow, the way it works is as the list below. a. Order entry. b. Moving parts from AS/RS to Conveyor. c. Moving parts to Station R1 with Conveyor. d. Moving parts from Conveyor to CNC with R1. -If CNC is busy, put the part in the queue. e. Milling/Lathe process. f. Moving parts to Conveyor with R1. g. Moving parts to Station R2 with Conveyor. h. Moving parts from Conveyor to Assembly with R2. -If Assembly is busy, put the part in the queue. i. Assembly process. j. Moving parts to Quality Control with R2. -If Quality Control is busy, put the part in the queue. k. Quality Control process. l. Moving parts to Conveyor with R2. m. Moving parts to Station AS/RS with Conveyor. n. Package the order.

119

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

4. Simulation

It is also important to be able to test the production cell under real-world events as well as the creating the simulation model. In order to test the system to real-life conditions, the time between arrivals of orders were determined by Exponential Distribution, which is frequently found in the literature.

When an order enters the system, its delivery date instantly calculated by Normal Distribution. It can also usually determined by Uniform Distribution, both of them can be found in the literature.

The machining times can be differ for each part because of the customers’ specific orders. Obviously, this can be determined by Normal Distribution.

Transportation times, assembly times, quality control times and barcode reading times are determined as constant times. As you can see, these processes are routine and do not change in time.

System pauses and maintenance times have been ignored. It is necessary to carry out maintenance regularly for these types of systems to operate smoothly and this means unexpected system failures going to happen extremely rare. If maintenance time added the simulation, it will not change the results of our performance criteria because our criteria is calculating by percentage values; it will only reduce the finished order quantity and slows down the simulation.

Maximum capacity of AS/RS is 36 parts. At the beginning of the simulation, 18 pieces of initial stock are sending to AS/RS. After this stage, the raw material requirement of the system is provided by the "Order Arrivals" block. Assignments of orders entering the system are made in the "Assignments" block. Here, orders are assigned permanent attributes such as machining times, delivery dates, customer wishes and order code. Then the order should be sent to “Station R1” for processing. For this operation, the AS/RS's own transportation robot is activated, the parts are picked up from the AS/RS and placed in the conveyor.

Figure 6 – The first part of the simulation model

Barcode reading must be done before activating the conveyor, after reading the barcode, the conveyor activates and the parts are sent to “Station R1”.

120

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 7 – CNC part of the simulation model

When a part arrived to “Station R1”, bunch of “Decide” blocks in the simulation model become active and decides to where this part should send. Transportation robot 1 (R1) picks up the part. If the CNC is busy, R1 is trying to put the part in queue, which named “Rack L/M”. If racks are full, R1 is trying to put the part in the “Common Stock Area”. If there is full too, R1 is trying to put the part in the “Buffer 1”. If buffer is full too, the part should stay on the conveyor and system starts to accumulate. Even the conveyor can be operated in the reverse direction, identification of the parts become very difficult at this stage. R1 can easily pick the wrong part and put it in the CNC, this means delivery time will be exceed. If R1 can pick up the wrong part, the probability of picking the wrong part again is heavily increased. In automated production cells like this, we should avoid the picking the wrong part at all costs. One mistake brings another and another. That is why monitoring and control mechanism is essential.

Rest of the simulation model operating like shown in Figure 6 and Figure 7. The simulation model run 24 hours a day for 1 year. The warm-up period has been set to 2 days, and the statistics have been started to be kept at the end of the warm-up period. The simulation model has been run 6 times in total for both RFID and Barcode technologies under three different dynamic scheduling rules (FIFO, SPT, EDD).

For the performance criteria, we used the most frequent ones in the literature.

5. Results

At first glance, RFID technology appears more advantageous than Barcode. When we looked at RFID related works in the literature, it was seen that RFID was superior to Barcode technology in almost all cases. However, in this study we aimed to show the difference between RFID and Barcode

121

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY technologies in a more numerical way. As you can see in Table 2, RFID technology gives better results than Barcode. It is better to lower the values according to our performance criteria.

Table 2 – Simulation Results FIFO SPT EDD Late Delivered Order Ratio (%) 31.97 29.82 32.35 BARCODE Number of Active Orders in the Cell 9.65 9.43 9.66 Average Flow Time (minute) 48.19 47.07 48.22 Late Delivered Order Ratio (%) 21.01 19.65 20.63 RFID Number of Active Orders in the Cell 7.63 7.46 7.63 Average Flow Time (minute) 38.07 37.23 38.05

If we look at dynamic scheduling, SPT technique gives better results than other techniques in this cell. Normally EDD technique known as minimizing the Late Delivered Order Ratio, but in this system the occupancy rate in the queues is not too high.

As you can see, all elements in the system need to communicate with each other. This makes the autonomous systems more efficient. In today’s technology, this can be achieved with RFID.

References 1. MacCarthy, B. L. and Liu, J., (1993), Addressing the gap in scheduling research: a review of optimization and heuristic methods in production scheduling, International Journal of Production Research, 31(1), 59– 79. 2. Shukla, C. S. & Chen, F. F., (1996), The state of the art in intelligent real-time FMS control: a comprehensive survey, Journal of Intelligent Manufacturing, 7, 441–455. 3. Sabuncuoğlu, I. & Bayız, M., (2000), Analysis of Reactive Scheduling Problems in a Job Shop Environment, European Journal of Operational Research, 126(3), 567-586. 4. Cowling, P. I. & Johansson, M., (2002), Using real-time information for effective dynamic scheduling, European Journal of Operational Research, 139(2), 230–244. 5. Ersoz S. (2006) Sabit Otomasyondan Esnek Otomasyona Geçiş, TÜBİTAK ve DPT destekli CIM Laboratuvarı Projesi, Kırıkkale University. 6. Wang, S. J., Liu, S. F. & Wang, W. L., (2008), The simulated impact of RFID-enabled supply chain on pull-based inventory replenishment in TFT-LCD industry, International Journal of Production Economics, 112(2), 570-586. 7. Turker, A. K., (2011), Üretim ve Hizmet Sistemlerinde Simülasyon ve Arena, Ankara Kral Matbaa. 8. Wei, J. & Leung, S. C. H., (2011), A simulation modeling and analysis for RFID-enabled mixed-product loading strategy for outbound logistics: A case study: Computers & Industrial Engineering, 61, 209-215. 9. Guojia, L., & Dingwei, W., (2014), Design and Simulation on RFID-Enabled Hybrid Pull/Push Control Strategy for Multi-Echelon Inventory of Supply Chain, 26th Chinese Control and Decision Conference, 4323-4328. 10. Xiaoju, H. & Wang, D., (2016), Simulation on RFID-enable CONWIP control strategy for multi-echelon inventory of supply chain, Control and Decision Conference, 246-250. 11. Dyson, N., (n.d.)., Working in a smart world - Industry 4.0, https://www.tuv-sud.co.uk. 12. Ar Apps for Industry. (n.d.)., http://www.igs.com.ar/ar-apps-for-industry. 13. Smiley, S., (2014), Monostatic vs Bistatic RFID Systems https://blog.atlasrfidstore.com/ monostatic-vs- bistatic-rfid. 14. AB&R®, Advantages of RFID vs Barcodes, (2018), http://www.atlasrfid.com/jovix-education/auto-id- basics/rfid-vs-barcode.

122

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Dynamic Optimization in A Dynamic Industry Adil Baykasoğlu

Department of Industrial Engineering, Faculty of Engineering, Dokuz Eylül University, Izmir, Turkey

[email protected]

Every engineering problem in life (with appropriate objective function, constraint functions, variables and parameters) can be defined as an optimization problem (The only limitation is our imagination). Things that are not optimally designed (product, system, etc.) either do not work, or work inefficiently and cannot be sustainable. Everything has a life-cycle, the issue is to extend its useful life. The only thing that distinguishes the civil engineer from the other building builders is to make optimally designed buildings. The same thing applies to those who design and operate machines, circuits, plants…However, in many engineering departments optimization is not taught at all! In the majority of optimization problems, it is assumed that the problem parameters, variables, constraints, and problem definition set are precisely known beforehand and do not change during the process. Is this assumption true for real life problems? Certainly not! We are accustomed to hearing this statement: “This job is very URGENT !!!”…or “One of the machines have been broken down. Revise the plans”…similarly, “We got new orders and some were canceled. We need to optimize the plan again. In the meantime, do we have enough time and resources?”. Branke and Schmeck (2003) reported that: "However, almost all publications deal with optimization in static, non-changing environments, whereas many real-world problems are actually dynamic: new jobs have to be added to the schedule, machines may break down or wear out slowly, raw material is of changing quality, etc.". In a more recent study, Nguyen, Yang & Branke (2012) indicated that: "Optimization in dynamic environments is a challenging but important task since many real-world optimization problems are changing overtime." So, what is the dynamic domain (environment)? In this context, what is the Dynamic Optimization Problem (DOP)? What are the main differences with other problems? Some typical examples can be easily found in production. New job arrivals, job cancellations, machine failures, changes in production constraints, changes in lot sizes, delivery times, changes in short-mid-term plans such as job-work assignment, production planning, are some frequently encountered dynamic events in real life production systems. The characteristic of these events is that, they cannot be known precisely and cannot be easily predicted. In general terms; the domains that change over time (time-varying) or change via some events (event- based) or time-varying variants are defined as dynamic environments (Branke, 1999; Branke & Schmeck, 2003). The problems with these features are referred to as dynamic optimization problems (DOPs) in the related literature. Nguyen et al. (2012) define a DOP as "Given a dynamic problem , an optimization algorithm G to solve , and a given optimization period , is called a dynamic optimization problem in the period if during the underlying fitness landscape that G uses to represent changes and G has to react to this change by providing new optimal solutions." Numerous other general definitions for DOPs can be found in the related literature. In this talk, several real life industrial applications of dynamic optimization from our previous studies are presented. These include; dynamic scheduling of heat treatment furnaces, dynamic part family formation for cellular manufacturing, dynamic scheduling of flexible manufacturing systems with flexible transportation abilities, dynamic load consolation for transportation operations, dynamic optimization of several CNC operations etc.

123

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Estimation of Daily and Monthly Global Solar Radiation with Regression and Multi Regression Analysis for Iskenderun Region İsmail Üstün1, Cuma Karakuş2

Department of Mechanical Engineering, Iskenderun Technical University, Turkey [email protected] [email protected]

Abstract

In this study, statistical analysis of monthly and daily meteorological data such as global solar radiation, relative humidity, cloudiness, soil temperature, sunshine duration and air temperature taken from the General Directorate of Meteorology Stations for Iskenderun region have been performed. Solar radiation estimation models have been obtained by using regression and multi regression analysis. Solar radiation estimation models such as linear, quadratic, cubic and multi type have been formed as a result of these analyses. However, multi type regression analysis has been applied only monthly average data. The performance of these new models have been compared with other solar radiation models exist in the literature and have been examined with statistical error analyses As a result, cubic type model gives better performance with a little difference to linear, quadratic and other type model for both monthly and daily estimation models. But multi type model (multi 2) show the best performance with big difference for monthly estimation models. Because multi type of model include much more meteorological parameters than other, it has better estimation capacity. In addition, meteorological parameters have great impact to estimation of solar radiation.

Keywords: Global Solar Radiation, Regression Analysis, Meteorological Data

1. Introduction

In recent years, the using of renewable energy sources has increased in order to meet the energy needs of countries due to the existence of limited fossil resources such as oil, natural gas and coal. As a result of the use of these limited resources as energy, the deterioration of ecology causes destructive consequences such as the accumulation of greenhouse gases that cause air pollution, climate change and perforation of the ozone layer [1]. Renewable energy sources can be shown as solar, wind, geothermal, biogas and hydroelectrical energy and Turkey has large amount of the theoretical potential of these renewable energies [2]. The energy consumption has been increased constantly in recent decades with stability and developing economy in Turkey. Whereas the ratio of the self-sufficiency in the electricity generation was 77% in 1980, this ratio dropped to 25,05% in 2014 [3] and in 2018 about 25 % [4]. The main reasons behind this situation is demanding of large amount of energy, huge amount of hard coal and natural gas import and use of renewable energy in deficient capacity [3]. However, Turkey could play very important role not only in reducing energy needs but also its dependence to other country by benefitting renewable energy unlike using fossil sources that are harmful to environment [5]. The percentage of electricity produced according to energy sources in Turkey is given in Figure 1. As can be seen from the Figure, Turkey has 37,17% of the natural gas, 32,79% of the coal are used to generate electricity from fossil sources. In addition to this, the ratio of renewable energies come from hydroelectric power plants is about 20% and the other renewable energies that contains only 1% solar energy are 10% [6].

124

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Hydropower Liquid fuel 19,58% 0,40% Waste heat Natural gas 1,00% 37,17% Solar 0,97%

Wind 6,02% Geothermal 2,06% Coal 32,79% Figure 1. The percentage of electricity produced according to energy sources [6] Turkey is a country that is in one of the sunny belt area of the world and has more solar potential than many European countries [7]. The estimation of solar radiation in the system design is very important. This knowledge is particularly needed in areas such as engineering, architecture, water and ocean science, and agriculture. [8].

Considering the studies in the literature, Yeşilbudak et al. analysed meteorological data such as global solar radiation, sunshine duration and average air temperature between 2007 and 2016 for province of Ankara. They have developed estimation models using polynomial, Gauss and Fourier curve fitting methods [9]. Karsu et al. have developed estimation models, that is include one-year data, using linear and Gaussian regression analysis in Matlab program for Zonguldak province. To determine the best performing model, statistical analyses have been done and they pointed out that Gaussian regression model showed less error in the solar radiation estimation [10]. Quej et al. have also developed new models using the Gaussian correlation for Yucatán’s solar radiation prediction [11]. Akpadio et al. generated solar radiation prediction models (Angstrom type) by establishing a relation between solar radiation and sunshine duration [12]. Charuchittipan et al. studied geological distribution of solar radiation and mentioned the importance of the amount of global solar radiation in production of biomass [13]. Bailek et al. have investigated the estimation performance by studying the estimation of monthly average diffuse solar radiation and reviewing 35 diffuse correlation for the Algerian region [14]. Nemotollahi and Kim have examined the amount of solar radiation and sunshine duration as monthly and yearly. They found that the area has 0,51 the annual average clearness index between 24 station and stated that the study would play a leading role in the future solar energy systems, projects and studies in this field [15]. Ekici et al. have studied a new estimation model based on Buckingham theorem with the help of the dimensionless pi parameters for some cities of Netherlands. As a result of comparation, they found that models give good performance both short and long term but better in long term [16].

In this study, estimation of monthly and daily global solar radiation models have been analysed thanks to data from General Directorate of Meteorology Stations for Iskenderun region. Linear, quadratic and cubic models have been developed for monthly and daily solar estimation models. However, Multi models have been developed from monthly data for estimation of monthly global solar radiation.

2. The location of Iskenderun and meteorological analysis

Iskenderun is a district in the Mediterranean Region and is located in south of Turkey. Whereas summer is hot and dry, winter is warm and rainy. The meteorological station of the station 17370 has 36,5924

125

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY latitude, 36,1582 longitude and 4 m altitude in Iskenderun. In this study, while the daily data covers 2003-10 (8 years), monthly data covers 2005-10 (6 years) data used in the regression analysis.

The distribution and deviation for the temperature (2015-2016) and radiation value (2005-2016) is given in Figure 2 with the monthly average meteorological data for Iskenderun. Whereas the region has 20.86 °Ϲ average annual temperature, it has 13,86 Mj/m2 average annual solar radiation. The greatest deviations for monthly average temperature are in winter months. However, the amount of lowest deviation is seen in summer month (7th and 8th). The greatest deviation for monthly average solar radiation 4th and 5th months are seen Figure 2.

Figure 2. The distribution and deviation for the temperature and radiation value

The change of meteorological data for Iskenderun region is given in Figure 3 to observe meteorological variation for the region. As can be seen from the Figure, the average monthly maximum sunshine duration is 6th month (10,11 hours) while the minimum sunshine duration is 12th month (6,25 hours). The ratio of annual average monthly humidity is around 60%. In addition, the amount of precipitation and cloudiness is observed in parallel from the Figure and also seen the amount of cloudiness has fallen by half in the first three months.

126

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 3. The change of the meteorological data for Iskenderun region

3. Method

a. Main mathematical expressions about solar radiation

Angstrom-Prescott type equation is generally used for calculating the monthly average daily global solar radiation [17-18]. 퐻 푠 = (푎 + 푏 ) (Eq. 1) 퐻0 푆0

Where H is the monthly average daily global solar radiation on horizontal surface. H0 is monthly average daily extra-terrestrial radiation outside of the atmosphere (calculated with Eq.5). s and S0 are the monthly average daily sunshine duration (calculated with Eq.4) and day length, respectively. a and b are empirical constant which is calculated during regression analysis. The declination angle is the angle of the sun’s rays with the equatorial plane. The declination angle varies throughout the year and is calculated by the following equation.

284+푛 훿 = 23,45푠푖푛 (360 ) (Eq. 2) 365 Where n is the number of days, starting from 1 January. Sunset hour angle is symbolized with 휔푠 and calculated with following equation. Here ∅ is latitude angle of the region.

−1 휔푠 = cos (−푡푎푛∅푡푎푛훿) (Eq. 3)

2 2 푆 = 푐표푠−1 − 푡푎푛∅푡푎푛훿 = 휔 (Eq. 4) 0 15 15 푠 24푥3600 360푛 2휋휔 퐻 = 푥퐺 [1 + 0,033푐표푠 ( )] 푥 [푐표푠∅푐표푠훿푠푖푛휔 + 푠 푠푖푛∅푠푖푛훿] (Eq. 5) 표 휋 푠푐 365 푠 360

Many solar radiation prediction models have been developed in the literature and some of them are given in Table 1- as linear, square, cubic and other types to compare developed models in this study.

127

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Table 1. Solar radiation models

퐻 푠 Model 1 = 0,175 + 0,552 [19] (Eq. 6) 퐻0 푆0

퐻 푠 Model 2 = 0,2456 + 0,3913 ( ) [20] (Eq. 7) 퐻0 푆0

퐻 푠 푠 2 Model 3 = 0,2545 + 0,5121 ( ) − 0,0864 ( ) [21] (Eq. 8) 퐻0 푆0 푆0

퐻 푠 푠 2 Model 4 = 0,148 + 0,668 ( ) − 0,079 ( ) [22] (Eq. 9) 퐻0 푆0 푆0

퐻 푠 푠 2 푠 3 Model 5 = −1,8843 + 12,645 ( ) − 21,871 ( ) + 12,367 ( ) [23] (Eq. 10) 퐻0 푆0 푆0 푆0

퐻 푠 푠 2 푠 3 Model 6 = 0,5121 − 0,4215 ( ) + 0,8012 ( ) − 0,324 ( ) [20] (Eq. 11) 퐻0 푆0 푆0 푆0

퐻 푠 Model 7 = 0,194 + 0,479 + 0,001푇 [24] (Eq. 12) 퐻0 푆0

퐻 푠 Model 8 = −0,107 + 0,70 ( ) − 0,0025푇 + 0,004푅퐻 [25] (Eq. 13) 퐻0 푆0

b. Error analysis methods

A few statistical analysis equations are frequently used in literature to see how much deviation occur between measured and calculated global solar radiation values.

Mean Square Error (MBE) is one of statistical analysis equation which gives information about long- term performance of the system by giving actual deviation between measured and calculated values. The ideal value of the MBE is zero and calculated with following equation [26]. [∑푛 (퐻 −퐻 )] 푀퐵퐸 = 𝑖=1 𝑖,푐푎푙푐푢푙푎푡푒푑 𝑖,푚푒푎푠푢푟푒푑 (Eq. 14) 푛

Mean Percentage Error (MPE) gives the calculation of the percentage average error between the measured and calculated values and calculated following equation [16]. 1 푛 (퐻𝑖,푐푎푙푐푢푙푎푡푒푑−퐻𝑖,푚푒푎푠푢푟푒푑) 푀푃퐸(%) = ∑푖=1 ( ) 푥100 (Eq. 15) 푛 퐻𝑖,푚푒푎푠푢푟푒푑

Root Mean Square Error (RMSE) gives information about short-term performance of the model and result are always positive. Optimum value is zero or the closest to zero [16]. 1 2 1/2 푅푀푆퐸 = [ ∑푛 (퐻 − 퐻 ) ] (Eq. 16) 푛 푖=1 푖,푐푎푙푐푢푙푎푡푒푑 푖,푚푒푎푠푢푟푒푑

Nash-Sutcliffe Equation (NSE) is given in following equation and the optimum value is 1 [27]. 푛 2 ∑𝑖=1(퐻𝑖,푚푒푎푠푢푟푒푑−퐻𝑖,푐푎푙푐푢푙푎푡푒푑) 푁푆퐸 = 1 − 푛 2 (Eq. 17) ∑𝑖=1(퐻𝑖,푚푒푎푠푢푟푒푑−퐻𝑖,푚푒푎푛.푐푎푙푐푢푙푎푡푒푑)

Correlation Coefficient (R2) is a statistical measure that calculates the strength of the relationships between variables and ranges between 0 and 1 [11]. 2 [∑푛 (퐻 −퐻 )푥(퐻 −퐻 ) ] 2 𝑖=1 𝑖,푐푎푙푐푢푙푎푡푒푑 𝑖,ℎ푑표 𝑖,푚푒푎푠푢푟푒푑 𝑖,ö푑표 푅 = 푛 2 푛 2 (Eq. 18) [∑𝑖=1(퐻𝑖,푐푎푙푐푢푙푎푡푒푑−퐻𝑖,ℎ푑표) 푥 ∑𝑖=1(퐻𝑖,푚푒푎푠푢푟푒푑−퐻𝑖,ö푑표) ]

128

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

4. Result and Discussion

a. Regression Analysis

The regression analysis of daily and monthly data taken from General Directorate of Meteorology Station is shown in Figure 4 and Figure 5. The models obtained from regression analysis are given. In this regression analyses, linear, quadratic and cubic regression are given with red, green and blue curve, respectively.

0,8 Linear Quadratic Cubic 0,7 R² = 0.8113 R² = 0.8251 R² = 0.8289 0,6

0,5

0,4 H/Hₒ 0,3

0,2

0,1

0 0 0,2 0,4 0,6 0,8 1 s/Sₒ Figure 4. Regression analysis of daily data

0,6 Linear Quadratic Cubic R² = 0.7903 R² = 0.7906 R² = 0.8013

0,5 H/Hₒ 0,4

0,3 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 s/Sₒ Figure 5. Regression analysis of monthly data

b. Developed regression models

Solar regression models are given below (Eq. 19-21) obtained from monthly data for Iskenderun region.

퐻 푠 Linear (Mod.9) = 0,2231 + 0,4045 ( ) 푅2 = 0,7903 (Eq. 19) 퐻0 푆0 퐻 푠 푠 2 Quadratic (Mod.10) = 0,208 + 0,4633 ( ) − 0,0542 ( ) 푅2 = 0,7906 (Eq. 20) 퐻0 푆0 푆0 퐻 푠 푠 2 푠 3 Cubic (Mod.11) = 0,5134 − 1,413 ( ) + 3,5739 ( ) − 2,2353 ( ) 푅2 = 0,8013 (Eq. 21) 퐻0 푆0 푆0 푆0

129

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Solar multiple regression models are given below (Eq. 22-23) obtained from monthly data for Iskenderun region. The fallowing equations include ST, RH, T, 퐶 and they symbolize monthly average soil temperature, relative humidity, air temperature and cloudiness, respectively.

푠 Multi 1 (Mod.12) 퐻 = −28,575 + 0,233푆푇 + 0,329푅퐻 − 0,235푇 + 0,953퐶 + 29,094 ( ) 푆0 R² = 0,842 (Eq. 22) 푠 Multi 2 (Mod.13) 퐻 = −9,259 − 0,139푆푇 + 0,104푇 + 0,327퐶 + 13,190 ( ) + 0,561 ∗ 퐻0 푆0 푅2 = 0,985 (Eq. 23) Solar regression estimation models are given below (Eq. 24-26) obtained from daily data of Iskenderun region.

퐻 푠 Linear (Mod. 14) = 0,2274 + 0,3943 ( ) 푅2 = 0,8113 (Eq. 24) 퐻0 푆0 퐻 푠 푠 2 Quadratic (Mod. 15) = 0,2005 + 0,6018 ( ) − 0,2261 ( ) 푅2 = 0,8251 (Eq. 25) 퐻0 푆0 푆0 퐻 푠 푠 2 푠 3 Cubic (Mod.16) = 0,1879 + 0,8617 ( ) − 0,9681 ( ) + 0,5435 ( ) 푅2 = 0,8289 (Eq.26) 퐻0 푆0 푆0 푆0

Table 2. The result of the statistical error analysis of monthly solar radiation models Models MBE (Mj/m²) MPE % RMSE (Mj/m²) NSE R² Model 1 1,1229 8,2587 1,5259 0,9173 0,9743 Model 2 0,3475 3,7240 0,8002 0,9773 0,9817 Model 3 1,7962 14,4052 1,9763 0,8614 0,9824 Model 4 1,5090 10,9224 1,9003 0,8718 0,9737 Model 5 0,6031 6,5846 1,5298 0,9169 0,9299 Model 6 0,3550 5,7117 1,0996 0,9571 0,9723 Model 7 1,0402 7,8035 1,4116 0,9293 0,9764 Model 8 1,1609 7,0109 1,7717 0,8886 0,9741 Model 9 -0,0818 0,3891 0,7568 0,9797 0,7903 Model 10 -0,0857 0,3631 0,7527 0,9799 0,7906 Model 11 -0,0817 0,3558 0,7463 0,9802 0,8013 Model 12 0,0222 2,4144 2,1120 0,8417 0,8420 Model 13 -0,0130 -0,0238 0,6447 0,9852 0,9850

The result of the statistical error analysis of monthly solar radiation models for Iskenderun region is given Table 2 with 13 estimation error result. As can be seen from the Table, Model 1 to 8 are selected model from literature and Model 9 to 13 are obtained model via the regression analysis.

Although Model 2 shows very good performance between 8 model from literature, Model 13 (multi-2) is by far the best performing model between all models. Model 12 give less performance between obtained models but better than literature models.

The result of the statistical analysis of daily solar radiation estimation for Iskenderun region is given in Table 3 with 9 estimation error result. When the table is examined, Model 1 to 6 are selected model from literature and Model 14 to 16 are obtained model via the regression analysis.

Model 2 gives better performance between 6 estimation model from literature. When examined obtained model, Model 16 (Cubic type) give the best performance with low error value.

130

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Table 3. - The result of the statistical analysis of daily solar radiation models Models MBE (Mj/m²) MPE % RMSE (Mj/m²) NSE R² Model 1 1,2134 10,1322 2,3293 0,8537 0,9304 Model 2 0,4163 9,3747 1,4956 0,9397 0,9444 Model 3 1,7304 18,8867 2,3180 0,8552 0,9462 Model 4 1,4804 10,4334 2,6675 0,8082 0,9276 Model 5 4,4798 129,6666 15,0283 5,0880 0,5320 Model 6 0,8752 27,8585 2,9097 0,7718 0,8050 Model 14 -0,0655 5,0408 1,4474 0,9435 0,8289 Model 15 -0,0458 4,2037 1,4110 0,9463 0,8251 Model 16 -0,0497 3,9201 1,4066 0,9467 0,8289

5. Conclusion

Before the insolation of solar radiation system project in the investigated region, some meteoritical data are needed such as amount of solar radiation and sunshine duration. Thanks to helping of these data, it is possible to answer the questions about how much energy can be obtained from the solar energy system and how long the system will pay itself. However, finding meteorological data and establishing meteorological stations for each region is costly and very difficult. In addition, these stations have requirements such as calibration, continuous data recording and qualified personnel. Therefore, solar radiation prediction models for the region should be developed. Between the obtained models both month and day, cubic model show better performance with a little difference to linear, quadratic and other type. But multi type of model (multi 2) show the best performance with big difference. Because multi type of estimation models include much parameters such as relative humidity, temperature and cloudiness, they give better performance than other. Because of these, meteorological parameters have great effect on solar estimation models. In addition, obtained models can be used in areas developed specifically for daily and monthly average daily solar radiation estimation or can be used in regions close to or similar climate characteristics to these regions.

References 1. Bulut, U., & Muratoglu, G. (2018). Renewable energy in Turkey: Great potential, low but increasing utilization, and an empirical analysis on renewable energy-growth nexus. Energy Policy, 123, 240-250. 2. Toklu, E. (2013). Overview of potential and utilization of renewable energy sources in Turkey. Renewable Energy, 50, 456-463. 3. Ozcan, M. (2019). Factors influencing the electricity generation preferences of Turkish citizens: Citizens' attitudes and policy recommendations in the context of climate change and environmental impact. Renewable Energy, 132, 381-393. 4. Internet, Eppen, Internet Access Date: 19.10.2018 http://www.eppen.org/index1.php?sayfa=Enerji%20Veri%20Taban%C4%B1&link=link8&makale=93 5. Ozcan, M. (2017). The role of renewables in increasing Turkey's self-sufficiency in electrical energy. Renewable and Sustainable Energy Reviews. 6. Internet, Teias, Access Date:19.10.2018 https://www.teias.gov.tr/sites/default/files/2018- 06/2017%20TE%C4%B0A%C5%9E%20%20Faaliyet%20Raporu.pdf 7. Dincer, F. (2011). Overview of the photovoltaic technology status and perspective in Turkey. Renewable and Sustainable Energy Reviews, 15(8), 3768-3779. 8. López-Lapeña, O., & Pallas-Areny, R. (2018). Solar energy radiation measurement with a low–power solar energy harvester. Computers and Electronics in Agriculture, 151, 150-155. 9. Yesilbudak, M., Colak, M., & Bayindir, R. Ankara İlinin Uzun Dönem Global Güneş Işınım Şiddeti, Güneşlenme Süresi ve Hava Sıcaklığı Verilerinin Analizi ve Eğri Uydurma Metotlarıyla Modellenmesi. Gazi Üniversitesi Fen Bilimleri Dergisi Part C: Tasarım ve Teknoloji, 6(1), 189-203. 10. Karasu, S., Altan A., Sarac, Z., Hacioğlu, R., Prediction of Solar Radiation Based on Machine Learning Methods. The Journal of Cognitive Systems, 2(1), 58-62. 11. Quej, V. H., Almorox, J., Ibrakhimov, M., & Saito, L. (2017). Estimating daily global solar radiation by day of the year in six cities located in the Yucatán Peninsula, Mexico. Journal of cleaner production, 141, 75-82.

131

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

12. Akpabio, L. E., & Etuk, S. E. (2003). Relationship between global solar radiation and sunshine duration for Onne, Nigeria. Turkish Journal of Physics, 27(2), 161-167. 13. Charuchittipan, D., Choosri, P., Janjai, S., Buntoung, S., Nunez, M., & Thongrasmee, W. (2018). A semi- empirical model for estimating diffuse solar near infrared radiation in Thailand using ground-and satellite- based data for mapping applications. Renewable Energy, 117, 175-183. 14. Bailek, N., Bouchouicha, K., Al-Mostafa, Z., El-Shimy, M., Aoun, N., Slimani, A., & Al-Shehri, S. (2018). A new empirical model for forecasting the diffuse solar radiation over Sahara in the Algerian Big South. Renewable Energy, 117, 530-537. 15. Nematollahi, O., & Kim, K. C. (2017). A feasibility study of solar energy in South Korea. Renewable and Sustainable Energy Reviews, 77, 566-579. 16. Ekici, C., & Teke, I. (2018). Developing a new solar radiation estimation model based on Buckingham theorem. Results in Physics, 9, 263-269. 17. Angstrom, A. (1924). Solar and terrestrial radiation. Report to the international commission for solar research on actinometric investigations of solar and atmospheric radiation. Quarterly Journal of the Royal Meteorological Society, 50(210), 121-126. 18. Prescott, J. A. (1940). Evaporation from a water surface in relation to solar radiation. Trans. Roy. Soc. S. Aust., 46, 114-118. 19. Bahel, V., Bakhsh, H., & Srinivasan, R. (1987). A correlation for estimation solar radiation. Energy, 12(2), 131–135. 20. Berkama, B. (2012). Isparta’da yatay düzlem üzerine gelen günlük global güneş radyasyonunun tahmini, Süleyman Demirel Üniversitesi Fen Bilimleri Enstitüsü, Yayımlanmış Yüksek Lisans Tezi, Isparta. 21. Bakirci, K. (2009). Correlations for estimation of daily global solar radiation with hours of bright sunshine in Turkey. Energy, 34(4), 485-501. 22. Aksoy, B. (1997). Estimated monthly average global radiation for Turkey and its comparison with observations. Renewable Energy, 10(4), 625-633. 23. Genç, Y. A. (2015). Osmaniye ili için yatay düzleme gelen güneş radyasyon tahmininde yeni model geliştirilmesi, Osmaniye Korkut Ata Üniversitesi Fen Bilimleri Enstitüsü, Yayımlanmış Yüksek Lisans Tezi, Osmaniye. 24. Sambo, A. S. (1986). Empirical models for the correlation of global solar radiation with meteorological data for Northern Nigeria. Solar & Wind Technology, 3(2), 89–93. doi.org/10.1016/0741-983X(86)90019-6 25. Maghrabi, A. H. (2009). Parameterization of a simple model to estimate monthly global solar radiation based on meteorological variables, and evaluation of existing solar radiation models for Tabouk, Saudi Arabia. Energy conversion and management, 50(11), 2754-2760 26. Bayrakçı, H. C., Demircan, C., & Keçebaş, A. (2017). The development of empirical models for estimating global solar radiation on horizontal surface: A case study. Renewable and Sustainable Energy Reviews. 27. Zhao, N., Zeng, X., & Han, S. (2013). Solar radiation estimation using sunshine hour and air pollution index in China. Energy conversion and management, 76, 846-851.

132

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Gender Estimation from Iris Images Using Tissue Analysis Techniques Tuğba Açıl, Yakup Kutlu

Iskenderun Technical University, Iskenderun,Turkey

[email protected] [email protected]

Abstract Due to the increasing number of people in recent years, the determination of the gender of the person sought when searching in the database means reducing the number of data in the database by half, which will provide us with great convenience in terms of time. In addition, the gender detection system can be used in the creation of marketing strategies that address only a specific gender group in security applications that require gender-based access control. When we look at these, it is seen that the gender detection system has a wide range of applications. Many biometric features such as fingerprint, face, voice and signature are used in person identification systems. However, due to the unique structure of iris, it is thought to be a more reliable system than other biometric properties. Therefore, in this study, gender prediction is tried to be made by using iris structure. The iris images used for gender estimation were taken from the ND_GFI database. 750 women and 750 men, a total of 1500 images were applied on the application. Feature extraction were made with general texture, regional texture and partial texture analysis methods from the iris images. These attributes were classified using the K-Near Neighbor, Naive Bayes, Decision Tree, Multilayer Perceptron classifiers and classification with a performance ratio of 70%. Keywords: Iris, Gender Estimation, Texture Analysis

1. Introduction Biometric systems are systems that are developed by using biometric characteristics of individuals such as face, iris, fingerprint, voice, signature shot shape to determine identity (Şamlı and Yüksek 2009). Biometric features such as fingerprint (Ceyhan,Sağıroğlu and Akyıl 2014)( Ceyhan,Sağıroğlu and Akyıl 2013), walking pattern (Tunalı and Şenyer 2012)(Yu,Tani Huang,Jia and Wu 2009)(Chang and Wu 2010), face (Jain, Huang and Fang 2005)(Guo, Lin and Nguyen 2010)(Stawska and Milczarski 2017), sound (Kotti and Kotropoulos 2008),heart sound (Dal, Coşğun and Özbek 2015), iris (Bansal, Agarwal and Sharma 2014)(Thomas, Chawla, Bowyer and Flynn 2007)(Kuehlkamp,Becker and Bowyer 2017) were used for gender classification in the literature. Iris is a part of the transparent layer of the eye which is behind the transparent layer of the eye and gives its color to the eye. Although iris color is genetic and shows similarity among family members, iris pattern is different in every individual. Iris is a lifelong unchanging organ that is least affected by genetic events and is not affected by hereditary diseases (Çakır, Volkan and Akbulut 2013). Therefore, iris is a highly reliable biometric property. Fig. 1 shows the structure of the human eye.

133

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Fig. 1. Human Eye Anatomy (Amrolkar and Tugave 2015) The gender classification has different uses field, such as security and advertising. For example, it is very important for security forces to search the database for a wanted criminal when halving the data to be searched by sex determination, especially because of the increasing population in recent times. Another area is security applications that require gender-based access control. As another area the gender identification system will be in high demand in the creation of marketing strategies that will address a specific gender group (Tapia and Aravena 2017). Fairhurst at al. (Da Costa-Abreu, Fairhurst and Erbilek 2015) have made a gender estimate on the iris images of 210 people using the BioSecure database. The Daughman algorithm was used for iris normalization. Three different attribute extraction methods have been tried as iris attributes. These are the geometric properties of the iris, the texture properties and the combination of these two attributes. The Gabor Wavelet was used to extraction the texture of the iris. Support Vector Machine, Multilayer Perceptron, Decision Tree, K Nearest Neighbor methods are used for gender classification. Amrolkar et al. (Amrolkar and Tugave 2015) iris images of 356 people from the ND-Iris database were used. Iris segmentation was performed using Circular Hough transformation. Local Binary Pattern, Local Phase Quantization and Uniform Local Binary Pattern methods have been tried for attribute extraction. For gender classification, Decision Trees, K Nearest Neighbor and Support Vector Machine methods were tried. Tapia at al. (Tapia, Perez and Bowyer 2016) were made gender classification on the iris images of 1500 people using GFI_database . After the necessary pretreatments has made such as normalization and segmentation for iris, feature extraction was performed by using Gabor Filter. Mutual information based feature extraction methods were used. In this study, by using the ND_GFI database, after the feature selection with the methods of tissue analysis, the performance rates of gender classification are presented using K Nearest Neighbor, Naive Bayes, Decision Tree and Multi Layer Perceptron classifiers. The diagram in Figure 2 shows the steps followed for gender classification.

Fig. 2. Flow Diagrams of Iris Recognition System

134

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

2. Materials and Methods 2.1.Database The iris images required for the project were taken from the ND_GFI database of the University of Notre Dame. There are iris images of person 1500 (750 women - 750 men) in this database. The images are 480x640 in size and gray scale, obtained with the LG 4000 sensor, there are right and left iris images of each person (Tapia, Perez and Bowyer 2016). 2.2.Iris Segmentation In order to use the images taken in the iris recognition system, some pre-treatments must be performed. In the iris recognition system, process have done using the cross-sectional area between the iris and the pupil. In this study, the radius, center coordinates of the iris and the pupil were determined by using the Daughman algorithm (Oad and Ahmad 2012). Then the within parts of the pupil and outside parts of iris were made in black color. In the study, features were obtained by using fields other than black in this cross - sectional area (Fig. 3).

Iris Segmentation

Fig. 3. Segmentation of the Iris Image

2.3. Feature Extraction In this study, be done tissue analysis on iris images and feature extraction was performed. The structures used when extraction the feature were determined as general iris, regional iris and partial iris regions, respectively, as shown in Fig. 4. Firstly, general tissue analysis was performed from the image of all iris and 7 properties were extracted for each image. Then, each iris was divided into 4 parts and a total of 28 properties were extracted and regional tissue analysis was performed. Finally, the iris image was divided into 16 parts and with a total of 112 features were made partial texture analysis. These 7 properties are the max, min, mean, standard deviation, variance, skewness and skewness values of iris.

a) b) c) Fig. 4. Structures Used to Feature Extraction a) General Iris b) Regional Iris c) Partial Iris Regions

135

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

2.4. Classification and Performance Criteria

Classification is the process of distributing new incoming data into defined classes from the classes defined in a dataset. In this study, gender classification was made by K-NN, Naive Bayes, Multilayer Perceptron and Decision Tree classification methods. In the K-Near Neighbor (K-NN) algorithm, all of the samples in the training set are kept in a n- dimensional sample space. When a new sample is received, it is decided to determine the k samples that are closest to the new sample coming from the training set and to look at the class label of the new sample and the majority of the class k of the nearest neighbor (Taşcı and Onan 2016). The Naive Bayes classification detect the class of new data presented to the system through transactions made on the data presented to the system for teaching purposes. Naive Bayes has a simple structure and is a high-performance classification algorithm (Orhan and Adem, 2012). Multilayer Perceptron is the model in which the outputs are expected to be produced in response to inputs. Multilayer Perceptron consists of input layer, hidden layer and output layer. The information in the input layer is transferred to the hidden layer. The information from the input layer is processed in the hidden layer. In the output layer, the output values for the inputs are calculated (Gör 2016). The Decision Tree algorithm is a powerful classifier that allows users to interpret unknown data. The Decision Tree structure consists of nodes in which root node, leaf node and the links are made by branches. Starting from the root node, the decision nodes are tested one by one and go to the leaf node (Alan 2014). Confusion Matrix is used on a test data where the actual data is known to define the performance of the classification models used in machine learning (Table 1).

Table 1. Confusion Matrix Predicted Values Positive Negative Positive True Positive (TP) False Negative (FN) Actual Values Negative False Positive (FP) True Negative (TN)

According to the definitions in this confusion matrix, the overall performance ratio of the result is calculated by the following formula.

Accuracy = 3. Results In this study, gender classification was made by using the attributes extracted from iris images with a performance ratio of 70%. The low performance rate may be due to the high loss of information and the lack of meaningful information in the extracted attributes. When other studies in the literature are examined, feature extraction was performed using different transform techniques and good results were obtained. In this study it is thought to improve the result using gabor filter. In the study, firstly the features obtained from the whole picture, then the features obtained by dividing the picture into 4 parts and the features obtained by dividing the picture into 16 parts were subjected to the classification. Finally, the properties obtained separately by combined and were made classification. Figure 7 shows the steps to classify iris.

136

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

1x7 Features 4x7 Features 16x7 Features

All Feature

Classification

Feature Selection

Fig. 7. Flow Diagram of the Gender Classification Model

Table 2 shows the performance ratios obtained from the different features and the classifiers. In this table, S1 is the combination of seven features from the whole picture, S2 is the combination of 28 properties obtained by dividing the picture by four, S3 is the combination of 112 features obtained by dividing the picture by sixteen, and SAll is the combination were obtained of all feature. Finally, classification were made by feature selection from these features.

137

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Table 2. Classification Performance Table FEATURES CLASSIFIERS S1 S2 S3 SAll After Feature Selection K-NN (KStar) %60.2 %50.8 %58.1 %61.3 %61.8 DECISION TREE(J48) %61.6 %51.6 %57.2 %63 %64.4 NAIVE BAYES %62.4 %50.6 %53.8 %61.4 %68.6 MULTILAYERPERCEPTRON %62.4 %52.8 %59.2 %67.3 %66.6

References:

Alan, M. A. (2014). Karar ağaçlarıyla öğrenci verilerinin sınıflandırılması. Atatürk Üniversitesi İktisadi ve İdari Bilimler Dergisi, 28(4).

Amrolkar, K. ve Tugave, A.S., (2015). Gender Classification from Iris Using Machine Learning Techniques

Bansal, A., Agarwal, R., & Sharma, R. K. (2014). Predicting gender using iris images. Research Journal of Recent Sciences, 3(4), 20-26.

Ceyhan, E. B., Sağiroğlu, Ş., & Akyil, M. E. (2013, April). Statistical gender analysis based on fingerprint ridge density. In Signal Processing and Communications Applications Conference (SIU), 2013 21st (pp. 1-4). IEEE.

CEYHAN, E. B., SAĞIROĞLU, Ş., & AKYIL, E. (2014). PARMAK İZİ ÖZNİTELİK VEKTÖRLERİ KULLANILARAK YSA TABANLI CİNSİYET SINIFLANDIRMA. Journal of the Faculty of Engineering & Architecture of Gazi University, 29(1).

Chang, C. Y., & Wu, T. H. (2010, November). Using gait information for gender recognition. In Intelligent Systems Design and Applications (ISDA), 2010 10th International Conference on(pp. 1388-1393). IEEE.

Çakır, A., Volkan, A., & Akbulut, F. T. (2013). Iris Tanıma Sistemleri ve Uygulama Alanları. 5. Akademik Bilişim Konferansları, 13.

Da Costa-Abreu, M., Fairhurst, M., & Erbilek, M. (2015, September). Exploring gender prediction from iris biometrics. In Biometrics Special Interest Group (BIOSIG), 2015 International Conference of the (pp. 1-11). IEEE.

Dal, F., Coşğun, S., & Özbek, İ. Y. (2015, May). Gender detection with heart sound. In Signal Processing and Communications Applications Conference (SIU), 2015 23th (pp. 2362-2365). IEEE.

Gör, İ. (2016) Çok Katmanlı Algılayıcı Yapay Sinir Ağı ile Lineer Diferansiyel Denklem Sisteminin Çözümü XVIII. Akademik Bilişim Konferansı

Guo, J. M., Lin, C. C., & Nguyen, H. S. (2010, July). Face gender recognition using improved appearance-based average face difference and support vector machine. In System Science and Engineering (ICSSE), 2010 International Conference on (pp. 637-640). IEEE.

Jain, A., Huang, J., & Fang, S. (2005, July). Gender identification using frontal facial images. In Multimedia and Expo, 2005. ICME 2005. IEEE International Conference on (pp. 4-pp). IEEE.

Kotti, M., & Kotropoulos, C. (2008, December). Gender classification in two emotional speech databases. In Pattern Recognition, 2008. ICPR 2008. 19th International Conference on(pp. 1-4). IEEE.

Kuehlkamp, A., Becker, B., & Bowyer, K. (2017, March). Gender-From-Iris or Gender-From-Mascara?. In Applications of Computer Vision (WACV), 2017 IEEE Winter Conference on (pp. 1151-1159). IEEE.

Percy, O., & Waqas, A. (2010). Iris localization using Daugman’s algorithm. Google Scholar, 1-48.

138

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Orhan, U., & Adem, K. (2012) Naive Bayes Yönteminde Olasılık Çarpanlarının Etkileri The Effects of Probability Factors in aive Bayes Method. Ionosphere, 351, 34.

Stawska, Z., & Milczarski, P. (2017). Support vector machine in gender recognition. Information Systems in Management, 6.

Şamlı, R., & Yüksel, M. E. (2009). Biyometrik güvenlik sistemleri. Akademik Bilişim, Şanlıurfa-Türkiye.

Tapia, J. E., Perez, C. A., & Bowyer, K. W. (2016). Gender classification from the same iris code used for recognition. IEEE Transactions on Information Forensics and Security, 11(8), 1760-1770.

Tapia, J., & Aravena, C. (2017). Gender classification from NIR iris images using deep learning. In Deep Learning for Biometrics(pp. 219-239). Springer, Cham.

Taşcı, E., & Onan, A. (2016). K-en yakın komşu algoritması parametrelerinin sınıflandırma performansı üzerine etkisinin incelenmesi. Akademik Bilişim.

Thomas, V., Chawla, N. V., Bowyer, K. W., & Flynn, P. J. (2007, September). Learning to predict gender from iris images. In Biometrics: Theory, Applications, and Systems, 2007. BTAS 2007. First IEEE International Conference on (pp. 1-5). IEEE.

Tunalı, İ., & Şenyer, N. (2012, April). Gender recognition from gait using RIT and CIT approaches. In Signal Processing and Communications Applications Conference (SIU), 2012 20th (pp. 1-4). IEEE.

Yu, S., Tan, T., Huang, K., Jia, K., & Wu, X. (2009). A study on gait-based gender classification. IEEE Transactions on image processing, 18(8), 1905-1910.

139

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Interactive Temporal Erasable Itemset Mining İrfan Yildirim, Mete Celik

Department of Computer Engineering, Erciyes University, Turkey [email protected] [email protected]

Abstract

Mining erasable itemset is one of the popular extension of frequent itemset mining. In time several algorithms have been presented to solve the problem of mining erasable itemset, efficiently. However, all these studies does not take into account the time information of the erasable itemsets. In this study, the concept of the temporal erasable itemset mining is first introduced. Then two algorithms named ITEMETA and ITEVME algorithms are proposed to discover a complete set of temporal erasable itemsets, interactively. The comparison of these two algorithms in terms of running time is also presented. Keywords: erasable itemset mining, itemset mining, temporal mining, interactive mining

1. Introduction

Frequent itemset mining (FIM) (Agrawal & Srikant, 1994) is one of the most popular topics in the data mining area. It was introduced as the problem of frequent itemset mining which aims to discover the complete set of frequently occurred itemsets in a given transactional database. Erasable Itemset Mining (EIM) is an interesting variant of FIM, originates from production planning (Deng, Fang, Wang & Xu, 2009). The goal of EIM is the find all itemsets satisying with a given threshold in a product database. Assume that a factory produces many type of products and each product compose of a number of items (raw metarials). The factory needs to puchase all these items in order to fabricate all the products. However, for some reason, the factory may experience financial problems, thus it may lose the ability of supplying all the items as usual. In such case, unfortunately, the factory should decide which items should not be purchased anymore for sustainability of the factory. In other words, the managers should make a desicion about manufacturing of which products will be stopped. However, this desicion is hard because the factory earns different profits from the sale of different products. Besides, it is very complex and challenging especially if the factory produces a large number of products and some products are depended to some common items (raw materials). On the other hand, it is very crucial for the sustainability of the factory.

The problem of EIM was first introduced and solved by the META algorithm (Deng et al., 2009). Since then, many algorithms have been presented fort the efficiency of the problem of EIM, such as VME (Deng & Xu, 2010), MERIT (Deng & Xu, 2012), dMERIT+ (Le, Vo & Coenen, 2013), MEI (Le & Vo, 2014). However, all these studies assume that each product is produced throughout the year. But, this is not true for many real-world manufacturers. For example, some products may only seasonally produced. Now, consider that the managers predict that the factory will solve its financial problems in three months, and they try to make the most effective decision about stopping the supplying for which products in three months. In such case, all of the above-mentioned studies are insufficient to respond to this problem. Therefore, designing an algorithm handling the EIM with in a time interval is important issue.

In this study, we first introduce the concept of the temperol erasable itemset mining (TEIM). Then, to discover the problem of TEIM, interactively, two algorithms are proposed named ITEMETA and ITEVME based on classic erasable itemset mining. These two algorithms allows a user to mine

140

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY itemsets which are erasable in a different time intervals and/or different temporal profit thresholds, interactively. Here, while saying interactive, we mean a data structure that allows many mining request after it is once constructed. Because, the managers may want to know and compare erasable itemset for different time intervals under different profit threshold values to make the best desicion. According to our knowledge, this is the first study that considers the renevues of products with time information.

The rest of the study is organized as follows. Section 2 gives the basic concept and the problem definition. Section 3 describes the concept of TEIM and proposed algorithms. Section 4 presents the runtime analysis of the proposed algorithms. Finally, Section 5 concludes the paper with some future study suggestions.

2. Basic Concept

In this section, some preliminary definitions and properties related to problem of erasable itemset mining are briefly given based on META algorithm (Deng et al., 2009).

Assume that the factory suplies m items and let I= {i1, i2, …, im} be a list that includes these items, where each item ik (1 ≤ k ≤ m) is a raw meterial of at least one product in its product database. Let TPDB= {P1, P2, …, Pn} be the product database, where each product Pk (1 ≤ k ≤ n) is in the form of , where pid is the product identifier of Pk, items are components of Pk, and gain is the profit when the factory earns by selling Pk, and time indicates that at which time interval the factory earns the gain by selling Pk.

Consider Table 1 as an example TPDB of a factory. There are 10 products, 8 different items, and 4 different time stamps in the dataset. Each product consists of various combinations of items. For example, to manufacture the product P5, the factory needs three items which are A, C and G, together. On the other hand, the factory earns 100 million $ by selling P5 at the time T2.

Definition 1: Let X  I be an itemset. The gain of X (Gain(X)) is calculated as follows:

( ) ∑ 푮풂풊풏 푿 = {푃푘|푃푘 .Items  X.Items = Ø } 푃푘. 푉푎푙 (Eq. 1)

For example, let say X consists of items A and C, X = {A, C}. Based on example dataset given in Table 1, P1, P2, P3, P4, P5, and P6 are the products that includes at least one of X`s items. Hence, the Gain(X) = P1.Val + P2.Val + P3.Val + P4.Val + P5.Val + P6.Val = 800 million $.

Table 1 – A sample product database, TPDB PID Items Val (million $) Time P1 {A, B, C} 140 T1 P2 {A, B, D} 60 T1 P3 {A, C, D} 200 T1 P4 {A, C, D, E} 180 T2 P5 {A, C, G} 100 T2 P6 {C, H} 120 T2 P7 {D, E} 100 T3 P8 {D, G} 300 T3 P9 {F, H} 700 T4 P10 {G, H} 500 T4

Definition 2: The total gain of the TPDB (TotalGain(TPDB)) is calculated as follows:

( ) ∑ 푻풐풕풂풍푮풂풊풏 푻푷푫푩 = {푃푘|푃푘 ∈ PDB} 푃푘. 푉푎푙 (Eq. 2)

141

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

For example, the total profit a factory is calculated by summing of gains of all products of the factory. So, TotalGain(TPDB) = P1.Val + P2.Val + P3.Val + P4.Val + P5.Val + P6.Val + P7.Val + P8.Val + P9.Val + P10.Val = 2400 million $.

Definition 3: Considering a pre-defined maximum profit threshold maxT (in %). An itemset X is erasable if:

푮풂풊풏(푿) ≤ 푇표푡푎푙퐺푎푖푛(푃퐷퐵)푥 maxT (Eq. 3)

For example, assume that maxT is set to 30%. Therefore, X = {A, C} is not a erasable itemset since ( Gain(X) = 800 ) > ( 720 = 2400 x 30 / 100 ).

Definition 4: Based on the definitions given above, the problem of erasable itemset mining is described as finding all erasable itemsets whose gains are less than TotalGain(PDB) x maxT in a given database.

Property 1: Let X  I and Y  I be two itemsets. If X  Y then (X) ≤ Gain (Y) (Deng et al., 2009).

Property 2: (anti-monotone property). Based on Property 1, if X is not an erasable itemset then Y cannot be an erasable itemset (Deng et al., 2009).

Example: Since Gain({A, C}) = 800 > 720, {A, C} amd none of {A, C}’s supersets can be EI. Therefore, the search space can be pruned fort he supersets of {A, C}.

3. Interactive Temporal Erasable Itemset Mining

However, for some reasons, the managers of the factory may only be interested in itemsets which are erasable within a certain time period. Therefore, in such a case, the results of classical EIM algorithms may be insufficient or useless to help managers make decisions. The below definitions are related to temporal erasable itemset mining.

Definition 5: Consider that a TPDB has N different time information. A time interval tInterval can be expressed as [TSTART – TEND], such that 1 ≤ TSTART ≤ TEND ≤ N.

For example, the TPDB in given Table 1 has 4 different time information which are T1, T2, T3, and T4. The [T1 – T3] represents a time interval from T1 to T3. Therefore, the tInterval consist of T1, T2 and T3 for this example. If the tInterval is expressed as [T1 – T1], then it represents the mining need is related only these products which are profitable in T1. For the running example, there are 10 possible time intervals which are [T1 – T1], [T1 – T2], [T1 – T3], [T1 – T4], [T2 – T2], [T2 – T3], [T2 – T4], [T3 – T3], [T3 – T4], and [T4 – T4].

Definition 6: Let X  I be an itemset. The gain of X in the tInterval (Gain(X, tInterval)) is calculated as follows:

( ) ∑ 푮풂풊풏 푿, 풕푰풏풕풆풓풗풂풍 = {푃푘|(푃푘 .Items  X.Items = Ø ) ∧ (푃푘.푇푖푚푒 휖 푡퐼푛푡푒푟푣푎푙)} 푃푘. 푉푎푙 (Eq. 4)

For example, Gain ({D, E}, [T2 – T3]) = P7.Val + P7.Val + P8.Val = 580.

Definition 7: The total temporal gain of the PDB in the tInterval (TotalTempGain(PDB, tInterval)) is calculated as follows:

( ) ∑ 푻풐풕풂풍푻풆풎풑푮풂풊풏 푷푫푩, 풕푰풏풕풆풓풗풂풍 = {푃푘|(푃푘 ∈ PDB) ∧ (푃푘.푇푖푚푒 휖 푡퐼푛푡푒푟푣푎푙)} 푃푘. 푉푎푙 (Eq. 5)

142

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

For example, TotalTempGain(PDB, [T2 – T3]) = P4.Val + P5.Val + P6.Val + P7.Val + P8.Val = 800.

Definition 8: Based on the definitions given above, the problem of temporal erasable itemset mining is described as finding all erasable itemsets whose gains are less than TotalTempGain x maxT in a given product database for in the related tInterval.

Note that, Property 1 is also hold in temporal erasable itemsets, so that Property 2 can be used to prune search space of TEIM.

While saying interactive mining we refer that construct a data structure once for a given TPDB and then mine many time using the constructed data structure. In the interactive TEIM, the database is stable, but time intervals and/or the threshold may changed interactively.

Therefore, to mine temporal EIs interactively, we proposed two algorithms which are ITEMETA and ITEVME. These proposed algorithms are the extensions of classic META and VME algorithms, respectively. The following subsections describes proposed algorithms.

3.1 The proposed ITEMETA algorithm Basically, the ITEMETA was designed based on the working principle of the META. However, there are differences between them. The major difference is that items are stored with their all necessary information in a data structure named ITIL (Item Time Information List). Besides, profits earned by the factory in different time stamps are kept in another data structure named TGL (Time Gain List). Thanks to ITIL and TGL, the ITEMETA easily deals the interactive TEIM. Moreover, the result of ITEMETA is the set of itemsets which are erasable in a given tInterval under a pre- defined threshold. The pseudocode of the ITEMETA is given in Algorithm 1. Algorithm 1: ITEMETA Algorithm Input: a produt database TPDB, Output: the set of all erasable itemsets EIs 1. For each procuduct P in TPDB 1.1. If TGL does not contain P.Time then - TGL.Add(P.Time); 1.2. TGL(P.Time).Gain = P.Val: 1.3. For each item I in P 1.3.1. If ITIL does not contain I then - ITIL.Add(I); 1.3.2. ITIL(I).Times does not contain P.Time then - ITIL(I).Times.Add(P.Time); 1.3.3. ITIL(I).Times(P.Time).Gain += P.Val;

// for each mining request, run followings lines. 3. Save tInterval and maxT (%) based on the mining request. 4. Calculate the TotalTempGain using the TGL based on tInterval. 5. Th = TotalTempGain x maxT (%); //Max. profit threshold 6. Obtain the TEI1 using the ITIL; // each item in TEI1 has a gain which is not greater than the Th value in the tInterval. 7. EIs = META(TEI1, Th, tInterval);

The ITEMETA starts with a database scan to construct TGL and ITIL data structures. The constructed TGL and ITIL for the TPDB given in Table 1 are illustareted in Figure 1 and Figure 2, respectively.

143

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 1: The TGL representation of the TPDB given in Table 1

From the TGL given in Figure 1, we understand that the factory has gains at times T1, T2, T3, and T4 which are 400, 400, 400, and 1200, respectively. Therefore, the gain of the factory for any time interval can easily be calculated using the TGL. For example, TotalTempGain(PDB, [T3 – T4]) is calculated as TGL(T3).Gain + TGL(T4).Gain = 400 + 1200 = 1600.

Figure 2. The ITIL representation of the TPDB given in Table 1

From the constructed ITIL, we can easily understand which items have how much gains for a time interval. For example, let time interval ve [T2 – T3]. Therefore, Gain({D}, [T2 – T3]) is obtained as ITIL(D).Times(T2).Gain + ITIL(D).Times(T3).Gain = 180 + 400 = 580.

After the TGL and ITIL are constructed once, EIs can be discovered over and over for different time intervals under different thresholds based on the mining requests. When a mining request is occurred, the algorithm calculates the maximum profit threshold Th based on the time interval tInterval and the maxT (%) values related to the mining request. To calculate the Th, the TGL data structure is used. Next, utilizing the ITIL data structure, the set of temporal erasable 1-itemsets (TEI1) in the tInterval are obtained. After that, the mining phase of the algorithm is processed to obtain the set of all erasable itemsets in the tInterval. For this reason, the META algorithm is called with three parameters which are TEI1, Th, and tInterval. For algorithm steps of the META, please refer to original paper (Deng et al., 2009). However, our META implementation has two differences comparing the original algorithm. First, our implementation starts with generating the set of temporal candidate erasable 2-itemsets (TECI2) since it is already called with TEI1. Therefore, it saves the time which is required for determining TEI1. Second, it scans the database ignoring the products which are not related to given tInterval.

Overall, The ITEMETA explores TEIs in the following manner. At first, the given TGL is scanned to calculate the total gain of the factory for the given tInterval, and the Th is calculated. Afterwards, utilizing ITIL, based on the threshold value, the set of temporal erasable 1-itemsets (TEI1) are filtered. Next, items in TEI1 are joined to generate the set of temporal candidate erasable 2-itemsets (TECI2) and a database scan is performed to obtain the temporal gains of each item in TECI2 in the related tInterval. Then, TEI2 are filtered from the TECI2. Next, itemsets in TEI2 are joined to generate TECI3 to find TEI3, and so on. This level-wise iterative search approach utilizing the join and prune strategies until no TECIN+1 can be found.

3.2 The proposed ITEVME algorithm

144

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

In this section, the proposed ITEVME algorithm is described. The ITEVME is an extension of VME algorithm to handle the problem of interactive TEIM. Before, describing the ITEVME, the PID_lists structure presented by VME is mentioned.

Definition 9: For any itemset X  I, the PID_lists of the X contains each pair of values such that for each Pi, where X  Pi.Items.

For example, let say itemset X consists of items A and C, X = {A, C}. Based on example dataset given in Table 1, at lease one item from{A, C} is occurred in products P1, P2, P3, P4, P5, and P6. Thefore, The PID_lists of {A,C} is a list of elements obtained from these products, such that PID_lists of {A,C} = (<1:140> , <2:60>, <3:200>, <4:180>, <5:100>, <6:120>).

Definition 10: The gain of an itemset X can be obtained by summing of profit of each PID_list of its PID_Lists. For example, the gain of {A,C} can be easily calculated utilizing its PID_lists. The gain of {A,C} is the sum of 140 + 60 + 200 + 180 + 100 + 120 = 800.

Definition 11: The PID_lists of an itemset X is the union of the PID_lists of all items that X conntains of. For example, The PID_lists of {A, D} is (<1:140> , <2:60>, <3:200>, <4:180>, <5:100>, <7:100> , <8:300> ). The PID_lists of {A, C} is given above example. By joining the PID_lists{A, D} and the PID_lists{A, C} and, we can obtain the PID_lists{A, C, D} as (<1:140> , <2:60>, <3:200>, <4:180>, <5:100>, <6:120>, <7:100> , <8:300>).

However, to deal the interactive TEIM, ITEVME algorithm stored each PID_list of each item with their time information. For this reason, we propose the ITIPIDL (Item Temporal Information PID_lists) data structure. Besides, TGL is also used. The pseudocode of the ITEVME is given in Algorithm 2.

The ITEVME starts with a database scan to construct TGL and ITPIDL data structures. The constructed TGL is the same as given in Figure 1. The ITPIDL for the TPDB given in Table 1 are illustareted in Figure 3.

Figure 3. The ITPIDL representation of the TPDB given in Table 1

145

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

From the constructed ITPIDL, we can easily understand which items have how much gains for a time interval. Besides, we also understand which items are used in manufacturing of which products. For example, let time interval be [T2 – T3]. Therefore, Gain({D}, [T2 – T3]) is obtained as ITIL(D).Times(T2).Gain + ITIL(D).Times(T3).Gain = 180 + 400 = 580. Besides, we can easily obtain the PID_lists of the item D in the tInterval [T2 – T3] as (<4:180>, <7:100>, <8:300>).

Algorithm 2: ITEVME Algorithm Input: a produt database TPDB, Output: the set of all erasable itemsets EIs 1. For each procuduct P in TPDB 1.1. If TGL does not contain P.Time then - TGL.Add(P.Time); 1.2. TGL(P.Time).Gain = P.Val: 1.3. For each item I in P 1.3.1. If ITPIL does not contain I then - ITPIL.Add(I); 1.3.2. ITPIL(I).Times does not contain P.Time then - ITPIL(I).Times.Add(P.Time); 1.3.3. ITPIL(I).Times(P.Time).Gain +=P.Val; 1.3.4. ITPIL(I).Times(P.Time).Add ();

// for each mining request, run followings lines. 8. Save tInterval and maxT (%) based on the mining request. 9. Calculate the TotalTempGain using the TGL based on tInterval. 10. Th = TotalTempGain x maxT (%); //Max. profit threshold 11. Obtain the PID_Lists of each item i that is erasable in the [Tstart – Tend] from its ITPIDL // ∀i ∈ TEI1 12. EIs = VME(PID_lists of each erasable item, Th);

After the TGL and ITPIL are constructed once, EIs can be discovered with different time intervals and/or different thresholds values, repeatedly. When a mining request is occurred, the algorithm calculates the maximum profit threshold Th as in ITEMETA. After that, for all items which are erasable in the given tIntervals, their PID_lists are constructed. Here, the PID_lists of an item contains each PID_list related to given tInterval and they sorted ascendingly based on their TID values.

Next, to discover all erasable itemsets in the given tInterval, the VME algorithm is called with constructed PID_Lists of each items of TEI1. For algorithm steps of the VME, please refer to original paper (Deng & Xu, 2010). However, our VME implementation does not need to scan the TPDB since we send the necessary information obtaining from the ITPIL. Therefore, it directly joins PID_Lists of temporal erasable k-itemsets to get temporal erasable (k+1)-itemsets until all temporal erasable itemsets are found.

4. Experimental Results

In this section, the experimental analysis of ITEVME comparing with ITEMETA algorithm is given. Both of two algorithms were coded in C# programming language and all experiments run on in the same PC.

146

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Experimental analysis is conducted on the chess datasets which consists of 3196 transactions (or products). The chess includes 75 distinct items. This dataset was downloaded from http://sdrv.ms/14eshVm (Le & Vo, 2014). The make it like a TPDB, a column was added to indicate the time of the profit earned by each product. To generate time information of each product a randomly selected from the set of times which included four different time stapms such that (T1, T2, T3, T4).

The experiments were conducted under various maxT (%) values, including 5%, 10%, and 15%, 20%. For each maxT (%) value, both of ITEMETA and ITEVME algorithms with all possible time intervals were executed and the total running time is reported. Figure 4 gives the experimental result.

chess dataset with 4 different time stamps 120 ITEMETA ITEVME 100 112,58 80 60

40 21,808 20 2,358 1,63 runtime runtime seconds) (in 0,5 0,3 0 5% 10% 15% maxT (%)

Experimental results on chess dataset with 4 different time stamps show that ITEVME outperforms the ITEMETA in terms of runtime. It is also observed that, the gap between two algorithms gets larger when the maxT(%) values are increased. The reason is that, with the increasing maxT(%) values, the number of discovered temproral erasable itemset increases. Therefore, the performance of ITEMETA dramatically decreases since it needs more database scans when the maxT(%) value is increased. On the other hand, the ITEVME has a good scalability in terms of runtime.

5. Conclusion and Future Work

In this study, two algorithms named ITEMETA and ITEVME are introduced to handle interactive temporal erasable itemset mining. According to our knowledge, this is the first study that considers products with time information in the erasable itemset literature.

To compare the algorithm in term of running time, experiments were conducted on the chess database until various maximum profit thresholds. The experiment demonstrate that the ITEVME is more efficient than the ITEMETA. In future, we will study for more effective algorithms for interactive temporal erasable itemset mining. Moreover, mining top-rank-k TEIs and closed/maximal TEIs are seen as good topics will be studied for large temporal product databases.

147

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

References

1. Agrawal, R., & Srikant, R. (1994). Fast algorithms for mining association rules in large databases. In Proceeding of the 20th International Conference on Very Large Data Bases, 487–499. Morgan Kaufmann Publishers Inc.. 2. Deng, Z.-H, Fang G.-D, Wang Z.-H., & Xu X.-R. (2009). Mining erasable itemsets. 2009 International Conference on Machine Learning and Cybernetics. IEEE. 3. Deng, Z., & Xu, X. (2010). An Efficient Algorithm for Mining Erasable Itemsets. Advanced Data Mining and Applications (pp. 214–225). Springer Berlin Heidelberg. 4. Deng, Z.-H., & Xu, X.-R. (2012). Fast mining erasable itemsets using NC_sets. Expert Systems with Applications, 39(4), 4453–4463. Elsevier BV. 5. Le, T., & Vo, B. (2014). MEI: An efficient algorithm for mining erasable itemsets. Engineering Applications of Artificial Intelligence, 27, 155–166. Elsevier BV. 6. Le, T., Vo, B., & Coenen, F. (2013). An Efficient Algorithm for Mining Erasable Itemsets Using the Difference of NC-Sets. 2013 IEEE International Conference on Systems, Man, and Cybernetics. IEEE.

148

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Jute Yarn Consumption Prediction by Artificial Neural Network and Multilinear Regression Zeynep Didem Unutmaz Durmuşoğlu, Selma Gülyeşil

Department of Industrial Engineering, Gaziantep University, Turkey [email protected] [email protected]

Abstract

In today’s increasing competitive market conditions, the companies operating in the production and service sectors should meet the demand of the customers in a timely and completely manner. Therefore, all resources (raw materials, semi-finished products, energy sources, etc.) should be planned and supplied at the right time, at the right place and at sufficient quantity based on an accurate forecast of the demand. In the literature, there have been few studies about forecasting of raw material consumption in a production sector. In this study, ANN method was employed to predict the raw material consumption of a carpet production company. The relevant variables of the actual data belonging to 2015-2016 and 2017 were used. In addition, a multiple linear regression (MLR) model was also established to compare the performance of ANN method. The results show that ANN method produces more accurate forecasts when compared to MLR method.

Keywords: Artificial Neural Network (ANN), Multiple Linear Regression (MLR), Raw Material Consumption

1. Introduction

ANN method is one of the most widely used methods in recent years. Actually, in the literature, there are so many studies about ANN method from different science areas and also for distinct application studies. ANN method was used to solve MRP (Material Requirement Planning) problem of lot-sizing (Gaafar & Choueiki, 2000). Hwarng (2001) studied about ANN to understand better the modelling and forecasting ability of back propagation neural networks (BPNNs) on a special class of time series to improve performance. As a result of this study, when compared Box-Jenkins model BPNNs generally performed well (Hwarng, 2001). Zhang et. al. presented an article about a simulation study of artificial neural networks for nonlinear time series forecasting. In this study, they examined the effects of three factors which are input nodes, hidden nodes, and sample size. As a result of this study, the number of input nodes is much more important than the number of hidden nodes (Zhang, Patuwo, & Hu, 2001). In another study ANN method was used for ABC classification of stock keeping units (SKUs) in a pharmaceutical company and compared with the multiple discriminate analysis (MDA) technique. The results indicated that ANN models had higher predictive accuracy than MDA (Partovi & Anandarajan, 2002). In 2002, Daniel J. Fonseca and Daniel Navaresse studied about ANN method for job shop simulation. As a result of this study, ANN-based simulations were able to fairly capture the underlying relationship between jobs’ machine sequences and their resulting average flowtimes (Fonseca & Navaresse, 2002). G. Peter Zhang studied about a hybrid methodology that combines both ARIMA and ANN models in linear and nonlinear modeling. The results of this study showed that the proposed combined model can be an effective way to improve forecasting accuracy (Zhang, 2003). In another study, G. Peter Zhang and Min Qui studied about ANN method forecasting for seasonal and trend time series. They investigated the effectiveness of data preprocessing , that includes deseasonalization and detrending (Zhang & Qi, 2005). Another combined method of ANN was studied by Azadeh et. al. In this study, integrated genetic algorithm (GA) and ANN methods were used to estimate and predict electricity demand using stochactis procedures (Azadeh, Ghaderi, Tarverdian, & Saberi, 2007). Henry C. Co and Rujirek Boosarawongse studied about forecasting Thailand’s rice export by using ANN with exponential smoothing and ARIMA models. The results showed that ANN

149

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY performed relatively well (Co & Boosarawongse, 2007). In another study, ANN prediction method was used to optimize work-in-process inventory level for wafer fabrication. This study about finding an optimal work-in-process (WIP) value of wafer fabrication processes, which can be properly used to trigger the decision of when to release wafer lots, by integrating ANN and the sequential quadratic programming (SQP) method. The result of this study show that the offered integrated method is greatly improved finding the optimal WIP level (Lin, Shie, & Tsai, 2009).

The focus of this study is the accurate estimation of the consumption amount of the jute, which is one of the most basic raw materials in carpet production. The raw material of jute grows in India. This means carpet production companies supply jute from India by sea shipment. The fact that jute is provided in about 6 months by sea trade makes it important to estimate the amount of jute that will be needed. Accurate estimation of the amount of raw material consumption ensures that customer demands are met in a timely and complete manner, thus leading to competitive market conditions.

The most basic methods used in the estimation are statistical estimation methods. However, these statistical methods may be inadequate in view of increasing uncertainties and dynamic economic conditions. At this point, the method of ANN which have been successful and widespread in recent years, steps in.

In this study, two methods are used in order to forecast raw material consumption. One of the methods is ANN, which is the most popular and wide-ranging methods in recent years. The method of ANN has many advantages. The main advantage is that the ANN method can be used for nonlinear problems and most of the real world problems are nonlinear. Also, ANN method is applicable to different problems from distinct areas for both service and production sectors. In addition to this, a neural network has more fault tolerance when compared to traditional methods because of its parallel connections of neurons. The other method is one of the statistical methods of MLR. For ANN method, the Neural Network Toolbox of MATLAB software was used while, the Minitab program software was executed for MLR approach.

The required data have been taken from one of the carpet production company located in Gaziantep. Data belongs to the years 2015- 2016 and 2017.

Graphs and figures have been used to illustrate the results of two employed approaches. In order to compare the performance of two methods, MSE, R2 and adjusted R2 performance indicators have been used. And the results show that ANN values for performance indicators are better than MLR method.

2. Application

a. Data Used in The Study

In this study, the amount of the jute yarn used, which is one of the most basic raw materials used by a carpet production company, has been estimated for the future. For this purpose, 246 data points (actual values) were used for 11 different types of jute on a monthly basis.

For analysis of the raw material consumption of the future, 6 input variables were selected. These 6 input variables were selected because of their relation with the total jute consumption and decided with supply chain director of the company. These variables and their statements are as given in Table 1.

150

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Table 1 –Variables definition table

Variable Variable Name Statement Symbol

There are 11 types of jute to be used in different carpet models. X1 Jute Type The main differences between different types of jutes are the thickness and the number of yarn layers.

The purchase price of jute of different quality (different thickness and number of floors) is different. In addition, a jute type of X2 Purchasing Price the same quality can be obtained from different companies. In this case, the weighted average of purchase prices based on order quantity was used.

As the buying prices are in $ terms, the exchange rate input is used. X3 Exchange Rate This data is taken from the official web address of the Central Bank of the Republic of Turkey.

X4 Amount of Stock The monthly stock of each type of jute varies.

Different types of jute are used in the production of different carpet models. The production m2 of these carpets are kept separately in the X5 Production % ERP system used by the company. The ‘production %’ shows percentage of the stated jute in the total amount.

In different carpet models, even if the same type of jute is used, the amount of consumption can be different. While the average 2 X6 Average Consumption in 1 m consumption is taken in this data type, weighted usage is calculated on the basis of m2 produced.

Y Amount of Consumption on the Basis of Jute Type Dependent variable calculated according to 6 independent input variables

An example of data that has been used is given in Table 2.

151

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Table 2– Sample of the raw data used in application

Jute Type Purchasing Exchange Amount of Production % Average Purchasing Price ($) Rate Stock (tons) consumption Amount in 1 m2 (tons)

1 1.38 2.46 15.639 4.228 0.52313 26

2 1.18 2.46 32.225 15.123 0.80209 78

3 1.15 2.46 89.413 20.474 0.77556 104

4 0 2.46 41.217 0.529 0.75833 0

5 1.125 2.46 5.414 8.556 0.73667 52

Since the types of data used ($, ton, kg) are different, they must be subjected to normalization before being used in the software. In addition, normalization of the data, makes ANN method much more meaningful to be used in the solution of nonlinear problems. Min_Max normalization method was used to normalize the data. In this method, Min is the smallest value in a data type; Max represents the highest value in this data type. With the Min_Max method, the data is reduced to a range of 0 to 1. The formula used for normalization is shown in Eq.1:

푋푖−푋푚푖푛 푋′ = (Eq.1) 푋푚푎푥−푋푚푖푛

The data shown in Table 2 after the normalization process is as in Table 3.

Table 3 –Normalized values of the data shown in Table 2 Jute Type Purchasing Exchange Amount of Production Average Purchasing Price Rate Stock % Consumption Amount in 1 m2 0 0.951724 0 0.05145 0.11142 0.52313 0.1 0.1 0.813793 0 0.11588 0.39856 0.80209 0.3 0.2 0.793103 0 0.29416 0.53958 0.77556 0.4 0.3 0 0 0.13560 0.01394 0.75833 0 0.4 0.775862 0 0.01781 0.22549 0.73667 0.2

b. Forecasting with Artificial Neural Network Method

In this study, ANN method was used for the estimation of the amount of raw material for jute feed. Feed forward and back propagation multilayer ANN model is preferred for the solution of this problem. In a feed forward neural network, the direction is towards from input layer to output layer. Most widely used ANNs in forecasting problems are multilayer perceptron (MLPs) (Hamzacebi, 2008). MATLAB software was used for the training and testing of the networks created for this purpose.

152

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

As the first stage of the problem solving, data were divided as training, validation, and testing as in the following percentages. These percentages are default values of Neural Network Toolbox of MATLAB software:

• Training = 70 % • Validation = 15 % • Testing = 15 %

The number of neurons, activation function and training algorithm parameters in the hidden layer constituting the structure of the ANN affect the performance of the resulting network. For this reason, models with different network structures and learning parameters were established and experiments were made and the obtained results were compared. The most suitable network for the problem has been found by trial and error. As a result, the properties of most appropriate network are given in Table 4.

Table 4 –Properties of network structure

Network Type Feed forward – back propagation

Training Function TRAINLM

Adap. Learning Function LEARNGDM

Number of Hidden Layers 1

Number of Neurons 10

Transfer Function LOGSIG

Epoch 500

MSE and regression graphs that have been obtained with the most appropriate ANN are shown in Figure 1 and Figure 2, respectively:

Figure 1 – MSE graph of neural net

153

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 2– Regression graphs of neural net

MLR method was used in order to compare performance of the ANN method. The performance indicators to compare two methods are MSE, R2, and adjusted R2.

Adjusted R2 formulation is given in Equation 2. 푛−1 푅 − 푠푞(푎푑푗) = 1 − (1 − 푅2)( ) (Eq.2) 푛−푘−1

where; n = sample size k = number of independent variables R2 = coefficient of determination

All three performance indicators of ANN method is given in Table 5.

Table 5 –Performance indicators of the neural net Performance indicator MSE 0,00265 R2 0,9604 Adjusted R2 0,9845

154

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

c. Forecasting with Multiple Linear Regression Method

In forecasting with MLR approach, Minitab software is used. The same 246 data points were used for MLP study and performance indicators that are MSE, R2 and adjusted R2 were calculated in order to compare with ANN method.

In order to apply MLR method for any problem, the independent variables in the model should satisfy multicollinearity constraint. One of the methods to measure multicollinearity is known as the variance inflation factor (VIF). When multicollinearity increases, the VIF value increases.

In this study, 246 data points were executed in the Minitab software and obtained VIF results are given in Table 6.

Table 6 –VIF values of independent variables Variable Name VIF Value Variable Name VIF Value Purchasing price 1,26 Jute type 4 2,15 Rate 1,01 Jute type 5 2,22 Amount of stock 1,2 Jute type 6 2,51 Production % 3,43 Jute type 7 2,23 Avg. consumption in 1 1,47 Jute type 8 2,17 m2 carpet Jute type 1 1,3 Jute type 9 2,29 Jute type 2 2,44 Jute type 10 3,33 Jute type 3 2,62 Jute type 11 2,41

As it can be seen from Table 6, all of the VIF values for independent variables are less than 5. Therefore, it is concluded that there is no multicollinearity among all x variables and ensure necessary condition.

In addition to multicollinearity condition, significance of overall regression model should be checked to ensure whether the overall model is significant. The F-test is a method for testing whether the regression model explains a certain proportion of the variation in the y, dependent variable. The formulation of F-test is given in Equation 3:

푆푆푅 푘 퐹 = 푆푆퐸 (Eq. 3) 푛−푘−1

where; SSR = sum of squares regression SSE = sum of squares error n = sample size k = number of independent variables degrees of freedom = D1 = k, and D2 = (n-k-1)

The result of analysis of variance is given in Table 7.

155

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Table 7–Analysis of variance result for MLR Source DF Adj SS Adj MS F-value P-value Regression 15 6,16582 0,41105 53,89 0,000 Error 230 1,75451 0,00763 Total 245 7,92033

The null and alternative hypotheses are given below:

H0 = β1 = β2 = β3 = β4 = β5 = β6 = 0

HA = At least one βi ≠ 0

If the null hypothesis is true and all the slope coefficients of independent variables are equal to zero, the overall regression model is not appropriate for predictive applications.

Other necessary parameters to perform F-test are given as follows:

Criticial α = 0.05 k = 6 n – k – 1 = 239 degrees of freedom

F-table value is equal to approximately 2,14.

F = 53,89 > 2,14

Therefore, H0 is rejected and it can be concluded that the overall regression model explains a significant proportion of the variation.

As a result; MSE, R2 and adjusted R2 values for multiple linear regression method is as given in Table 8.

Table 8 – Performance indicators of MLR Performance indicator Regression results MSE 0,00713 R2 0,7785 Adjusted R2 0,7640

156

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 3 – Histogram plot of residuals

Figure 4 – Normal probability plot of residuals

Relative results of ANN and MLR models according to MSE, R2 and adjusted R2 is given in Table 9. As it can be seen from the table, the ANN model more successful than MLR method according to performance indicators’ values.

Table 9–Relative results of ANN and MLR methods Performance indicator Results of ANN model Results of MLR model MSE 0,00265 0,00713 R2 0,9604 0,7785 Adjusted R2 0,9845 0,7640

4. Conclusion

In this study, a real world business problem has been solved by using two methods, ANN and MLR. As a result, it is concluded that the results of ANN method are much closer to actual values when

157

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY compared MLR and therefore ANN seems to be more appropriate for dynamic real world conditions. In this study, the ANN method has been shown to be a reliable method in estimating jute consumption. In addition, by using this method customer demands can be estimated roughly and thereby they can be met in timely manner. In this respect customer satisfaction can be increased. As a future work, this method can be extended to include several macro level parameters like the interest rate. The proposed approach can be applied to other manufacturing sectors for forecasting the possible raw material consumption.

References 1. Azadeh, A., Ghaderi, S. F., Tarverdian, S., & Saberi, M. (2007). Integration of artificial neural networks and genetic algorithm to predict electrical energy consumption. Applied Mathematics and Computation, 186(2), 1731-1741. https://doi.org/10.1016/j.amc.2006.08.093

2.Co, H. C., & Boosarawongse, R. (2007). Forecasting Thailand’s rice export: Statistical techniques vs. artificial neural networks. Computers & Industrial Engineering, 53(4), 610-627. https://doi.org/10.1016/j.cie.2007.06.005

3.Fonseca, D. J., & Navaresse, D. (2002). Artificial neural networks for job shop simulation. Advanced Engineering Informatics, 16(4), 241-246. https://doi.org/10.1016/S1474-0346(03)00005-3

4.Gaafar, L. K., & Choueiki, M. H. (2000). A neural network model for solving the lot-sizing problem. Omega, 28(2), 175-184. https://doi.org/10.1016/S0305-0483(99)00035-3

5.Hamzacebi, C. (2008). Improving artificial neural networks’ performance in seasonal time series forecasting. Information Sciences, 178(23), 4550-4559. https://doi.org/10.1016/j.ins.2008.07.024

6.Hwarng, H. B. (2001). Insights into neural-network forecasting of time series corresponding to ARMA(p,q) structures. Omega, 29(3), 273-289. https://doi.org/10.1016/S0305-0483(01)00022-6

7.Lin, Y.-H., Shie, J.-R., & Tsai, C.-H. (2009). Using an artificial neural network prediction model to optimize work-in-process inventory level for wafer fabrication. Expert Systems with Applications, 36(2), 3421-3427. https://doi.org/10.1016/j.eswa.2008.02.009

8.Partovi, F. Y., & Anandarajan, M. (2002). Classifying inventory using an artificial neural network approach. Industrial Engineering, 16.

9.Zhang, G. P. (2003). Time series forecasting using a hybrid ARIMA and neural network model. Neurocomputing, 50, 159-175. https://doi.org/10.1016/S0925-2312(01)00702-0

10.Zhang, G. P., Patuwo, B. E., & Hu, M. Y. (2001). A simulation study of arti"cial neural networks for nonlinear time-series forecasting. Operations Research, 16.

11.Zhang, G. P., & Qi, M. (2005). Neural network forecasting for seasonal and trend time series. European Journal of Operational Research, 160(2), 501-514. https://doi.org/10.1016/j.ejor.2003.08.037

12.http://www.tcmb.gov.tr/

158

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Latest Trends in Textile-Based IoT Applications Duygu Erdem1, Münire Sibel Çetin2

1 Department of Fashion Design, Selcuk University, Turkey [email protected] 2Independent Researcher, İzmir, Turkey [email protected]

Abstract

Internet of Things (IoT) can be basically defined as the connection of any devices with each other or to the Internet. It can be used for remote sensing, data collection, control and processing, etc. There are many applications in this area from agriculture, energy and healthcare industry to architecture and textile industry.

In this study, textile based IoT applications, which have recently become a technological trend, have been examined and advantages and disadvantages of these applications are summarized. Then, the required aspects of textile based IoT applications are indicated.

Keywords: Internet of Things, IoT, wearables, smart textiles, textile sensors, textile antennas

1. Introduction

The “Internet of things” (IoT) is becoming an increasingly growing topic and touching almost every corner of the globe. Today, more devices have wi-fi capabilities, the Internet is more widely available, there is a wide range of sensors existing and technology costs are going down. Also, develop and usage of smartphones is rising rapidly. They are all establishing a basis for the IoT (Morgan, 2014).

IoT can be basically defined as the connection of any devices such as sensors, actuators and smartphones with each other or to the Internet (Morgan, 2014; Sethi and Sarangi, 2017; Irea et al., 2010). There are many application areas of the IoT from different industries. Aerospace and aviation industry automotive industry, telecommunication industry, medical and healthcare industry, independent living, pharmaceutical industry, retail, logistics and supply chain management, manufacturing industry, process industry, environment monitoring, transportation industry, agriculture and breeding, media and entertainment industry, ensuring industry and recycling can be given as examples of these areas. Some of them are still in the research phase while some of them are already used in our everyday life (Bandyopadhyay and Sen, 2011). Besides these areas, there is also an increasing demand for wearable technologies and smart textiles parallel to developments in the IoT applications. Smart textiles are now able to sense various signals from the environment such as temperature, heart rate, humidity, light intensity etc. and respond them as flexible devices integrated into clothes (Jeong et al., 2017).

In this paper, IoT applications in textiles were examined and outlined. Based on the examined studies, mostly used applications are defined and their advantages and challenges are presented.

2. Examples of IoT Applications in Textiles

Textile studies, especially studies about electronic textiles, are carried out for a range of applications including human healthcare in tandem with developing IoT applications (Moinudeen et al., 2017). The IoT systems basically comprise of objects, sensor devices, communication infrastructure, computational and processing unit (Khan et al., 2012). In connection with this structure, textile studies related to IoT technologies generally focus on integrating sensors and electronic equipment on garments.

159

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Some researchers use conventional components and garment together to generate a textile-based system while some others produce sensors, antennas etc. directly on garments. Textiles are great base materials because of their soft structures and using textile products for everyday life applications provide the users to feel comfortable (Moinudeen et al., 2017).

Textile based IoT applications generally use sensors and antennas and enable monitoring of human factors including health, wellness, behaviors and other data (Hiremath et al., 2014). Most of the studies examined in this paper are about textile antennas (Lee and Choi, 2017; Pavec et al., 2017; Agbor et al., 2018; Van Baelen et al., 2018; Huang et al., 2016; Sipilä et al., 2016; Shahariar et al., 2018). It is expected to textile antennas be an important part of the next generation wearable systems. However, they must be durable, washable, flexible, breathable and are needed to be upgraded in terms of using conditions, washing and heat cycling over time. The most challenging issue about producing antennas is incompatibility of mechanical properties of conductive materials and soft textile materials (Shahariar et al., 2018). Researchers are used conductive textile materials as yarns, fabrics, and inks to overcome this problem. In some researches, it is seen that textile antennas were directly sewn or embroidered on fabric/garment with conductive threads (Pavec et al., 2017; Huang et al., 2016). Screen printing method is another alternating production method for textile antennas. It is seen that flexible polymer, copper oxide ink, and silver conductor are used for screen printing onto different materials as 100% cotton material, nonwoven substrate and fireproof fabric (Pavec et al., 2017; Sipilä et al., 2016; Shahariar et al., 2018). One another possible solution to this problem is to use directly conductive fabrics. Conductive fabrics (felts) can be sewn as layers to form an antenna as Lee and Choi did in their study or commercial conductive fabrics as Cobaltex, ShieldT and Copper Taffeta can be attached by adhesives onto fabrics as Agbor et al. and Van Baelen et al. did in their study (Lee and Choi, 2017; Agbor et al., 2018; Van Baelen et al., 2018).

Textile based sensors are another focus of interest area for electronic textiles and wearables. There are several studies about development and production of tactile sensors, pH sensors, piezoresistive sensors, strain sensors, proprioceptive sensors and so on (Hasegawa et al., 2017; Coyle et al., 2009; Guo et al., 2011; Yuen et al., 2014). The aim of the sensor studies which is done in the past decades was to integrate electronic equipment into a textile and obtain softer and more comfortable products. With the progress in the internet of things, researchers started to address the topic as a whole and flipped the direction of research towards this direction. In spite of every textile-based sensor study could be integrated into IoT based systems with required addition, it is seen that some studies are directly focused on IoT based system. For instance, Navratil et al. have developed temperature sensors with printing the carbon nanotubes ink on 100% fluorocarbon, 100% nylon fiber and narrow polyimide stripe, and Hallfors et al. have developed fabric electrocardiogram (ECG) sensors made from Nylon coated with reduced graphene oxide (Navratil et al., 2018; Hallfors et al., 2018).

Another issue where textile based IoT studies are predominantly concentrated on is monitoring. Trending topics in the field of monitoring are measuring of body temperature, heart rate, breathing rate, ECG and oxygen saturation, fall detection, body position and motion detection and gesture recognition (Ferreira et al., 2016; Wu et al., 2017; Yang et al., 2016; Doukas and Maglogiannis, 2012; Abtahi et atl., 2018; Poupyrev et al., 2016; Kokalki et al., 2017). . It is seen that gathered data from the wearable IoT systems are transferred to a smartphone application or Cloud infrastructures and saved there. The most commonly used wireless communication technologies are remarked as Bluetooth, Zigbee and 3G. As mentioned before, it is observed from the examined studies that some system is produced from fully textile materials when the others piece together the conventional electronic components and garments. Textiles are especially preferred for such applications because they are flexible, stretchable and comfortable.

160

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

3. Results and Discussion

It is desired from wearable technologies to provide features such as being flexible, durable, configurable, graceful, holistic, interactive, jazzy, kinesthetic, lightweight, multifunctional and being suitable for everyday use. Textile materials meet all the required attributes of wearables.

In addition to these attributes, textiles have other positive characteristics which will be described below: ● Clothes are the most universal interfaces for wearable IoT technologies. Because dress up is the most basic need of people. Furthermore, fabrics are used in many areas of our everyday life such as furniture, cars, carpets and so on. ● Textiles are flexible. Therefore, they can be transformed into any desired shape with bending up and they can return to their original shape without any structural damage. ● Like a second skin, textiles can be produced to fit upon each body shape and each climatic condition. For instance, they can be produced from different and special materials to protect the wearer from cold in winter or from chemicals in some working environments, etc. ● By combining pieces of fabrics with different characteristics and sizes, various surfaces can be formed for integration of sensors, processors and other electronic equipment. ● Textiles can be manufactured using a variety of threads, fabrics and production techniques (e.g. weaving, knitting, nonwovens, and printing) since therefore the required flexibility is preserved even after the integration of electronic equipment. Besides, they can be manufactured in a relatively cost-effective manner compared to traditional printed circuit boards. ● The data buses or communication pathways can be produced from directly textile materials or integrated into the fabric and by this way the problems associated with entanglement and snags when using the textile based IoT systems are eliminated.

In addition to these advantages, there are also difficulties in the use of textiles in IoT applications. These disadvantages can be summarized as incompatibility of mechanical properties of conductive materials and soft textile materials, prone to damage during washing, and heat cycling over time and encountered inability to establish reliable connections between the components placed at different locations on the textile.

4. Conclusion

In this study, various textile based IoT applications in the literature have been examined. Their advantages and challenges are summarized. As can be seen from previous studies; textiles are very effective materials for IoT applications. By reason of there are a variety of materials and production techniques for producing of textiles and they can be manufactured in many different forms in terms of size, shape, etc.

There are some usage challenges such as durability, incompatibility of material properties (conductive and soft materials) and inability to establish reliable connections. However, due to the high number of advantages of textile based IoT applications, studies should be increased to eliminate these difficulties.

In this way, by integrating IoT technologies into the products made from textile materials in every area of our lives (bed, furniture, carpets of cars, workplaces and gyms, clothes we wear and etc.); it will be possible to reach a broader data bandwidth and gather all kinds of data about all the activities from the moment we wake up to the time we sleep.

161

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

References Morgan, J. (2014). A simple explanation of' The Internet of Things'. Retrieved November, 20, 2015. Sethi, P., & Sarangi, S. R. (2017). Internet of things: architectures, protocols, and applications. Journal of Electrical and Computer Engineering, 2017. Iera, A., Floerkemeier, C., Mitsugi, J., & Morabito, G. (2010). The internet of things. IEEE Wireless Communications, 17(6), 8-9. Bandyopadhyay, D., & Sen, J. (2011). Internet of things: Applications and challenges in technology and standardization. Wireless Personal Communications, 58(1), 49-69. Jeong, S. M., Song, S., Seo, H. J., Choi, W. M., Hwang, S. H., Lee, S. G., & Lim, S. K. (2017). Battery‐Free, Human‐Motion‐Powered Light‐Emitting Fabric: Mechanoluminescent Textile. Advanced Sustainable Systems, 1(12), 1700126. Moinudeen, G. K., Ahmad, F., Kumar, D., Al-Douri, Y., & Ahmad, S. (2017). IoT Applications in Future Foreseen Guided by Engineered Nanomaterials and Printed Intelligence Technologies a Technology Review. International Journal of Internet of Things, 6(3), 106-148. Khan, R., Khan, S. U., Zaheer, R., & Khan, S. (2012, December). Future internet: the internet of things architecture, possible applications and key challenges. In Frontiers of Information Technology (FIT), 2012 10th International Conference on (pp. 257-260). IEEE. Hiremath, S., Yang, G., & Mankodiya, K. (2014, November). Wearable Internet of Things: Concept, architectural components and promises for person-centered healthcare. In Wireless Mobile Communication and Healthcare (Mobihealth), 2014 EAI 4th International Conference on (pp. 304-307). IEEE. Lee, H., & Choi, J. (2017, July). A compact all-textile on-body SIW antenna for IoT applications. In Antennas and Propagation & USNC/URSI National Radio Science Meeting, 2017 IEEE International Symposium on (pp. 825-826). IEEE. Pavec, M., Bystričky, T., Moravcova, D., Reboun, J., Soukup, R., Navratil, J., & Hamacek, A. (2017, May). A comparison of embroidered and screen-printed Ultra-Wideband antennas. In Electronics Technology (ISSE), 2017 40th International Spring Seminar on (pp. 1-4). IEEE. Agbor, I., Biswas, D. K., & Mahbub, I. (2018, April). A comprehensive analysis of various electro-textile materials for wearable antenna applications. In 2018 Texas Symposium on Wireless and Microwave Circuits and Systems (WMCS) (pp. 1-4). IEEE. Van Baelen, D., Lemey, S., Verhaevert, J., & Rogier, H. (2018). A Novel Manufacturing Process for Compact, Low-Weight and Flexible Ultra-Wideband Cavity Backed Textile Antennas. Materials, 11(1), 67. Huang, J. S., Jiang, T. Y., Wang, Z. X., Wu, S. W., & Chen, Y. S. (2016). A novel textile antenna using composite multifilament conductive threads for smart clothing applications. Microwave and Optical Technology Letters, 58(5), 1232-1236. Sipilä, E., Liu, J., Wang, J., Virkki, J., Björninen, T., Cheng, L., ... & Ukkonen, L. (2016, April). Additive manufacturing of antennas from copper oxide nanoparticle ink: Toward low-cost RFID tags on paper-and textile-based platforms. In Antennas and Propagation (EuCAP), 2016 10th European Conference on (pp. 1- 4). IEEE. Shahariar, H., Soewardiman, H., Muchler, C. A., Adams, J. J., & Jur, J. S. (2018). Porous textile antenna designs for improved wearability. Smart Materials and Structures, 27(4), 045008. Hasegawa, Y., Shikida, M., Ogura, D., & Sato, K. (2007, June). Glove type of wearable tactile sensor produced by artificial hollow fiber. In Solid-State Sensors, Actuators and Microsystems Conference, 2007. TRANSDUCERS 2007. International (pp. 1453-1456). IEEE. Coyle, S., Morris, D., Lau, K. T., Diamond, D., & Moyna, N. (2009, June). Textile-based wearable sensors for assisting sports performance. In Wearable and Implantable Body Sensor Networks, 2009. BSN 2009. Sixth International Workshop on (pp. 307-311). IEEE. Guo, L., Peterson, J., Qureshi, W., Kalantar Mehrjerdi, A., Skrifvars, M., & Berglin, L. (2011). Knitted wearable stretch sensor for breathing monitoring application. In Ambience'11, Borås, Sweden, 2011. Yuen, A. C., Bakir, A. A., Rajdi, N. N. Z. M., Lam, C. L., Saleh, S. M., & Wicaksono, D. H. (2014). Proprioceptive sensing system for therapy assessment using cotton fabric-based biomedical microelectromechanical system. IEEE Sensors Journal, 14(8), 2872-2880.

162

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Navratil, J., Rericha, T., Soukup, R., & Hamacek, A. (2018, May). Aerosol Jet Printed Sensor on Fibre for Smart and IoT Applications. In 2018 41st International Spring Seminar on Electronics Technology (ISSE) (pp. 1-4). IEEE. Hallfors, N. G., Alhawari, M., Jaoude, M. A., Kifle, Y., Saleh, H., Liao, K., ... & Isakovic, A. F. (2018). Graphene oxide: Nylon ECG sensors for wearable IoT healthcare—nanomaterial and SoC interface. Analog Integrated Circuits and Signal Processing, 1-8. Ferreira, A. G., Fernandes, D., Branco, S., Monteiro, J. L., Cabral, J., Catarino, A. P., & Rocha, A. M. (2016, March). A smart wearable system for sudden infant death syndrome monitoring. In Industrial Technology (ICIT), 2016 IEEE International Conference on (pp. 1920-1925). IEEE. Wu, T., Wu, F., Redoute, J. M., & Yuce, M. R. (2017). An autonomous wireless body area network implementation towards IoT connected healthcare applications. IEEE Access, 5, 11413-11422. Yang, Z., Zhou, Q., Lei, L., Zheng, K., & Xiang, W. (2016). An IoT-cloud based wearable ECG monitoring system for smart healthcare. Journal of medical systems, 40(12), 286. Doukas, C., & Maglogiannis, I. (2012, July). Bringing IoT and cloud computing towards pervasive healthcare. In Innovative Mobile and Internet Services in Ubiquitous Computing (IMIS), 2012 Sixth International Conference on (pp. 922-926). IEEE. Abtahi, M., Gyllinsky, J. V., Paesang, B., Barlow, S., Constant, M., Gomes, N., ... & Mankodiya, K. (2018). MagicSox: An E-textile IoT system to quantify gait abnormalities. Smart Health, 5, 4-14. Poupyrev, I., Gong, N. W., Fukuhara, S., Karagozler, M. E., Schwesig, C., & Robinson, K. E. (2016, May). Project Jacquard: interactive digital textiles at scale. In Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems (pp. 4216-4227). ACM. Kokalki, S. A., Mali, A. R., Mundada, P. A., & Sontakke, R. H. (2017, September). Smart health band using IoT. In 2017 IEEE International Conference on Power, Control, Signals and Instrumentation Engineering (ICPCSI) (pp. 1683-1687). IEEE.

163

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Localization and Point Cloud Based 3D Mapping With Autonomous Robots Selya Açikel1, Ahmet Gökçen2

1 Institute of Engineering and Science, Iskenderun Technical University, Turkey [email protected] 2 Department of Computer Engineering, Iskenderun Technical University, Turkey [email protected]

Abstract

Emerging technology is keeping its pace with the ever-increasing needs of people and expanding to different domains. Robotics, which is one of the domain of Artificial Intelligence (AI), has become the indispensable part of technology. It extends services such as saving human power, minimizing errors in production, working in microscopic or gigantic dimensions ultimately saving time and cost. The most important field of Robotics is Intelligent Autonomous Robots (IAR) which has the ability to detect and make decisions. Space science, industry, health, mining and many other sectors benefited from autonomous robots.

In this study, three-dimensional modelling of an environment was made and the location of the intelligent autonomous robot was monitored. A laser sensor, Lidar Lite V3, was used for modeling. Ultrasonic distance sensors have been used to ensure that the vehicle travels without hitting obstacles. The data obtained from the autonomous vehicle and the environment modeling and vehicle location tracking were performed in a three-dimensional virtual environment designed using OpenGL Library. The Moving Average Filter have been used to remove noise from measurement errors. During the mapping, motion tracking was performed by using OpenCV library to detect moving objects in the environment. Intelligent autonomous robot moved to three different points of the environment and obtained data from different perspectives and these data were combined with algorithms and three- dimensional modeling of the environment was performed.

Keyword(s): Mapping, localization, SLAM, Moving Average Filter, Lidar

1. Introduction

With the rise and emerging of technologies, autonomous robots have become indispensable in human life. As in many areas, autonomous robots are used for mapping. In an environment where an autonomous robot does not have knowledge, the problem of mapping the environment and finding its own position is referred to in the literature as Simultaneous Localization and Mapping (SLAM) (Dissanayake, Newman, Durrant-Whyte, Clark & Csorba, 2000; Altuntaş, Uslu, Çakmak, Amasyalı, & Yavuz, 2017). The SLAM problem was first mentioned in 1986 at the IEEE Robotics and Automation Conference (Durrant-Whyte & Bailey, 2006). The most important material used to solve the SLAM problem is the scanners, cameras or distance sensors used to scan the environment. Said material directly affects the amount of noise in the generated map. Noise is a factor that reduces the ability of the system to predict. In order for the system to produce the results that are closest to reality, images should be cleaned as much as possible from noise. Since the 1990s, probabilistic methods have been used to eliminate noise (Thrun, 2002). Some commonly used probabilistic filters are Kalman Filter, Particle Filter and Popular Expectation Maximization. Most of the three-dimensional mapping with distance sensors uses point cloud topology. The mapping with point cloud is realized by combining numerous points and forming meaningful shapes. Rusu, R. et al. In 2008, based on the densities of the clustered points, they obtained successful results using the MLESAC (Maximum Likelihood Estimation Sample

164

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Consensus) algorithm for identifying objects from point clouds (Rusu, Marton, Blodow, Dolha, Beetz, 2008).

Sensors that make point measurements are used to create point clouds. The ultrasonic sensors, which were created by the French physicist Paul Langevin and measuring the distance with sound waves, were first introduced in 1917 (Graff, 1981). Infrared sensors measure distance based on the intensity of the infrared light they emit into their field of view. Today, devices that produce the most accurate results are laser distance sensors that do not require over-calculation. Lidar sensors, which are a kind of laser distance sensors and operate with radar logic, are high-range and low-error sensors (Fowler, 2000).

The autonomous robot needs to record his movements step by step in order to map his environment and also to determine his own position. It should be able to make certain calculations in each new movement and have information about the current moment. Various materials can be used to determine the position of the robot. GPS sensors, motors with encoders, GYRO sensors or compass sensors are some of the materials that can be used for motion and location tracking. GPS modules are not preferred for indoor mapping because they are inefficient in closed environments.

There are many studies in the literature on mapping with autonomous robots. Kurt Z. et al. used an infrared sensor as a distance sensor and suggested the Sequential Monte Carlo method, one of the particle filter applications, to reduce noise in the image. He successfully applied one of the statistical estimation methods to SLAM problems (Kurt, 2007).

In 2010, Ankışhan H. and Efe M. compared the results of simultaneous positioning and mapping of the square root odorless and square root difference from the Kalman Filters to those obtained from the Extended Kalman Filter and produced an alternative solution by saving the margin of error and time spent (Ankışhan & Efe, 2010).

One of the problems encountered in SLAM problems is the motion estimation. The movements, coordinates and position of the autonomous robot reduce the calculation quality. In 2015, Carlone L. et al. suggested Exposure Graph Optimization based on nonlinear estimation methods such as Gauss and Newton in their estimation of rotation. They have succeeded in minimizing the errors caused by the movement of the autonomous vehicle (Carlone, Tron, Daniilidis, & Dellaert, 2015).

In 2018, Lee, K. et al. they have worked on a pole star model and have made a study that the direction finding problem will be solved in the case of an object or structure that is to be noticed from everywhere such as pole star in the environment tobe mapped. They proposed the Direction Landmark- based SLAM (DLSLAM) algorithm to detect the pole star model they called Direction Landmark (DL) and determine the direction accordingly (Lee, Ryu, Nam & Doh, 2018).

Hosseinzadeh M. et al. they made a study that transforms the point sets of the environment mapped with an autonomous robot into meaningful objects with deep learning methods. By recognizing the objects previously taught, the robot provided more information about the environment. (Hosseinzadeh, Latif & Reid, 2018).

In this study, a solution was made to the SLAM problem in closed environments by measuring the distance with the Lidar Lite V3 integrated on the autonomous robot. The data has been received from the Lidar sensor is shown in 3D in the OpenGL supported desktop application with Point Cloud topology through the Moving Averages Filter. Unlike the literature, the errors that may occur in rotation and other movements have been resolved by means of processing some data received from a digital compass sensor and encoders integrated into the wheels of the autonomous vehicle. The GPS module on the autonomous robot is used to show the real coordinates of the environment to be mapped. The data obtained from the symmetrical three points of the environment are combined to display all the surfaces

165

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY of the objects in the environment. Co-ordination of all processes is provided by Raspberry Pi 3. In order to reduce the errors in measurement by detecting the movement of the objects measured in the distance during measurement, motion detection was done with OpenCV supported image processing algorithms on the images taken from the camera placed on Raspberry.

2. Materials and Methods

2.1 Lidar Lite V3

The Lidar Lite V3 is a laser distance sensor with a wavelength of 905nm. It is capable of measuring a maximum of 270 times per second and can measure up to 40 meters. It has a margin of error of ± 2.5 centimeters over distances longer than one meter (Maulana, Rusdinar, & Priramadhi, 2018). The Lidar sensors make the distance calculation based on the reflection time of the laser beam. Equation 1 shows how to calculate the distance. The value indicated by c represents the speed of the laser beam, the value indicated by t represents the reflection time of the light, and the value indicated by d represents the calculated distance.

푐×푡 푑 = (Eq. 1) 2

2.2 Moving Average Filter

The Moving Average Filter is a filter that takes a continuous average of a set of values. It is used to improve sensor data, to reduce sharpness in pictures or to remove noise. Moving Average Filter, a linear filter, is used in this study to minimize errors in a set of measurement data from the lidar sensor. In Equation 2, the basic equation of Moving Average Filter is shown.

1 푥(푘) = ∑푁−1 푥(푘 − 1) (Eq. 2) 푁 푖=0

The N value specified in Equation 2 indicates the number of elements of the sequence to be processed, ie the number of iterations to which the filter will be applied. 푥(푘) represents the final value produced by the system (Golestan, Ramezani, Guerrero, Freijedo, & Monfared, 2014).

2.3 Calculation of Coordinates

In order to draw a point in three-dimensional space in a virtual environment, the coordinates of the point are required. The coordinates of the point are calculated with three parameters. These: • The shortest distance to the origin, • The angle made in the horizontal axis. • The angle made in the vertical axis (Üstün, 1996).

Figure 1 shows the parameters required to display a point in the three-dimensional coordinate system. The origin represents the point at which the autonomous robot is located. Angle parameters on the horizontal and vertical axis are taken from the servo motors which provide the movement of the lidar sensor.

166

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 1 – Showing a Point in a Three Dimensional Coordinate System

In the Equation 3, it is shown that the coordinate calculations made with the parameters obtained from the lidar sensor and servo motors are shown.

푥 푐표푠휑 × 푐표푠휃 [푦] = 푑 × [푐표푠휑 × 푠푖푛휃] (Eq. 3) 푧 푠푖푛휑

The movements of the autonomous robot affect the condition of the stage. In this case, every action must be reflected on the stage. The projection of the robot movements to the scene was carried out with transformation matrices. The movements of the robot in the x or y axes are reflected to the scene using the translational matrix. Equation 4 shows the translation matrix in the x and y axes (Siciliano, Sciavicco, Villani & Oriolo, 2010).

1 0 0 푡푥 0 1 0 푡 푇 = [ 푦] (Eq. 4) 푥,푦 0 0 1 1 0 0 0 1

In addition to the translation process, the rotation of the autonomous robot in the environment must be reflected on the stage. The rotation movements also affect the coordinates of the point. Since the origin of the scene is the point where the autonomous robot is located, the rotation is applied to the origin. The rotational movement of the autonomous robot takes place in the z-axis. The rotation matrix is shown in Equation 5. The angle θ, is the difference between the last angle of the robot and the first angle (Siciliano et al., 2010). A three-axis digital compass sensor is used to obtain the angle difference.

푐표푠휃 − 푠푖푛휃 0 푅푧(휃) = [푠푖푛휃 푐표푠휃 0] (Eq. 5) 0 0 1

2.4 Motion Detection

Motion detection can be defined as the detection of differences between the initial state of the system and its current state. If the first image of an environment is determined as a background, then any changes in the environment can be easily detected. Motion detection is made by the difference between the image set as a background and the new frames (Bradski & Kaehler, 2008). When the pixel set changes are detected, the pixels in which changes occur can be labeled and movements in the environment can be detected.

167

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

In this study, motion detection was made by using pixel group difference method between images using OpenCV Library. The purpose of motion detection is to minimize the noise and error in the environment map by ignoring the dynamic objects inside the environment to be mapped.

3. Results and Discussion

Findings from the measurements show that the Moving Average Filter is highly effective on sensor data. For each 1 degree movement of the servo motors, ten distance measurements were made and the measurements were added to the sequence of the Moving Averages Filter respectively. The results are shown in Table 1. Table 1 – Moving Average Filter Effect Iteration Sensor Data Moving Average Filter 1 149,3 149,3 2 150,3 149,657 3 152,6 150,708 4 148,1 149,777

5 151,2 150,285 6 150,5 150,362

7 149,2 149,947 8 153,1 151,073 9 149,9 150,654 10 149,8 150,349

Table 1 shows the sensor measurements from a point of 150 centimeters and the effect of the Moving Averages Filter on these measurements. A 0.2 centimeter error of the final value of the sensor measurements does not mean that the result is healthy in every measurement. The measured distance can be any of the ten measurements in Table 1. Therefore, the data is filtered. The value used for the coordinate account is the last value obtained from the filter. The effect of the Moving Averages Filter on the sensor data is shown in Graph 1.

Graph 1 – Moving Average Filter Effect MOVING AVERAGE FILTER

Sensor DataEFFECTMoving Average Filter 154 152 150

CM 148 146 144 1 2 3 4 5 6 7 8 9 10 ITERATION

In Graph 1, it is concluded that the amount of sensor data is linearized by Moving Average Filter. After the Moving Average Filter has been applied to all the data, a significant decrease has been observed in the noise. Figure 2 shows the results obtained from the measurements by applying the Moving Averages Filter.

168

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 2 – Applying the Moving Averages Filter to All Data

Figure 2 shows that the Moving Averages Filter removes noise from measurement errors and clarifies objects in the map.

It is important not to create meaningless images when creating a map. In static environments, meaningless images resulting from incorrect measurements are removed with the help of filters. In dynamic environments, only filtering is insufficient. Meaningless lines or objects can be created on the map, although the sensor data is measured correctly. The motion detection algorithm developed by using the OpenCV Library has been added to the autonomous robot in order to avoid nonsense images caused by movements in dynamic environments. If the robot detects a moving object while it is taking its measurement from the environment it is mapped in, it is allowed to pause the measurement process until the environment is restored. Figure 3 shows the detection of moving objects in the environment.

Figure 3 – Detecting Moving Objects

The movement of autonomous robot is monitored by dc motors with encoder and digital compass sensor. The motion in the environment was transferred to the scene with data from the encoder motors. The rotational movements have been realized by reflecting the angle values obtained from the digital compass sensor to the scene. Figure 4 shows the mapping of a room and the movement of the autonomous robot in the environment.

169

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 4 – Three-dimensional Mapping of Environment

Figure 4 shows the images taken from different perspectives. The red lines are formed by the autonomous robot's own position in the environment as a result of the movements.

4. Conclusion

In Epilogue to this study, a solution has been developed to prevent the formation of meaningless images in three-dimensional SLAM problems with dynamic motion detection. The success of the Moving Average Filter has been shown to eliminate sensor errors. By taking measurements from three different points of an environment, the meaninglessness of the images in different perspectives was prevented. The movements of the autonomous robot have been accurately analysed to prevent image shifts during the measurements.

References 1. Altuntaş, N., Uslu, E., Çakmak, F., Amasyalı, M. F., & Yavuz, S. (2017, October). Comparison of 3- dimensional SLAM systems: RTAB-Map vs. Kintinuous. In Computer Science and Engineering (UBMK), 2017 International Conference on (pp. 99-103). IEEE. 2. Ankışhan, H., & Efe, M. (2010). Eşzamanlı konum belirleme ve harita oluşturmaya Kalman filtre yaklaşımları. DÜMF Mühendislik Dergisi, 1(1), 13-20. 3. Bradski, G., & Kaehler, A. (2008). Learning OpenCV: Computer vision with the OpenCV library. " O'Reilly Media, Inc.",(pp. 265-271) 4. Carlone, L., Tron, R., Daniilidis, K., & Dellaert, F. (2015, May). Initialization techniques for 3D SLAM: a survey on rotation estimation and its use in pose graph optimization. In Robotics and Automation (ICRA), 2015 IEEE International Conference on (pp. 4597-4604). IEEE. 5. Dissanayake, M. G., Newman, P., Durrant-Whyte, H. F., Clark, S., & Csorba, M. (2000). An experimental and theoretical investigation into simultaneous localisation and map building. In Experimental robotics VI (pp. 265-274). Springer, London. 6. Durrant-Whyte, H., & Bailey, T. (2006). Simultaneous localization and mapping: part I. IEEE robotics & automation magazine, 13(2), 99-110. 7. Fowler, R. A. (2000). The lowdown on LIDAR. Earth Observation Magazine, 9(3), 5. 8. Golestan, S., Ramezani, M., Guerrero, J. M., Freijedo, F. D., & Monfared, M. (2014). Moving average filter based phase-locked loops: Performance analysis and design guidelines. IEEE Transactions on Power Electronics, 29(6), 2750-2763. 9. Graff, K. F. (1981). A history of ultrasonics. In Physical acoustics (Vol. 15, pp. 1-97). Academic Press. 10. Hosseinzadeh, M., Li, K., Latif, Y., & Reid, I. (2018). Real-Time Monocular Object-Model Aware Sparse SLAM. arXiv preprint arXiv:1809.09149. 11. Kurt, Z. (2007). Eş zamanlı konum belirleme ve haritalamaya yönelik akıllı algoritmaların geliştirilmesi (Doctoral dissertation, YTÜ Fen Bilimleri Enstitüsü).

170

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

12. Lee, K., Ryu, S. H., Nam, C., & Doh, N. L. (2018). A practical 2D/3D SLAM using directional patterns of an indoor structure. Intelligent Service Robotics, 11(1), 1-24. 13. Maulana, I., Rusdinar, A., & Priramadhi, R. A. (2018). Aplikasi Lidar untuk Pemetaan dan Navigasi pada Lingkungan Tertutup. eProceedings of Engineering, 5(1). 14. Rusu, R. B., Marton, Z. C., Blodow, N., Dolha, M., & Beetz, M. (2008). Towards 3D point cloud based object maps for household environments. Robotics and Autonomous Systems, 56(11), 927-941. 15. Siciliano, B., Sciavicco, L., Villani, L., & Oriolo, G. (2010). Robotics: modelling, planning and control. Springer Science & Business Media. 16. Thrun, S. (2002). Robotic mapping: A survey. Exploring artificial intelligence in the new millennium, 1(1-35), 1. 17. Üstün, A. (1996). Datum dönüşümleri. Yüksek Lisans Tezi, Yıldız Teknik Üniversitesi Fen Bilimleri Enstitüsü, İstanbul

171

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

NoSQL Database Systems: Review and Comparision Kerim Melih Çimrin1, Yaşar Daşdemir2

Iskenderun Technical University, Engineering and Natural Sciences Faculty, Computer Engineering, Hatay/Turkey [email protected] [email protected]

Abstract Today, the use of the Internet consists of many different purposes. These are business, commercial and social media applications. As a result of all this, a large amount of data is generated. Big data; social media shares, network logs, photos, video, log files, such as all the data recovered from different sources, is converted to a meaningful and processed form. The fact that large data can be processed and made meaningful is particularly important for many organizations, such as companies. Large data consists of structural, semi-structural and non-structural data. The non-structural data are difficult to process in RDBMS. NoSQL (not only SQL system) systems have been developed as it is more efficient and logical to process non-structural data. NoSQL systems work on both non-structural and semi-structural data. In this study, we aimed of comparison to NoSQL systems. It is much more efficient and effective to explain how graph database and data are linked to other data. Column-oriented is used when more reporting and modification are more significant. A key-value store database is used with a simple and limited set of processes. The most important feature of the document database is that they are flexible and the data corresponding to a key are stored in objects called documents.

Keywords: Graph database, Column-oriented database, RDBMS, NoSQL, Key-value database, Document database, Cypher, CQL, API

1.Introduction

Big data is a term that describes the large volume of data both structured and unstructured that business on a day-to-day basis. The fact that large data can be processed and made meaningful is particularly important for many organizations, such as companies. Structured Query Language (SQL), which is a computer language for storing, manipulating and retrieving data is widely used in a relational database. It is not possible to process large data (unstructured) with meaningful relational databases. In other words, RDBMS systems are not suitable for unstructured data. Therefore, the development of NoSQL systems was needed. NoSQL systems have been created for different purposes from SQL systems. In fact, there are advantages and disadvantages of both solutions [1]. NoSQL systems can be divided into four categories. These can be listed as key-value storage, document storage, column- oriented storage and graph storage respectively. The purpose of this study is to compare the NoSQL systems to the users and to show which system is more suitable for what purpose.

2.Types of NoSQL

In this section, we briefly summarize types of NoSQL. NoSQL systems have four different categories. These NoSQL systems are key-value (e.g., RIAK), document (e.g., MongoDB) column- oriented (e.g., Cassandra) and graph databases (e.g., Neo4j). In the following sections, we present these four major types.

172

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

3.Key-Value Store Databases The key-value data stores are very efficient and flexible model. The key-value has a API (application programming interface) [2]. At the same time key-value stores has very important role for many different websites like Memcached at Facebook, Twitter Voldemort at LinkedIn. Values are identified and accessed by means of a key. Stored values can be numbers, strings, counters, JSON, XML, HTML, binaries, images, and short videos. The key-value store is very flexible model, because the application has complete control over in the value. At the same time, key-value store has very high performance and scalability. Firstly, key-value store databases works very simple. A key-value database stores a data record using one primary key. On the other hand, key-value store is often used for caching data [3]. Secondly, key-value store has also high-level scalability. Key-value store functions works very simplicity for that reason key value store databases scalability level high rate. Some areas of use; user suggesting advertising according to shopping preferences, caching for applications and user preference and profile information storage.

Figure 1 Key-Value Store (www.intel.com) 4.Document Store Databases

Document store databases are different works from key-value store databases. Basically document store databases means that store the data in the form of documents [4]. Document store databases assigns a key-value to each document. Documents are encoded in a standard data exchange format, such as a JavaScript Option Notation or binary json[1]. Document store databases are approved as a powerful, flexible and nimble tool to store Big Data. Document stores the values are not ambiguous to the system and can query also [4]. All in all, document store databases performance and flexibility level is high but confusion is low level. When it comes to scalability and functionality, it is variable. Some areas of use; sensor data from devices and logging of applications.

173

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 2. Document Store (www.lennilobel.wordpress.com)

Table 2 – The Technical comparison about the leader NoSQL products (Simp: Simplication, Lic: License, Lang: Language) [5] Simp Search Originat Storage Protocol Model Lic Lang e MongoD Yes Yes - Memory BSON Document AGPL C++ B mapped b- trees CouchDB No No Dynamo COW- HTTP/RE Document Apache Erlang BTree ST Riak Yes Yes Dynamo Pluggable: HTTP/RE Key/Valu Apache Erlang InnoDB, ST or e LevelDB, TCP/Prot Bitcask o Redis No No - In memory, TCP Key/Valu BSD C++ snapshot to e disk Voldemor No No Dynamo Pluggable: - Key/Valu Apache Java t BSV, e MySQL, in Cassandra Yes Yes BigTabl Memtable/ TCP/Thrif Wide Apache Java e, SStabl e t Column Dynamo HBase Yes Yes BigTabl HDFS HTTP/RE Wide Apache Java e ST or Column

5.Column-oriented Store Databases

Column-oriented store databases is the type most similar to relational database model. Although column databases and column extensions share the concept of column-column storage in row-based databases, column stores do not store data in tables but the data is stored in highly distributed architectures. In column stores, each key is associated with one or more attributes, and a column store stores data quickly with less I / O activity [2]. Inshort, column-oriented databases are based on hybrid

174

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY approach and it is based on the declaration features and various key-value stores schema and relational databases[6]. Some areas of use; e-mail operations and short message service operations.

Figure 3 Column-oriented Store (www.researchgate.net) 6.Graph Store Databases Basically, a graphical database is a database designed to treat relationships between data as equally important and graph databases are suitable to store. It is intended to keep data without collapsing to a predefined model. Instead, the data is stored in the graph we draw first, and shows how each node is interconnected or associated with others [7]. Graphical databases also store, store, and display data in graphical form. In the graphical database, the nodes consist of objects and edges. Nodes and edges are like the relationship between objects. An index-free contiguous technique is used, which means that each node consists of a direct pointer to another node. Some areas of use; social media applications, route and finding the closest route, fraud detection operations and investment management and finance area.

Figure 4 Graph Store (www.neo4j.com)

175

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Table 2 – NoSQL DATABASE TYPES[8]

KEY-VALUE COLUMN- DOCUMENT GRAPH ORIENTED Performance High High High Variable

Scalability High High Variable(high) Variable

Flexibility High Middle High High

Confusion - Low Low High

Functionality Variable Minimal Variable(low) Graph Theory

7.Conclusion

In conclusion, NoSQL systems are very important for big data problems. Every type of NoSQL system submits different approach solution in big data problems. Most important things users should choose a suitable store. And it is only possible to know the disadvantages of each system and thus choose the one that suits itself. For example, key-value store disadvantages are indexing shortcomings, data querying like create, read, update, delete and data searching capacity lack. On the other hand, document store disadvantages are performance tuning, lack of join concept and data consistency problems on the graph disadvantages are API application requirement and data querying optimization problems. Last one column-oriented disadvantages are data writing and updating problems, performance tuning and data compression. Key-value store databases have high performance, scalability, and flexibility but with it functionality value is variable. Column-oriented store have high performance and high scalability but flexibility, confusion and functionality scales are not good. Graph and document stores are the same scale in flexibility. However, the scale confusion of the graphics database is high while the document database is low. Similar difference is also valid in performance. While the performance in the document database is high, this scale is variable in the graphics database.

References [1] A. Oussous, F. Benjelloun, A. A. Lahcen, and S. Belfkih, “Comparison and Classification of NoSQL Databases for Big Data,” Lab. G´enie des Syst`emes, ENSA, Ibn Tofail Univ. Kenitra, Morocco, vol. 1, no. 1, pp. 1–6, 2009. [2] A. Nayak, A. Poriya, and D. Poojary, “Type of NOSQL Databases and its Comparison with Relational Databases,” Int. J. Appl. Inf. Syst., vol. 5, no. 4, pp. 16–19, 2013. [3] T. T. Nguyen and M. H. Nguyen, “Zing Database: high-performance key-value store for large- scale storage service,” Vietnam J. Comput. Sci., vol. 2, no. 1, pp. 13–23, 2015. [4] “Investigation in mysql database and neo4j database zahraa mustafa abdulrahman al-ani june 2015,” no. June, 2015. [5] Y. Gökşen and H. A. Ş. A. N, “Veri Büyüklüklerinin Veritabanı Yönetim Sistemlerinde Meydana Getirdiği Değişim : NoSQL Induced Change of Data Size in Database Management Systems : NoSQL,” pp. 125–131, 2015. [6] S. Sharma, U. S. Tim, S. Gadia, J. Wong, R. Shandilya, and S. K. Peddoju, “Classification and comparison of NoSQL big data models,” Int. J. Big Data Intell., vol. 2, no. 3, p. 201, 2015.

176

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

[7] “Neo4j.” [Online]. Available: www.neo4j.com. [Accessed: 03-Dec-2018]. [8] M. Enst, S. Tez, K. E. Enstit, and A. Dal, “VERİTABANLARI KULLANARAK TÜRKİYE ’ DE TERÖR,” 2017.

177

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Response of Twitter Users to Earthquakes in Turkey Türkay Dereli1,2, Nazmiye Çelik2* ,Cihan Çetinkaya3

1 Iskenderun Technical University, 31200 Hatay, Turkey [email protected] 2 Gaziantep University, Industrial Engineering, 27310 Gaziantep, Turkey [email protected] [email protected] 3Adana Science and Technology University, Department of Management Information Systems 01250, Adana, Turkey [email protected]

Abstract Response of people to disasters may be different and it can vary from region to region. The expectations of individuals and their response to events can be shaped by their demographic structure. Twitter has an increasing popularity and people show their reactions to disasters on twitter. Can discourses related to disasters differ regionally on twitter? Are people more fatalistic according to their region or not? Twitter raises situational awareness and draws attention to these questions. In this work, major earthquakes and the places with frequent earthquakes in Turkey are studied. Turkish has been used as the search language, since Turkey is chosen as the studied area. “Van, deprem”, “Samsat, deprem”, “17 Ağustos, deprem”, “Gölcük, deprem”, “Çanakkale, deprem”, and “Marmara, deprem” are used for searching under six main titles. All of the searches are created in the word clouds by data visualization. Latent Dirichlet Allocation (LDA model) and Bag of Ngram words methods are used to obtain important subjects. The results show that people are talking about spiritual subjects rather than concrete subjects, but this is changing regionally. It is seen that the level of earthquake awareness of the society is not a sufficient level. Keywords: Bag of Ngram, Earthquake, LDA Model, Turkey, Twitter 1. Introduction Twitter is a popular microblog service and people can express their ideas through twitter easily. A tweet is limited to 140-caharacter. In crisis situations and emergency events, information can be obtained from social media, especially from twitter. In case of disaster, response of people to disaster may vary according to the age, gender, education, region etc. Feelings and requests may not be the same for every person but it can be questioned whether people of the same region feel common feelings or not? In literature, there are remarkable studies considering the demographic structure of disaster victims based on twitter. Demographic structure includes age, education, gender, income, occupation, race and region factors. For example, the authors analyzed tweets related to Hurricane Sandy to investigate risk awareness and concerns according to the age, ethnicity, language group and sex (Lachlan et al., 2014). Differences for using social media and communication forms before the tornado disaster were analyzed in terms of age, education, and years in Tuscaloosa (Stokes & Senkbeil, 2017). In another study, the authors analyzed the natural disasters based on at different levels such as location level, and national level considered male and female specific analysis in the light of natural language processing and opinion mining (Bala et al., 2017). Sheldon & Antony (2018) examined how college students respond to both natural and human made disaster scenarios considering gender differences in a college campus. While women notify others so that they protect themselves, men share emergency warns to provide confidence. Cvetojevic & Hochmair (2018) presented how twitter community react to Paris attacks as unexpected event around the world. Study was analyzed using exploratory methods and regression models. An approach was proposed by (Kumar et al., 2014) that identified whether a tweet is created from crisis regions using a behavior analytics. Their study included tweet characteristics and behavior observations from crisis region. Comunello & Mulargia (2017) focused on utilization models of social media, grasp of social media and obstacles to wider social media adoption. They performed number of

178

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

16 interviews in the Italian region as a county level. In a study, according to the authors, the big problem for the social media is determining posts related to the area and attaching them on a map (Venturini & Corso, 2017). Authors aimed to evaluate number of tweets that includes geographical information and determine useful information for disaster management and relief activities. So, a real time geo-targeted event observation was built by (Tsou, et al., 2017). A new system was proposed by (Kim et al., 2016) to examine spatiotemporal patterns to determine local topics using latent spatiotemporal relationships. The results showed that certain keywords had a powerful spatiotemporal closeness even though they were not in the same message. Anand & Narayana (2014) introduced a probabilistic model based on spatiotemporal model to find the location of event. Semantic analysis was done to determine negative and positive classes into tweets. How to understand demographic structures of twitter users? Mislove et al. (2011) examined twitter user population by using machine learning algorithms through gender, race and axes of geography. They obtained a non-uniform sample of population. What are people's thoughts about the earthquakes in Turkey? What are most mentioned topics about the earthquake in Twitter? There is no study examining important earthquakes together in Turkey using twitter. In this study, a content analysis is done related to major earthquakes and the places with frequent earthquakes in Turkey through twitter by using topic extraction models. Remainder of this paper is organized as follows: section 2 presents collection of necessary data and preprocessing step. Section 3 presents content analysis based on word frequencies and visualization of the data. Section 4 and 5 presents LDA model and Bag of Ngram words methods. Finally, section 6 summarizes the conclusion.

2. Data Collection and Preprocessing Turkey has been exposed to major earthquakes throughout the history. The significant earthquakes are “17 Ağustos Gölcük” and “Van” earthquakes in the near history. Number of 18.373 people died in “17 Ağustos Gölcük” earthquake in 1999, and thousands of people were injured and homeless (AFAD, 2018/1). In the year 2011, 644 people died in “Van” earthquake and 1966 people was injured (AFAD, 2018/2). Furthermore, some locations in Turkey have been exposed to medium-sized earthquakes frequently such as “Çanakkale” and “Adıyaman-Samsat” regions. In this study, above mentioned earthquakes and expected “great Marmara earthquake” are considered. A major earthquake is expected as there are two unbroken fault lines in the Marmara Sea. “Van, deprem”, “Samsat, deprem”, “17 Ağustos, deprem”, “Gölcük, deprem”, “Çanakkale, deprem”, and “Marmara, deprem” syntaxes are used for searching under six main titles through twitter till 2nd October, 2018. A number of 160 tweets from “Van” earthquake, 58 tweets from “Samsat” earthquake, 553 tweets from “17 Agustos” earthquake, 60 tweets from “Gölcük” earthquake, 157 tweets from “Çanakkale” earthquake and 73 tweets from “Marmara” earthquake has been obtained. Tweets are obtained by using Python programming language via selenium library. Steps of preprocessing level of tweets can be seen in figure 1. This level is important to obtain accurate information about data.

179

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 1: Step of preprocessing level In the first step of preprocessing, hashtags have been cleaned, punctuations and urls have been removed, stop words have been eliminated and all texts have been converted to lower case. After the tokenization process stemming process has been conducted. All these stages have been performed by Python programming language. For the “17 Ağustos” earthquake, a sample of tokenized tweets can be seen in table 1. Table 1: Tokenized of Tweets

['17ağustos', 'depreminde', 'hayatını', 'kaybeden', 'vatandaşlarımızı', 'rahmet', 'minnetle', 'anıyoruz'] ['deprem', 'riski', 'altındaki', 'istanbul', '’', 'umuzun', '500', '’', 'e', 'yakın', 'deprem', 'toplanma', 'alanlarından', 'geriye', 'kalan', 'sadece', '77', 'onların', 'olduğunu', 'bilen', 'söyleyen', 'yok', '17', 'ağustos'] ['17ağustos1999', 'acının', 'tarihi', 'var', 'tarifi', 'yok'] ['19', 'yıl', 'geçti', 'yüreğimizdeki', 'acı', 'dinmedi', '17ağustos'] ['17ağustos1999', '45', 'saniye', 'sürdü', '17480', 'kişi', 'yaşamını', 'yitirdi', '23', 'bin', 'kişiden', 'fazla', 'yaralı', '500', 'binden', 'fazla', 'kişi', 'sakat', 'kaldı',] After preprocessing operations, all texts are analyzed and created word clouds.

180

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

3. Word Clouds of Data Word frequencies of texts are created as in the seen table 2. When the “Samsat” earthquake is considered, leading words are ‘get well soon, God bless, loss of life’. For the “van” earthquake, ‘magnitude, time, and location’ words are seen. For the “17 Ağustos” earthquake, the words are based on ‘mercy, don’t forget, mention’. When the “expected Marmara” earthquake considered, “17 Ağustos” earthquake has been mentioned and especially ‘Istanbul’ is the focal point. The earthquake ‘Gölcük’ is the same with ‘17 Ağustos’ earthquake because of the being same earthquakes. Different keywords are used for handling better results. For the earthquake “Çanakkale”, ‘magnitude, earthquake depth, date, and time’ words can be seen like “Van” earthquake. All of the texts are created in the word clouds by data visualization. Table 2: Word Frequencies “Samsat” “Van” “17 Ağustos” “Marmara” “Gölcük” “Çanakkale” adıyaman vanda 17agustos istanbul 17ağustos saat olsun büyüklüğünde rahmetle denizi bir büyüklüğünde geçmiş meydana depreminde 17ağustos 17ağustos1999 derinlik meydana derinlik hayatını marmarada kocaeli büyüklük kaybı saat anıyoruz 17ağustos1999 ağustos canakkale can büyüklük yaşatmasın büyüklüğünde rahmetle meydana korusun yer unutmadık meydana marmara tarih

Word clouds of “Samsat, Van, 17 Ağustos, Marmara, Gölcük, and Çanakkale” earthquakes can be seen figure 2 to 7.

Figure 2: Word cloud of “Samsat” earthquake

181

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 3: Word cloud of “Van” earthquake

Figure 4: Word cloud of “17 Ağustos” earthquake

182

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 5: Word cloud of “Marmara” earthquake

Figure 6: Word cloud of “Gölcük” earthquake

Figure 7: Word cloud of “Çanakkale” earthquake

All earthquakes are shown compatible with their maps.

183

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

It is seen that the level of earthquake awareness of the society on twitter is not at a sufficient level. Because, plans for disaster management, resource usage and relief activities etc. have not been mentioned sufficiently. Normally, people should talk about these factors. But, people only have made spiritual wishes. Mercy words have been used for dead people. There are wishes for no recurrence of disaster again. Where the earthquake happened and what is the magnitude of earthquake are the other subjects that mentioned in the texts. It seen that people in Turkey are not ready for disaster management phases such as mitigation, preparedness, response, and relief activities. There are four phases defined for disaster operation management: mitigation, preparedness, response, and recovery phases (McLoughlin, 1985). These phases refer operations that need to make before, during and after the disasters. 4. Latent Dirichlet Allocation Blei, Ng, & Jordan (2003) proposed a generative probabilistic model of a corpus referred as Latent Dirichlet Allocation (LDA model). Each topic is described by a distribution over words. Topics created by texts with topic distribution can be seen in table 3.

Table 3: Topics Created by LDA Model

“Samsat” “Van” “17 Ağustos”

Word Score Word Score Word Score "deprem" 0.070279 "deprem" 0,098371 "17ağusto" 0,067278 "samsat" 0.064519 "van" 0,079662 "bir" 0,028077 "adıyaman" 0.057606 "vanda" 0,019312 "17agusto" 0,024015 "olsun" 0.029957 "büyüklüğünd" 0,015691 "rahmetl" 0,018719 "geçmiş" 0.027652 "meydana" 0,012674 "depremind" 0,015362 "meydana" 0.012675 "saat" 0,011467 "hayatını" 0,014831 "kaybı" 0.012675 "derinlik" 0,011467 "anıyoruz" 0,013773 "gelen" 0.011523 "yer" 0,010863 "1999" 0,013067 "rabbim" 0.011523 "büyüklük" 0,010863 "deprem" 0,012891 "korusun" 0.011523 "bir" 0,010260 "yaşatmasın" 0,012361 “Marmara” “Gölcük” “Çanakkale” Word Score Word Score Word Score "deprem" 0,095580 "deprem" 0,07744 "deprem" 0,104032 "marmara" 0,084829 "gölcük" 0,07390 "çanakkal" 0,078674 "istanbul" 0,015534 "17ağusto" 0,02026 "saat" 0,020156 "denizi" 0,013144 "bir" 0,02026 "büyüklüğünd" 0,016906 "17ağusto" 0,010755 "17ağustos1999" 0,01907 "derinlik" 0,016255 "marmarada" 0,009560 "koca" 0,01668 "büyüklük" 0,015605 "büyüklüğünd" 0,008365 "ağusto" 0,01549 "canakkal" 0,014955 "17ağustos199 0,008365 "rahmetl" 0,01430 "yer" 0,013655 "yalova"9" 0,007170 "1999" 0,01430 "tarih" 0,013655 "meydana" 0,007173 "marmara" 0,01072 "meydana" 0,013655 Extracted topics are almost same as in the word frequencies. The words in the same topic tend to be parallel. So, words can be labeled with topic names. For the “Samsat” earthquake, topic is about praying. “Van” earthquake is about technical information related earthquake. “Çanakkale” earthquake is also about technical information like “Van” earthquake. For the “Marmara” earthquake, topic is about expected earthquakes in Istanbul and Yalova. “17 Ağustos” and “Gölcük” earthquakes are the same

184

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY earthquakes. Topic for “Gölcük” is the about “17 Ağustos” and topic for “17 Ağustos” is about remembering those days and respect. 5. Bag of Ngram Words A bag-of-n-grams model enrolls the number of times that each n-gram seems in every single document of a corpus (MathWorks, 2018). In table 4, topics obtained with Bag of Ngram model can be seen. Table 4: Bag of Ngram Model

“Samsat” “Van” “17 Ağustos” "deprem" "samsat" "deprem" "meydana" "sesimi" "duyan" "adıyaman" "geldi" "var" "adıyaman" "samsat" "büyüklük" "derinlik" "depremind" "hayatını" "deprem” "deprem" "kaybeden" "geçmiş" "olsun" "vanda" "büyüklüğünd" "17aðusto" "depremind" "adıyaman" "deprem" "hayatını" "deprem" "büyüklüğünd" "deprem" "rahmetl" "anıyoruz" "adıyaman""samsat" "meydana" "17aðusto" "samsat" "ilçesind" "derinlik" "deprem" "hayatını" "kaybedenleri" "meydana" "yer" "rahmetl" “Marmara” “Gölcük” “Çanakkale” "17ağustos1999" "marmara" "gölcük" "17ağusto" "büyüklük" "derinlik" "deprem" "deprem" "deprem" "deprem" "meydana" "17ağusto" "deprem" "meydana" "geldi" "deprem""şehitlerini" "geldi" "beşiktaş" "medya" "rahmetl" "anıyor " "geçmiş" "olsun" "grup" yakınlarına" "çanakkal" "marmara" "denizi" "deprem" "gölcük" "büyüklüğünd" "deprem" "büyüklük" “17ağusto" "meydana" "denizi" "büyüklük" "17ağustos1999" "gölcük" "bir" "deprem" "derinlik" "deprem" "meydana"

Ngram length is 3 as seen in table. Successive words can provide logical information about texts. For example, for the “17 Ağustos” earthquake, the slogan of “sesi mi duyan var mı? (In Turkish) / Is there anyone who can hear my voice?” is at the forefront. In general, same things have been mentioned in texts as in the latent dirichlet allocation. 6. Conclusion In this study, important earthquakes and the places with frequent earthquakes in Turkey are analyzed in the twitter context. “Van, deprem”, “Samsat, deprem”, “17 Ağustos, deprem”, “Gölcük, deprem”, “Çanakkale, deprem”, and “Marmara, deprem””syntaxes are used for searching under six main titles. After the preprocessing stages, word frequencies are determined and visualized through word clouds. Latent Dirichlet Allocation (LDA model) and Bag of Ngram words methods are used to obtain important subjects. Results show that people are not conscious against earthquakes on twitter. Resource use and gathering places after earthquakes are not discussed by people. Things that need to be done in the earthquake phases are not mentioned. People in Turkey are not ready for disaster management phases such as mitigation, preparedness, response, and relief activities. People are talking about spiritual subjects rather than concrete subjects. People always wish not to reoccurrence of earthquakes. The words of mercy and prayer are used frequently. It is seen that the level of earthquake awareness of the society is not a sufficient level.

185

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

In future studies, people from all Turkey cities can be examined from the perspective of the earthquake conscious on twitter. References

AFAD.(2018/1,October14).Retrieved from AFAD:https://deprem.afad.gov.tr/tarihteBuAy?id =37

AFAD. (2018/2, October 14). Retrieved from AFAD: https://www.afad.gov.tr/tr/2385/Van-Depremi-Hakkinda

Bala, M. M., Rao, M. S., & Babu, M. R. (2017). Sentimend trends on natural disasters using location based twitter opinion mining. International Journal of Civil Engineering and Technology (IJCIET), 8:9-19.

Blei, D. M., Ng, A. Y., & Jordan, M. I. (2003). Latent Dirichlet Allocation. Journal of Machine Learning Research, 993-1022.

Comunello, F., & Mulargia, S. (2017). A #cultural_change is needed. Social media use in emergency communication by Italian local level institutions. Proceedings of the 14th ISCRAM Conference, (pp. 512-521). Albi, France.

Cvetojevic, S., & Hochmair, H. H. (2018). Analyzing the spread of tweets in response to Paris attacks. Computers, Environment and Urban Systems, 71:14-26.

Kim, K.-S., Kojima, I., & Ogawa, H. (2016). Discovery of local topics by using latent spatio-temporal relationships in geo-social media. International Journal of Geographical Information Science, 30:1899–1922.

Kumar, S., Hu, X., & Liu, H. (2014). A Behavior Analytics Approach to Identifying Tweets from Crisis Regions. Proceedings of the 25th ACM conference on Hypertext and social media, (pp. 255-260). Santiago, Chile.

Lachlan, K. A., Spence, P. R., & Lin, X. (2014). Expressions of risk awareness and concern through Twitter:On the utility of using the medium as an indication of audience needs. Computers in Human Behavior, 35:554-559.

MathWorks. (2018, August 30). Retrieved from MathWorks: https://www.mathworks.com/help/textanalytics/ref/fitlda.html

McLoughlin, D. (1985). A framework for integrated emergency management. Public Administration Review, 165- 172.

Mislove, A., Lehmann, S., Ahn, Y.-Y., Onnela, J.-P., & Rosenquist, J. N. (2011). Understanding the Demographics of Twitter Users. Proceedings of the Fifth International AAAI Conference on Weblogs and Social Media, (pp. 554- 557).

Anand,S. & Narayana, K. (2014). Earthquake reporting system development by tweet analysis. International Journal of Emerging Engineering Research and Technology, 96-106.

Sheldon, P., & Antony, M. G. (2018). Sharing Emergency Alerts on a College Campus: How Gender and Technology Matter. SOUTHERN COMMUNICATION JOURNAL, 83:167-178.

Stokes, C., & Senkbeil, J. C. (2017). Facebook and Twitter, communication and shelter, and the 2011 Tuscaloosa tornado. Disasters, 41:194-208.

Tsou, M.-H., Jung, C.-T., Allen, C., Yang, J.-A., Han, S. Y., Spitzberg, B. H., & Dozier, J. (2017). Building a Real-Time Geo-Targeted Event Observation (Geo) Viewer for Disaster Management and Situation Awareness. In Advances in Cartography and GIScience, Lecture Notes in Geoinformation and Cartography (pp. 85-98). Springer International Publishing.

Venturini, L., & Corso, E. D. (2017). Analyzing spatial data from Twitter during a disaster. IEEE International Conference on Big Data, (pp. 3779-3783). Boston, MA, USA.

186

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Review of Trust Parameters in Secure Wireless Sensor Networks Cansu Canbolat1, İpek Abasıkeleş-Turgut2

1,2Department of Computer Engineering, Iskenderun Technical University, Turkey [email protected] [email protected]

Abstract

Since Wireless Sensor Networks (WSNs) are commonly used in military applications for monitoring hostile environment, the information collected from the network has to be trustworthy. However, unguarded nature of wireless medium and limitation of the resources of WSN nodes necessitate of developing security solutions for these networks. Routing attacks are among the most dangerous attacks on WSNs that disrupt the flow of data originating from the sensor nodes and are routed to a central station, called the base station. The studies in literature on routing attacks are divided into two groups based on intrusion detection or trust-based schemes. The focal point of intrusion detection systems is finding the resource of the attacker and hence, recover the network. However, designing such a system is an energy consuming solution and is not suitable for sensor nodes with limited battery power. On the other hand, trust-based schemes enable data of the sensor nodes to by-pass the attacked area by using safety paths. Since the purpose of this approach is not to detect the attack but to discover the safe routes around it, it is considered to be an energy effective solution. For carrying out a safe transmission, data needs to be transmitted through trustworthy nodes. The trust value of a node can be calculated by various ways. In this study, the trust parameters, including social (intimacy and honesty) and QoS (energy and unselfishness) trust; sending rate factor, consistency factor and packet loss rate factor, used in secure WSNs are investigated. In future work, various trust values will be modelled and compared in cluster based WSNs. Keywords: Wireless Sensor Networks, Trust-based Routing, Survey, Security

1.Introduction Wireless Sensor Networks (WSN) are typically consist of tiny collaborating sensor nodes, which collect desired information from a monitored area and transmit to a central node, called base station, through wireless links. Base station either utilizes data locally or send to Internet by using gateways (Rawat, Singh, Chaouchi, Bonnin, 2014). One of the advantages of WSNs is providing low cost solutions to various real-world problems, including military applications, environmental applications, smart homes, health monitoring etc (Tripathi, Gaur, Laxmi, ,2013). However, these low-cost sensor devices are equipped with limited storage, battery, computational and communicational capabilities, which turns into a problem in these networks (Patil, Khanagoudar, 2012). As the size of a WSN, which is especially projected for mission-critical tasks, increases, security becomes one of the most important issues (Chen, Makki, Yen and Pissinou, 2009). Although the advantages of WSNs bring about extensive usage in critical applications, including military and commercial, these networks catch intruder’s interest due to deploying in hostile environments. A security hole in a military application would strengthen enemies’ hand and cause loss of territorial conflict or war. Similarly, a confidential patient health record should be protected from unauthorized usage in a heath care application (Butun, Morgera, Sankar,2014). In view of the constrained resources of sensor nodes, WSNs are vulnerable to various attacks (Sedjelmaci, Senouci, & Feham,2012). Routing attacks, which targets to disrupt the functioning of network layer, are one of the most dangerous attacks that threatens WSNs (Virmani, Soni, Chandel, & Hemrajani,2014). These attacks, including sybil, blackhole, sinkhole, selective forwarding, hello flood, grey hole and wormhole, damage the network by injecting false routing packets, modifying, blocking or replicating data packets (Zin,

187

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Anuar, Kiah,Ahmedy,2015). Hence, providing a secure routing becomes a primary goal in WSN applications (Zin, Anuar, Kiah,Pathan, 2014). The studies in literature focusing on routing attacks are either based on intrusion detection systems or trust-based routing. An intrusion detection system (IDS) tracks and analyses the behaviour of the network or a specific target node. If the IDS catches a suspicious activity, then an alarm is triggered to warn the network (Sedjelmaci, Senouci, & Feham,2012). The rest of the process is under the control of intrusion response systems (IRS). On the other hand, in trust-based approaches, the purpose is not capturing the attack but establishing a secure path from source node to destination node. For this purpose, the trust of the system nodes is calculated and the network traffic flows over the nodes having higher trust values. Considering the resource constraints of WSNs, including battery, memory, processor and wireless power, energy consuming defence mechanisms such as public key infrastructure or IDS techniques are not suitable for these networks (Bao, Chen, Chang, Cho, 2011). However, trust-based systems reduce the overhead of these approaches by allowing trusted nodes to communicate and block the untrusted ones from the network (Bao, Chen, Chang, Cho,2012). In this study, we have investigated recent trust-based approaches used in cluster-based WSN architectures. The studies are compared over trust parameters, the usage of trust mechanism and cluster head election techniques. Section 2 describes the clustering routing approach, while Section 3 explains trust based systems and compares the recent studies in literature. Finally, Section 4 concludes the paper. 2.Clustering Architecture Routing in WSNs can be classified as flat, hierarchical and adaptive. The nodes in the network have the same capabilities in flat routing, while they have different roles in hierarchical architecture. On the other hand, in adaptive approach, some system parameters are controlled according to the state of the network and available level of the energy. In hierarchical approaches, the network is composed of clusters to alleviate network operations, including data aggregation, routing and data query. Since hierarchical routing adjusts the capacity of the system and reduce the communication overhead, efficient energy usage is ensured in these systems (Ökdem, Derviş,2007; Patil vd. , 2012).

In cluster-based routing (Figure 1), neighbouring nodes cooperate to form clusters. A cluster head (CH) is elected in each cluster to manage operations. Member nodes (MN) transmit their sensed data to related CH, which directs aggregated data to a central authority, called Base Station (BS) (Tohma, Aydın, Abasıkeleş-Turgut, 2015).

Figure 1. A cluster-based WSN architecture

188

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

3. Trust Based Systems Some studies about secure routing in literature are based on traditional security mechanisms, such as cryptologic solutions. Since these approaches consumes much energy and have high memory usage, they are not suitable for WSNs with limited resources, including bandwidth, power, memory etc. Morever, cryptologic approaches can only solve external attacks, but does not offer a solution for internal attacks. Therefore, traditional security mechanisms are out of keeping with WSNs (Duan ,Yng, Zhu, Zhao, 2014). Trust management mechanisms are suggested in literature for solving these problems. The definition of trust is a double relationship between a subject and an object. A trust management system records the behaviors of the nodes, such as object or interaction, and uses these past experiences to calculate recent trust value of the nodes. The trust value of a node provides an insight into future behaviour of the node (Chen, Liang,Chen,2015). In this paper, five different trust-based solutions on cluster-based WSN is investigated. Table 1 shows the name and year of these studies, the trust parameters used in these papers, the purpuse of trust evaluation and CH election mechanisms. Table 1. Five different trust-based solutions on cluster-based WSN

NAME AND TRUST PARAMETERS PURPOSE CH ELECTION YEAR OF TRUST MECHANISM TRUST 1. Intimacy CH and MN The selection of CH is MANAGEMENT 2. Honesty monitoring beyond the scope of this PROTOCOL 3. Energy study. HEED (Younis, (Bao, Chen, Chang, 4. Unselfishness Fahmy, 2004) Cho, 2012) algorithm is used. 1. Sending Rate Factor Selection of CH is elected by trust TLES 2. Consistency Factor CH value, residual energy (Chen, 3. Packet Loss Rate of the nodes, and node Liang,Chen,2015) Factor density.

1. Intimacy CH and MN The selection of CH is HTBID 2. Honesty monitoring beyond the scope of this (Dhakne, Chatur, 3. Energy study. HEED (Younis, 2017) 4. Unselfishness Fahmy, 2004) algorithm is used. 1. Interactive trust CH and MN The node in cluster with TBHTA-WSN 2. Honesty trust monitoring the highest energy is (Sheera, 3. Content trust selected as CH. Maragatharaja,2018) TRUST-BASED 1. Positive transaction rate Selection of CH is elected by trust CH ELECTION 2. Negative transmission CH value and fuzzy logic. PROTOCOL rate (Salehi, Karimian, 2017)

189

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Bao et. al (2012) have proposed a trust evaluation by using both social network and communication network parameters. They have four different trust parameters: two of them are obtained from social trust values (including intimacy and honesty), while the rest are QoS trust values (including energy and unselfishness). The trust parameter of intimacy shows the interaction degree of two nodes; honesty evaluates whether the node is malicious; energy reflects the sufficiency of the node and unselfishness demonstrates the tendency of the node to increase the welfare of the system. By using the proposed trust mechanism MNs monitor the other MNs in the same cluster and report their trust evaluation periodically to CHs. Similarly, the trust evaluation of CHs about their neighbours are reported to base station. The selection of CH is beyond the scope of this study and HEED (Younis, Fahmy, 2004) algorithm is used. For evaluating the performance of proposed method, it is applied both on geographic routing and a trust- based IDS system. The results show that proposed trust based IDS yield better performance than traditional anomaly-based IDS, while trust-based geographic routing outperforms flooding-based routing in terms of delay and delivery rate of the messages.

Chen el. al (2015) have proposed an energy effective trust-based routing, called TLES, for WSNs. TLES is used to elect CHs. In the proposed architecture CHs are elected by their trust values, energy levels and neighbour numbers. Three different trust parameters, including sending rate factor, consistency factor and packet loss rate factor, are used. After CHs are elected and clusters are formed, data is flowed over a weighted tree. The relay CHs in the tree are chosen by neighbour numbers, energy levels and distance to base station of the nodes. The proposed method is compared with LEACH, LEACH-MF and CMRA and they have reported that TLES outperforms than the equivalent approaches in terms of energy consumption and malicious node elimination.

Dhakne and Chatur (2017) have proposed an IDS system called HTBID. The same trust parameters are used with Bao et. Al (2012). Three types of attacks are modelled, including black hole, selective forwarding and on-off. By using NS2 platform, simulations are conducted to obtain false positive rates, packet delivery ratios and attack detection rates. HTBID is compared with EDTM and the results have reported that HTBID yields better performance than EDTM.

Sheera and Maragatharajan (2018) have proposed a hierarchical trust-based algorithm for clustered WSNs, called TBHTA-WSN. Three types of trust parameters, including interactive trust, honesty trust and content trust, are used. A two levet trust mechanism is used. The first level trust evaluation is between MNs and CHs, while the second level is between CHs. By using MATLAB platform, simulations are conducted to estimate data rate, the number of trusted nodes and malicious nodes.

Salehi and Karimian (2017) have proposed a trust-based CH election for WSNs. After estimating the trust of neighbours, each node communicates within a cluster to elect the most trusted node as CH. If the trust value of a node is higher than a predefined threshold value, than the node becomes a candidate for CH. Final CHs are chosen by their energy levels and trust values. The trust of a node is estimated by using positive and negative transaction rates.

4.Result and Discussion

WSNs consist of tiny sensor devices with limited battery power, storage and communication capability. Due to unprotected environment and the resource constraints of the sensor nodes, data must be protected from malicious usage/access. Security becomes an important issue, especially as the size of WSN increases in military and commercial applications. Routing attacks damage the network by modifying, blocking and replicating data packets in the network layer. Therefore, a secure routing becomes the most important priority in the system. The key for overcoming the routing attacks is either IDS or trust-based systems. Since IDS tries to find the malicious acitivity in the network, it is an energy consuming approach. However, trust-based systems constructs a secure path around malicious regions and provides an effective solution. In this paper, we have examined recent trust-based routing protocols for cluster-based WSN architectures. The studies are compared over trust parameters, the usage of trust

190

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY mechanism and cluster head election techniques. In future work, various trust values will be modelled and compared in cluster based WSNs.

5.Reference Bao, F., Chen, R., Chang, M., & Cho, J. H. (2011, June). Trust-based intrusion detection in wireless sensor networks. In Communications (ICC), 2011 IEEE International Conference on(pp. 1-6). IEEE. Bao, F., Chen, R., Chang, M., & Cho, J. H. (2012). Hierarchical trust management for wireless sensor networks and its applications to trust-based routing and intrusion detection. IEEE transactions on network and service management, 9(2), 169-183. Butun, I., Morgera, S. D., & Sankar, R. (2014). A survey of intrusion detection systems in wireless sensor networks. IEEE communications surveys & tutorials, 16(1), 266-282. Chen, X., Makki, K., Yen, K., & Pissinou, N. (2009). Sensor network security: A survey. IEEE Communications Surveys and Tutorials, 11(2), 52-73. Chen, Z., He, M., Liang, W., & Chen, K. (2015). Trust-aware and low energy consumption security topology protocol of wireless sensor network. Journal of Sensors, 2015. Dhakne, A. R., & Chatur, P. N. (2017). Design of Hierarchical Trust based Intrusion Detection System for Wireless Sensor Network [HTBID]. International Journal of Applied Engineering Research, 12(8), 1772-1778. Duan, J., Yang, D., Zhu, H., Zhang, S., & Zhao, J. (2014). TSRF: A trust-aware secure routing framework in wireless sensor networks. International Journal of Distributed Sensor Networks, 10(1), 209436. Ökdem, S., & Derviş, K. (2007). Kablosuz Algılayıcı Ağlarında Yönlendirme Teknikleri. Akademik Bilişim. Patil, S. S., & Khanagoudar, P. S. (2012). Intrusion Detection Based Security Solution for Cluster .Based WSN. International Journal of Advanced Research in Computer Engineering & Technology (IJARCET), 1(4), pp- 331. Rawat, P., Singh, K. D., Chaouchi, H., & Bonnin, J. M. (2014). Wireless sensor networks: a survey on recent developments and potential synergies. The Journal of supercomputing, 68(1), 1-48. Salehi, M., & Karimian, J. (2017). A Trust-based Security Approach in Hierarchical Wireless Sensor Networks. I.J. Wireless and Microwave Technologies, 2017, 6, 58-67 Sedjelmaci, H., Senouci, S. M., & Feham, M. (2012, July). Intrusion detection framework of cluster-based wireless sensor network. In Computers and Communications (ISCC), 2012 IEEE Symposium on (pp. 000893- 000897). IEEE. Shari S.& Maragatharajan M.(2018). Trust-Based Hierarchical Trust Algorithm For Wireless Sensor Network. International Journal of Pure and Applied Mathematics, Volume 118, No. 24 2018. Tripathi, M., Gaur, M. S., & Laxmi, V. (2013). Comparing the impact of black hole and gray hole attack on LEACH in WSN. Procedia Computer Science, 19, 1101-1107. Tohma, K., Aydın, M. N., & Turgut, İ. A. (2015, May). Improving the LEACH protocol on wireless sensor network. In Signal Processing and Communications Applications Conference (SIU), 2015 23th (pp. 240-243). IEEE. Virmani, D., Soni, A., Chandel, S., & Hemrajani, M. (2014). Routing attacks in wireless sensor networks: A survey. arXiv preprint arXiv:1407.3987. Younis, O., & Fahmy, S. (2004). HEED: a hybrid, energy-efficient, distributed clustering approach for ad hoc sensor networks. IEEE Transactions on mobile computing, 3(4), 366-379. Zin, S. M., Anuar, N. B., Kiah, M. L. M., & Pathan, A. S. K. (2014). Routing protocol design for secure WSN: Review and open research issues. Journal of Network and Computer Applications, 41, 517-530. Zin, S. M., Anuar, N. B., Kiah, M. L. M., & Ahmedy, I. (2015). Survey of secure multipath routing protocols for WSNs. Journal of Network and Computer Applications, 55, 123-153.

191

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Superiorities of Deep Extreme Learning Machines against Convolutional Neural Networks Gokhan Altan1, Yakup Kutlu2

1,2 Department of Computer Engineering, Iskenderun Technical University, Turkey [email protected] [email protected]

Abstract

Deep Learning (DL) is a machine learning procedure for artificial intelligence that analyzes the input data in detail by increasing neuron sizes and number of the hidden layers. DL has a popularity with the common improvements on the graphical processing unit capabilities. Increasing number of the neuron sizes at each layer and hidden layers is directly related to the computation time and training speed of the classifier models. The classification parameters including neuron weights, output weights, and biases need to be optimized for obtaining an optimum model. Most of the popular DL algorithms require long training times for optimization of the parameters with feature learning progresses and back-propagated training procedures. Reducing the training time and providing a real- time decision system are the basic focus points of the novel approaches. Deep Extreme Learning machines (Deep ELM) classifier model is one of the fastest and effective way to meet fast classification problems. In this study, Deep ELM model, its superiorities and weaknesses are discussed, the problems that are more suitable for the classifiers against Convolutional neural network based DL algorithms.

Keyword(s): Deep Learning, Deep ELM, fast training, LUELM-AE, Hessenberg, autoencoder

1) Introduction

Deep Learning (DL) is a classification method that is a specific version of the artificial neural networks (ANN). The DL was applied to various types of disciplines including computer vision, object detection, semantic segmentation, diagnosis systems, time series analysis, image generating, and more. The ANN model has limitations on layer size, number of neurons depending of computation capability of the CPU. Hence, whereas the number of the ANN model increases, the classification parameters to optimize increase along with the model advancing. The DL has advantages of pre-defined parameters and using shared weights in training stages. It has been the easiest way of producing new ANN models using most effective framework background for last decades. The DL has effective algorithms that are group demanding on the unsupervised and supervised stages at the training phases. The DL is a two- step classifier. At the first stage, the classification parameters are defined without labels using an unsupervised algorithm including filtering, restricted Boltzmann machines, autoencoders, self- organizing maps, convolution, stage mapping and more. The pre-defined parameters are fed into the ANN model with similar structure of the unsupervised stage with the labels to optimize the classification parameters for minimum error using back-propagation algorithms including contrastive divergence, dropout, extreme learning machines (ELM) and more.

Convolutional neural networks (CNN) model is the most popular method in the DL. The CNN is utilized to handle for visual imagery including the computer vision and object detection solutions. It provides to collect deterministic features from the input images using low- and high-order features with specific filters. The convolution filter size and number of the filters enables to perform self-assessment on the images. The convolutional layer and extracting deterministic variable order features are called as

192

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY feature learning for the DL stages. The CNN was utilized to perform object detection, time-series signals, diagnosis of the different diseases using biomedical signals, face recognition and similar disciplines. Lawrance et al (1997) proposed a face recognition model and compared the self-organizing map approach with the CNN. They also performed hybrid classifier models on face images from 40 subjects. They highlighted the efficiency of the hierarchical set of layers on CNN for facial recognition models. Li et al. (2015) analyzed the face images to extract a face recognition model. They structured a cascade architecture on CNN. Their model had focused on multiple resolutions using central processing unit (CPU) and graphical processing unit (GPU) supports. Kalchbrenner et al. (2014) analyzed the sentences using CNN. They performed an accurate natural language processing model. The CNN has increased the performance of semantic word modelling system about 25%. Ciresan et al. (2011) suggested a CNN model for handwritten character optimization. They achieved 0.27% error rate of classification on MNIST dataset. Their analysis is a path for the next character optimization techniques with the simplicity of the model. The biggest step on modeling the CNN on big image datasets was ImageNET progresses and the analysis on it. Krizhevsky et al. (2012) has suggested ImageNet with novel models and layer sizes for image classification models. The analyzed image dataset is large and effective enough as a big data problem.

Deep ELM classifier is an alternative method to the DL algorithms. It is a fast and high- generalized capability model against the DL algorithms. It is based on the autoencoder modeling approach that creates representations in a selective output size. At the first stage, autoencoder kernel generates different presentations of the input data, and transfers the output weights into the queued layers of the deep model (Ding et al., 2015). At the last stage, last generated representations are fully connected to the supervised ELM classifier with the class labels. Deep ELM model was performed to analyze time- series for modeling diagnosis systems, image classification, character optimization and object detection (Altan and Kutlu, 2018a).

The remaining of the study is organized as definition and specifications of the Deep ELM and CNN in detail, applications of the classifiers, superiorities and performance comparison on various disciplines in the material method. The comparisons are based on the classification accuracy rates and training time for the same datasets in the literature. The advantages of the Deep ELM and CNN algorithms are discussed at the results section. 2) Materials and Methods

Classifier models are explained. The specifications of the classifier models with optimization parameters and requirements of the classifier models are shared for the beginner researchers on the DL applications. Even though the CNN and Deep ELM models are similar classifier models with the pre- defining specifications and supervised stages, the mathematical solutions and the workloads of the algorithms are completely different from each other. 2.1. Convolutional Neural Networks

The CNN is a main and frequently used type of DL for image-based approaches. Advantages of the CNN are standing feature learning with convolutional processes with iterated filters and sizes, and extracting the most significant bits at a specified range on the image using pooling process. The feature-learning phase of the CNN stands for the unsupervised stage of the model. The extracted pooling inputs are fed into the fully connected neural network model.

The CNN considers angular edges with different edge-finding filters, transitions from dark to light, light to dark and calculates each one separately. The edges are usually detected in the first layers

193

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY of a CNN. Whereas performing all of these convolutional calculations, there is a difference between the input size and the output size.

After convolutional layer, the calculation of the size difference between the input sign and the exit sign is possible. This is performed by adding extra pixels to the input matrix. This is called a padding. If the input matrix is n by n, the filter matrix (f by f) is the same size as the input matrix; (n + 2p-f + 1) * (n + 2p-f + 1) is applied. The value represented by id p is a matrix that is the pixel size added to the input matrix, the padding value. To determine this, p = (f-1) / 2 equation is used.

The stride stands for the weight matrix for the conversion operation will shift the filter in steps of one or more steps on the image. Stride is another parameter that affects the output size directly. It allows you to analyze low-level and high-level attributes on the entire image.

The max pooling method is usually used in the pooling layer. There are no parameters learned in this layer of the network. It decreases the height and width information by keeping the channel number of input matrix constant. Pooling layer performs decreasing image size for extracting distinctive features for smaller ones. At the fully connected layer, the bits of last convolved image is transferred as input values to the ANN. 2.2. Deep Extreme Learning Machines

ELM model is a feedforward neural networks structure with a single hidden layer. It is usually utilized as classifier, application of regression, clustering, decoding, encoding and feature learning. ELM does not require optimization for the classification parameters as different from ANN. The classification parameters including output weights, weight of hidden neurons are generated according to the input feature set and class labels. The generated output and input weights are never be tuned. It is a fast and stable algorithm for the classification stages. Hence, it is more stable and appropriate for real-time and instantaneous applications. Using only a single layer has limitations for detailed analysis of the input data.

The theorem of ELM (Huang et al., 2006; Huang and Chen, 2007) is randomly stated input weights provides fast learning and the universal estimation ability using least mean square algorithm for defining outputs. 훽 is output weight matrix, 푇 stands for training data, 퐻 is randomly generated hidden layer output matrix, 퐻† is the Moore-Penrose inverse of matrix of 퐻.

훽 = 퐻†푇 (Eq. 1)

Training process of the ELM is as follows:

1 −1 훽 = 퐻푇 ( + 퐻퐻푇) 푇 (Eq. 2) 휆

Deep ELM model is the enhanced type of the ELM. Deep ELM enables using multi-layer hidden nodes in the model with the ELM kernels. The autoencoder kernels are the basis of the transformation of the ELM into the Deep ELM models. An autoencoder is algorithm that generates the output weights in unsupervised ways effectively (Liou et al., 2014). The autoencoder generates a representation of input data with decreased or increased size of dimensionality models. The autoencoder algorithms are popular algorithms to enhance and accelerate DL classification stages (Altan and Kutlu 2018b). The ELM autoencoder gets the input matrix as also output matrix to generate the encoder vector. The encoder vector does not need optimization and carries deterministic and significant characteristic for input

194

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY feature set. The transpose of calculated output weight (훽푇) is used as the weight between two adjacent layers. At the last layer, the training progress is finalized with traditional ELM model in supervised ways. 3) Results and Discussion

The DL algorithms are effective ways for classification stages. Nevertheless, they requires optimization and feature learning stages that are time-consuming progresses. Especially, optimization of the classification parameters including output weights, learning rates, output functions, layer size, neuron size at each layer, convolution size and number of filtering channels, and more is a hard-to decide options for creating an accurate model for the disciplines. Reducing the number of the optimization parameters at training stage accelerates decision accuracy for a specific training range.

The CNN has an unchallengeable system performance on the different classification problems. Some studies are also performed ELM models to the convolutional features for achieving fast classification results. The Deep ELM has high and effective achievements on the same datasets in remarkable training time. The advantages of the Deep ELM is generalization performance, training speed and decoding application by the use of autoencoder models. It also has ability to analyze the input data in detail by using many hidden layers and autoencoder kernels. Each kernel and input with different sizes generates different presentations of the input feature set. The CNN has ability to extract low- and high-level features using convolutional layers. Nevertheless, using rectified linear unit activation function generates the sharp features from the images. CNN also utilized for assessing the time-series signal by converting them into the images. In our opinion, it is a disadvantage of the CNN, by an additional progress to the classification problems. For any signal, it is a handicap using non-standardized signals. Table 1 presents some studies on same problems.

Table 1 – Training speed and classification performance comparison of CNN and Deep ELM Method Accuracy Training Time MNIST character recognition Deep ELM 99.14% 281.37s (CPU) CNN 99.05% 19 hours (GPU) 3D Shape Classification Deep ELM 81.39% 306.4s (CPU) CNN 77.32% > 2 days (GPU) Traffic sign recognition Deep ELM 99.56% 209s (CPU) CNN 99.46% > 37 hours (GPU) CNN + ELM 99.48% 5 hours (CPU)

The CNN needs long time processes. The main improvements on DL routes the researchers to compose fast training methods. Deep ELM were proposed to advance the training capabilities for reducing time and generalization capabilities. Deep ELM model consists of autoencoder models, which are fast and effective models for generating different representations of the input data.

The ELM, which is a single layer neural network model, is formed to the many hidden layer using autoencoder kernels unsupervised ways. The proposed ELM autoencoder kernels including lower-upper triangularization (Altan and Kutlu, 2018a) and Hessenberg decomposition (Altan and Kutlu, 2018c)

195

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY have improved the generalization capability of the DL approaches. The Deep ELM provides reducing fast training with simplest mathematical solutions. Deep ELM is an adaptable method that can be integrated even for supervised and unsupervised stages of the classifiers. It is possible to integrate the ELM autoencoder kernels into the filtering progress to generate the significant presentations of the input data. Using max pooling for the Deep ELM structures would provide extracting heavy low- and high- level features from the intended sizes of the input signals. The Deep ELM is also more compatible for the time-series to analyze the long-term signals using encoding specifications.

References 1. Altan, G., Kutlu, Y, Pekmezci, AÖ, Yayik, A., (2018a), Diagnosis of Chronic Obstructive Pulmonary Disease using Deep Extreme Learning Machines with LU Autoencoder Kernel, 7th International Conference on Advanced Technologies (ICAT’18), 28 April-1 May 2018, p. 618-622 2. Altan, G., Kutlu, Y, (2018b), Generative Autoencoder Kernels on Deep Learning for Brain Activity Analysis, Natural and Engineering Sciences, 10 October 2018, ISSN: 2458-8989, Vol.3, Issue.3, pp.311–322, https://doi.org/10.28978/nesciences.468978 3. Altan, G., Kutlu, Y, (2018c), Hessenberg Elm Autoencoder Kernel For Deep Learning, Journal of Engineering Technology and Applied Sciences, Volume 3, Issue 2, 30 August 2018, pp. 141-151, e- ISSN 2548-0391, https://doi.org/10.30931/jetas.450252. 4. Ciresan, D. C., Meier, U., Gambardella, L. M., & Schmidhuber, J. (2011, September). Convolutional neural network committees for handwritten character classification. In Document Analysis and Recognition (ICDAR), 2011 International Conference on (pp. 1135-1139). IEEE. 5. Ding, S., Zhang, N., Xu, X., Guo, L. and Zhang, J.. (2015). Deep Extreme Learning Machine and Its Application in EEG Classification, Mathematical Problems in Engineering, vol. 2015, Article ID 129021, 11 pages, 2015. https://doi.org/10.1155/2015/129021. 6. Huang, G-B, Chen, L, Siew, C-K. 2006. Universal approximation using incremental constructive feedforward networks with random hidden nodes. IEEE Trans Neural Netw., 17(4):pp.879–92. 7. Huang, G-B, Chen, L. 2007. Convex incremental extreme learning machine. Neurocomputing. 70, pp.3056–62. 8. Kalchbrenner, N., Grefenstette, E., & Blunsom, P. (2014). A convolutional neural network for modelling sentences. arXiv preprint arXiv:1404.2188. 9. Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems (pp. 1097-1105). 10. LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), pp.436. 11. Lawrence, S., Giles, C. L., Tsoi, A. C., & Back, A. D. (1997). Face recognition: A convolutional neural-network approach. IEEE transactions on neural networks, 8(1), 98-113. 12. Li, H., Lin, Z., Shen, X., Brandt, J., & Hua, G. (2015). A convolutional neural network cascade for face detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 5325-5334). 13. Liou, C.-Y., Cheng, C.-W., Liou, J.-W., and Liou, D.-R., 2014. Autoencoder for Words, Neurocomputing, vol.139, pp.84–96 (2014), doi:10.1016/j.neucom.2013.09.055 14. Tang, J., Deng, C. and Huang, G., (2016) "Extreme Learning Machine for Multilayer Perceptron," in IEEE Transactions on Neural Networks and Learning Systems, vol. 27, no. 4, pp. 809-821, April 2016. doi: 10.1109/TNNLS.2015.2424995 15. Altan, G., & Kutlu, Y. (2018). Hessenberg Elm Autoencoder Kernel For Deep Learning. Journal of Engineering Technology and Applied Sciences, 3(2), pp.141-151. 16. Altan, G., Kutlu, Y., Pekmezci, A. Ö., & Yayık, A. (2018). Diagnosis of Chronic Obstructive Pulmonary Disease using Deep Extreme Learning Machines with LU Autoencoder Kernel. In 7th International Conference on Advanced Technologies (ICAT’18) vol.28, pp. 618-622

196

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

The Changing ERP System Requirements in Industry 4.0 Environment: Enterprise Resource Planning 4.0 Burak Erkayman

Department of Industrial Engineering, Atatürk University, Turkey [email protected]

Abstract Fourth Industrial revolution: Industry 4.0 represents the fourth industrial revolution in industrial markets with smart manufacturing environment. It is considered as a revolution that carries manufacturing processes to a new industrial level by introducing more flexible, more efficient and faster technologies to achieve higher industrial performance. Industry 4.0 refers to a transformation that focuses heavily on virtualization, interconnectivity, integration, machine learning and real-time data to create a more compact and better-integrated ecosystem for companies that concentrate on production and supply chain management. Industry 4.0 affects all business processes and functions; therefore, the entire IT infrastructure is obliged to be redesigned. This involves the systems: material handling, production planning and control, sales management and logistics. ERP (Enterprise Resource Planning) is a system that enables enterprises to use the labor force, machines and materials they need to produce their products or services efficiently. ERP systems play a critical role to execute business processes by taking the advantages of using common and defined data structure from one system. ERP systems provides access to corporate data through multiple activities using common structures, definitions and common user experiences. Industry 4.0 aims to manage complexity and improve connectivity through the communication channels of each element involved in the manufacturing. Industry 4.0 will not change the definition of ERP systems, but it will change the role of ERP systems. This paper provides an insight into the ERP systems, which will be used in Industry 4.0 environment. The managerial, technical and process based requirements and ERP related Industry 4.0 concepts are also discussed in this study.

Keywords: Industry 4.0, ERP

1. Introduction

ERP refers to the systems and software packages used by organizations to manage day-to-day business activities, such as accounting, procurement, project management and manufacturing. ERP systems tie together and define a plethora of business processes and enable the flow of data between them. By collecting an organization’s shared transactional data from multiple sources, ERP systems eliminate data duplication and provide data integrity by using a single accurate source of information (Oracle, 2017). ERP systems are known as planning and control tools at top management level. Today, industrial production is realized with the emergence of global competition and the need to adapt rapidly to ever- changing market demands. These requirements can only be met by radical developments in existing production technology(Erkayman, 2018).

Industry 4.0 is in Germany a common discussion topic in research, academic and industry communities at many different occasions. The main idea is to exploit the potentials of new technologies and concepts such as (Rojko, 2017):

• availability and use of the internet and IoT,

• integration of technical processes and business processes in the companies,

197

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

• digital mapping and virtualization of the real world,

• ‘Smart’ factory including ‘smart’ means of industrial production and ‘smart’ products.

Industry 4.0: New connections between machines, manufacturing processes and systems. Mobility, Big Data, Networking, Human-machine interaction, Dynamic and fast construction constitutes the content of the approach. Industry 4.0 calls for an agile and affordable production future triggered by technology providers such as IoT, 3D printing, cloud computing, mobile devices and big data.

Smart responses to events with proccesses and machines that can talk to each other and ensuring accurate decision-making with triggers and full-integrated systems make ERP's position in companies even more important. Conventional systems take into account fixed processing time and infinite capacity utilization. Companies use a secondary control mechanism.

The machines will be able to make decisions about the product in the production line (faulty, reproduction, control by the operator) and automatically determine the next step, thanks to the sensors and central software on it. In this way, the planned and actual values in manufacturing can be monitored in real time and rapid responses and corrections to disorders can be made automatically. The main goal is more productive manufacturing.

"Real-time Smart Factories" that can be managed by computer and internet, which can be self-managed, controlled and optimized. Communication between the material and the machine will find the most appropriate way to the finished product independently. The deficiencies in resource capacities, substitutes , downtime etc. will be taken into consideration.

Industry 4.0 is a project of encouraging traditional industry for computerization, equipping it with high technology and integrating smart devices into industry by communicating with each other. Processes are redefined in many areas, from order to production, distribution to after-sales services. The Industry 4.0 approach shows the importance of the impact of processes such as production and logistics over industrial operations. The interaction between the physical world of manufacturing and the world of planning will bring about automatic control and visibility in all operations. Enterprise resource planning (ERP) will become even more central for production in this environment. ERP system will become the backbone of the network; It will connect smart machines, logistics systems, production facilities, sensors and devices as products and machines.

2. MES and MOM and their Roles in Industry 4.0

MOM (Manufacturing Operations Management) is a holistic solution that provides full visibility for manufacturing operations, and in this way, we can improve our production performance in a stable manner. A manufacturing operation management system is a structure to improve the overall system of production, inventory, quality management, R & D management, maintenance, advanced planning and programming, and MES (Manufacturing Execution System). MOM is ensured by applications such as Manufacturing Implementation Systems (MES) that digitally transform operational processes and information to ensure both efficiency and transparency.

MES (Manufacturing Execution Systems) is a computer-aided system that is integrated online with all the methods used in production. This structure controls factory-level management and monitoring processes. MES takes all production information from the robots, machines and workers in real time.

198

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

The purpose of a production execution system is to increase productivity and reduce the duration of the production process (Lavi, 2017) .

MES (manufacturing execution system) is a comprehensive system that controls all activities taking place in the shopfloor. It starts with all orders from customers, MRP system, master schedule and other planning resources; and then the products are manufactured as possible as in the most efficient, cost effective, convenient and high quality ("What is MES?," 2018).

MES solutions support two-way implementation of the Industry 4.0 strategy: Horizontal and vertical integration. Horizontal integration: An smart and relational creation of all data in the manufacturing process. MES plays an important role in the collection of comprehensive and multidimensional data. Vertical integration: Ensuring interaction with the data in the entire supply chain. This is only possible with a business software solution (ERP and complementary), which is used to run all processes within the company. The use of IoT technologies in vertical integration becomes a necessity for infrastructure (Trovarit, 2018). MES will meet with other existing and developing production control systems. It is seen that Enterprise Resource Planning (ERP) solutions often have overlapping functionality with MES. MES is more interested in physical operation, while ERP is addressing logical planning for the organization. The ERP by itself does not have a clear knowledge of the individual production processes and capabilities, so an intermediate process is necessary to transform the ERP plan into something that can be performed. With MES, the information received from the ERP in accordance with the needs is used as the basis for creating custom production operations that MES will manage later (Software, 2018) . MES provides information that is more accurate to ERP to improve the capabilities that can be used to plan next operations. Operations in production are managed by MES while the planning process is done logically through ERP. Although production management systems are self-controlling, they are now integrated into Enterprise Resource Planning (ERP) software.

Although IoT data collection and digestion is far from a whole innovation, the combination of these data and the capabilities of ERP to actively develop production processes will help to build useful tool for Industry 4.0.

Today, industrial production is realized with the emergence of global competition and the need to adapt rapidly to ever-changing market demands. These demands can only be met by radical developments in the current manufacturing technology. The technical aspects of these demands are defined by mplementation of Cyber-Physical Systems (CPS) and Internet of Things (IoT) concepts to industrial manufacturing systems.

The need to reconsider changing ERP responsibilities through the context of Industry 4.0, two main concepts are encountered: Data management and IT governance. Data management organizes automated processes and ensures that the information generated is reliable and applicable.

3. ERP Requirements through Industry 4.0

Traditional ERP systems rely on a central database that allows all applications to receive information. This provides a single source of truth to all users and allows all departments and teams to work with the same set of information. Data were processed at regular intervals in these systems; this caused the system not to work in the desired performance. Therefore, new generation ERP systems require a data

199

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY model that makes it possible to access continously up-to-date data and a dynamic structure is desirable. Instant and integrated data flow must be provided vertically (from ERP- MES- PLC or opposite) and horizontally (from supplier to customer) through the system. It is important to respond immediately to any changes in work orders to achieve agility. Data accumulated in the digitalized production environment should be visualized for interpretability and intelligibility. ERP system must be able to collect data from different devices to create new information. Collected data from the various machine tools with the support of receivers should be used effectively and efficiently. The goal is maximizing the value of data. These collected data are enriched with historical data, costs, and demand data to generate information on business values. The main challenge is innovative data management.

The increased flexibility in manufacturing is crucial to enable making production at low cost and with high capacity utilization. ERP systems will still play a central role also in Industry 4.0, but it should be considered how to use it to maximize its expectations for Industry 4.0. ERP systems must establish effective connection with MES and PLC. This makes it easy to dynamically release production orders that aim to produce a wide variety of variants at low costs (Christian A. Hochmuth, 2017). Bidirectional data exchange in vertical integration facilitates the support of ERP system to dynamic structure. Mobile applications are very useful tools for Industry 4.0. Applicable to all business- related processes from any place at any time executable control processes should be created. The ERP system must provide mobile applications for user interaction through smartphones and tablets. Improved communication with suppliers and customers is aimed. Standardization and speeding of operations processes up along the supply chain are provided by this more integrated end to end processes. Some of the dedicated goals of MES is reducing manual operations and minimizing errors during data collection. In the processes where there is a high probability of error in human tracking, machines with user friendly applications are needed more. The applications must support the human activities. If there is any change in any process or data, it is important that all systems that have relationship to this process or data should detect the change as soon as possible. This will also provide a decision support mechanism for planning and control. Achieving process improvement with intelligent data analysis will make the processes more effective. All interrelated processes will be affected in a chained manner. The agility can only be achieved in this way.

4. Conclusions

The use of Industry 4.0 technologies is increasing in many areas, such as factories, supply chains and stores. With the Effects of Industry 4.0, the number of people in the factory and production environments will be reduced in a short time; robot, technology, programs and vehicles are expected to increase. Industry 4.0 has found a wide range of fields from production to agriculture, from medicine to society. The Industry 4.0 concept is not limited only to the manufacturing, it appears to be a harbinger of significant changes in societies. For example, a concept called “Society 5.0: the big societal transformation plan of Japan” is intorduced to the world by Japan.

Data processing technologies, mobile technologies, image processing technologies, data analytics, artificial intelligence and Internet of Things (IoT) are the main components of the new industrial era. Especially IoT has a key importance through Industry 4.0. Human machine interaction will play one of the most important roles in this change. Machines will interact with people to achieve agility in production processes and to minimize errors to support human activities or to manage fully automated processes.

200

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Manufacturing Operations Management (MOM), Manucaturing Execution System (MES), API (Application Programming Interface) and SCADA (Supervisory Control and Data Acquisition) will be the most common concepts for new industrial era. Especially, API is the liberator in overcoming the restrictions on the use of industrial terminals and the use of mobile devices by playing an independent or supportive role in the implementation of the ERP, MES and the Internet of Industrial Objects (IIoT) integration.

References

Christian A. Hochmuth, C. B., Carolin Schwägler. (2017). Industry 4.0

Is your ERP system ready for the digital era? Deloitte. Erkayman, B. (2018). Transition to a JIT production system through ERP implementation: a case from the automotive industry. International Journal of Production Research, 1-11. Lavi, Y. (2017). Industry 4.0: Harnessing the Power of ERP and MES Integration. Retrieved from https://www.industryweek.com/supply-chain-technology/industry-40-harnessing-power-erp-and-mes- integration Oracle. (2017). What Is ERP? Retrieved from https://www.oracle.com/applications/erp/what-is-erp.html Rojko, A. (2017). Industry 4.0 concept: background and overview. International Journal of Interactive Mobile Technologies (iJIM), 11(5), 77-90. Software, A. (2018). What Is “MES”? (Manufacturing Execution Systems). Retrieved from https://www.newcastlesys.com/blog/what-is-mes-manufacturing-execution-systems Trovarit. (2018). Endüstri 4.0 Yolculuğu: ERP vs MES/MOS & API. Retrieved from http://www.trovarit.com/tr/enduestri-4-0-yolculugu-erp-vs-mes-mos-api/ What is MES? (2018).

201

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

The Effect of Different Stimulus on Emotion Estimation with Deep Learning

Süleyman Serhan Narlı1 , Yaşar Daşdemir1, Serdar Yıldırım 2

1İskenderun Technical University, Engineering and Natural Sciences Faculty, Computer Engineering, Hatay-TURKEY [email protected] [email protected] 2Adana Science and Technology University, Engineering Faculty, Computer Engineering, Adana-TURKEY [email protected]

Abstract Emotion recognition system from brain signals is important for human-computer and human-machine interaction. There are many different methods to extract important attributes from these signals. Some of these methods are classification by machine learning, classification by artificial neural networks and classification by deep neural networks. In this study, the effects of these 3 different stimuli (Audio, Video, Audio and Video) on the effect on emotion estimation were investigated. For this study, 1125 samples (375 Audio, 375 Video, 375 Audio-Video) EEG signals were recorded from 14 channels with a sampling frequency of 128 Hz. In this experiment, signals indicating brain activation for 10 different emotions were recorded by considering the emotion evaluations of the participants. The participants were primarily played back emotionally with audio recordings, then the video was played without audio and finally, video and audio were watched together. In the preprocessing step MARA method and Independent component analysis (ICA) were used to eliminate the artifact in the signals. A specific model was created by using deep learning for feature extraction. The effect of different stimuli on performance was investigated for emotion recognition.

Keywords: EEG, Deep Learning, Audio stimuli, Video stimuli, Audio-Video stimuli, Convulutional Neural Network.

Introduction: Emotions are a physiological mechanism that affects our decisions in our daily lives and directs us. For this, it is important to investigate this area, there are many emotion estimation algorithms. There are many studies in this area by using machine learning (artificial intelligence) by processing the signals obtained from the brain. There are limited number of researches and articles that make estimation of emotions with convolution process, artificial neural networks. In addition, it is the data obtained from the outside world which affects emotions. What kind of stimulation is our feelings more affected?

202

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Which stimulant affect our emotions is more effective in predicting emotion when our brain perceives events? In this study, the effects of different stimulus types on emotion recognition will be investigated. Related Work: The success of convolutional neural networks on EEG signals depends on the data set(Lin, Li, and Sun 2017). The success of convective neural networks on EEG signals depends on the data set. Convolutional neural networks are frequently used in image processing (Bhattarai et al. 2017), there is not much study on the processing of EEG signals by this method. To process the EEG signals using the CNN (Convolutional Neural Network) method, it is correct to change the dimensional setting of the data set (Zeng et al. 2018), for this we need to convert the data set to image format and apply a convolution process (Zeng et al. 2018), there is a 3-layer RGB plane in image format and the structure of EEG signals it is similar to this structure on the basis of frequency because the number of time, frequency, and number of channels during the recording of EEG signals can be considered as a 3D plane (Acharya et al. 2018). The way the data set is created in the emotion recognition system is very important, The fact that the emotion stimulus type is an audio, video (Mohammadi, Frounchi, and Amiri 2017) or Audio-Video produces different results. There are many different used data sets(Liu and Sourina n.d.) and the most commonly used is the Deap dataset. In the content of the data set used in this study, the emotions affected by the different types of stimuli were recorded (Dasdemir, Yildirim, and Yildirim 2017) that is, only data, only video and audio-video are recorded together. 1- Dataset description: The data set used in the study was recorded with Emotiv EPOC (Emotiv Systems Inc., San Francisco, USA). 1 EEG signals from 25 volunteers were recorded over 14 channels for 60 seconds, with a total of 15 different clips for different emotion excitations, which were played in 3 different formats; Audio, Video and Audio-Video. As a result, 45 different stimuli were used for each subject. The data were recorded at 128 frequencies.(Table 1) Table 1. List of Stimulus

Stimuli Type Number of Number of Clips Number of Total Records: Volunteers Records 1125 Audio 25 15 375 Video 25 15 375 Audio - Video 25 15 375

Values for the Valence - Arousal - Dominance axes were calculated for the recorded data. In this study, Valence> 0.5 for 1, Valence <0.5 for 0, was used.

2- Data Preprocessing: During recording with the EPOC Emotiv device the raw data may be recorded loudly (blinking, muscle movements, eye movements, etc.)

203

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

With independent component analysis, some of these noises can be destroyed. In this study, MARA (Multiple Artifact Rejection Algorithm), an open source EEGLAB attachment (Winkler et al. 2011), was used for automatic artifact rejection using ICA. MARA is a supervised machine learning algorithm that solves a binary classification problem: It works as "accept or reject" the stand-alone component. 3- Methodology: 3.1. Editing the Data Set: The data set used represents 1125 samples in total (Table 2). Table 2. Sample list Total number Number of Total Sampling Sate Total recorded Raw data length of clips Volunteers Number of time Channels 45 25 14 128 60 1125 x 14 x 7680

For the 1125 samples in the data set; 375 x 14 x 7680 Audio, 375 x 14 x 7680 Video and 375 x 14 x 7680 Audio-Video. The data set was re-arranged for the learning model to be used and the new data set was defined as 3D; 1 x 14 x 120 x 64, thus, the new data set was made ready for processing. 3.2. Convolutional Neural Network (ConvNet): A simple ConvNet is a layer array, and each layer of a ConvNet converts one volume activation to another with a differentiable function. We use three main layer types to create ConvNet architectures: Convex Layer, Pool Layer, and Fully Linked Layer (exactly as seen in Normal Neural Networks). The parameters of the CONV layer consist of a series of learnable filters. Each filter is small (along width and height), but extends across the full depth of the inlet volume. 3.2.1. Convolutional Layer: The Conv layer is the basic building block of a Convolution Network that performs most of the calculation heavy lifting operations. Convolution is the first layer that extracts properties from an input image. Convolution maintains the relationship between pixels by learning image properties using small data input frames. Image matrix is a mathematical process that takes two inputs as filter or kernal. The convolution of an image with different filters can be applied by applying filters such as edge detection, blur and sharpening.

3.2.2. Non-Linear (ReLU) : ReLU stands for Straightened Linear Unit for a nonlinear operation. Output) (x) = max (0, x). ReLU's goal is to introduce nonlinearity in our ConvNet. Because real-world data does not want ConvNet to learn, there will be non-negative linear values.

204

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Other non-linear functions such as tanh, sigmoid or eLU can also be used instead of RELU. The majority of the data scientists use ReLU for their performance better than the other two groups. In this study "elu" was used as activation. 3.2.3. Pool Layer: The pool layers section will reduce the number of parameters when the images are too large. Spatial pooling also called sub-sampling or downsampling, which reduces the dimensionality of each map, but preserves important information. Different types of spatial pooling can be: Max Pooling, Average Pool, Sum Pooling Maximum pooling takes the largest element in the rectified property map. Taking the largest element can also take the average pool. The sum of all items in the property mapping is called ball pooling. 3.2.4. Fully Connected Layer: The layer, which we call the FC layer, flattened our matrix into the vector and fed it to a completely bound layer, such as neural networks. 3.2.5. Normalization process: In order to increase the stability of a neural network, batch normalization normalizes the output of the previous activation layer by subtracting the batch mean and dividing by the standard deviation of the batch. However, after these change / scale activation outputs by some randomly initiated parameters, the weights in the next layer are no longer optimal. SGD (stochastic gradient lowering) eliminates this normalization if there was a way to minimize the loss function. 3.3. Deep Learning Network: A model was developed for the creation of a deep learning network, and this model was run with Google Colaboratory (an open source service provided by Google) using the python (keras) software language.(Figure 1-2)

Figure 1- Schema of the CNN model

205

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 2- Model evalute result 4. The Conclusion:

According to the data used in the study, the performance rate of the Video-type stimulus seems to be better than the others. The results can be improved by changing the parameters of the model and at this point, it is important to show the difference between them by training different stimulus types on the same model, Validation performance is shown in Table 3. According to the results obtained, the audio stimuli performance curve and other types are shown below: (Figure 3-4-5) Audio Stimuli Rate :

Figure 3- Audio stimuli Accuracy

206

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Video Stimuli Rate :

Figure 4- Video stimuli accuracy

Audio – Video Stimuli Rate :

Figure 5- Audio-Video stimuli accuracy

Table 3 – Comparing accuracies between stimulus Stimuli Train Accuracy Validation Accuracy Audio %82,83 %65.30 Video %83,87 %58,28 Audio - Video %78 %59,44

207

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

References

Acharya, U Rajendra et al. 2018. “Deep Convolutional Neural Network for the Automated Detection and Diagnosis of Seizure Using EEG Signals.” Computers in Biology and Medicine 100(September 2017): 270–78. https://doi.org/10.1016/j.compbiomed.2017.09.017. Bhattarai, Smrity et al. 2017. “Digital Architecture for Real-Time CNN-Based Face Detection for Video Processing.” : 1–26. Dasdemir, Yasar, Esen Yildirim, and Serdar Yildirim. 2017. “Analysis of Functional Brain Connections for Positive–negative Emotions Using Phase Locking Value.” Cognitive Neurodynamics 11(6): 487–500. Lin, Wenqian, Chao Li, and Shouqian Sun. 2017. “Image and Graphics.” 10667(January 2018). http://link.springer.com/10.1007/978-3-319-71589-6. Liu, Yisi, and Olga Sourina. “EEG Databases for Emotion Recognition.” Mohammadi, Zeynab, Javad Frounchi, and Mahmood Amiri. 2017. “Wavelet-Based Emotion Recognition System Using EEG Signal.” Neural Computing and Applications 28(8): 1985–90. Zeng, Hong et al. 2018. “RESEARCH ARTICLE EEG Classification of Driver Mental States by Deep Learning.” Cognitive Neurodynamics 12(6): 597–606. https://doi.org/10.1007/s11571-018-9496-y.

208

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

The Investigation of The Wind Energy Potential of The Belen Region and The Comparison of The Wind Turbine with The Production Values Fatih Peker1, Cuma Karakuş1, İlker Mert2

1Department of Mechanical Engineering, Iskenderun Technical University, Turkey [email protected] [email protected] 2Osmaniye Vocational School, Osmaniye Korkut Ata University, Turkey [email protected]

Abstract

In this study, the potential of wind energy has been investigated in Belen region of Hatay province between 2013-2016. As a result of the study, it was aimed to compare the real field conditions with the predicted values and to enlighten the error analysis of the pre-feasibility reports of the investors who will invest in the region. In the research area, the annual production values are based on a known reference wind turbine. This wind turbine, which is already installed, has been analyzed with computer aided software considering environmental factors. Wind speed, temperature and pressure data were obtained from Belen Meteorology station, which is very close to the area where the turbine is located. The topographical data of the turbine and meteorological station were evaluated by using the WaSP (Wind Atlas Analysis and Application) program using the vector elevation maps of Hatay region. A wind atlas map of the region was created with the WaSP program. Considering the classification requirements of the European Wind Energy Association, it was evaluated that, Belen region could be included in the classes rated as good and very good.

Keywords: WAsP, Belen, Wind, Energy, Weibull Distribution

Symbols A : Area [m2] c : Weibull scale parameter [m/s] fW(v) : Weibull distribution function Γ() : Gamma function k : Weibull shape parameter RES : Wind power plant WAsP : Wind Atlas Analysis and Application NASA :National Aeronautics and Space Administration SRTM : Shuttle Radar Topography Mission

PW : Average power density for the Weibull distribution [W/m2] Mevbis : Meteorological data information sales and presentation system P(v) : Average wind power potential [W] R : Correlation coefficient ρ : Air density [kg/m3] σ : Standard deviation [m/s] v : Wind speed [m/s] vm : Average wind speed [m/s]

209

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

1. Introduction

Increasing population in the world, developments in technology enhance the need of energy day by day. The current depletion of fossil fuels used in the near future, price fluctuations, carbon emissions and transport problems have led countries to different sources of energy. As an alternative to fossil fuels, renewable energy sources are of great importance (1). Among renewable energy sources, wind energy is considered to be a good alternative to fossil fuels. The number of wind power plants in the world is increasing in recent years. In 2017, 52.6 GW wind power plants were established and total installed capacity reached 539 GW by January 2018. As of June 2018, the installed wind power plant in Europe is 182 GW. The 165 GW of these are onshore power plant and 16.9 GW are offshore power plant. As of January 2018, the total installed capacity is 6.98 GW in Turkey. In terms of installed capacity Turkey is the 12th largest country in the World and 6th in Europe (2-3).

Due to enhanced energy demand, Turkey's high wind energy potential and investment are increasing in importance. Energy imports, which is one of the biggest obstacles to the development of the country, increases the dependency on foreigners and becomes one of the most important items of the current account deficit. Turkey is trying to supply its energy needs with own resources and investments made in recent years. Turkey plans to invest about 3.3 GW of wind power by 2022 (3). With these investments, it is planned to reduce the emission of fossil fuels and to reduce the carbon emissions and thus to increase the employment. In 2015, with the energy generated from wind turbines, natural gas imports decreased by $ 574 million and carbon emissions decreased by 5.88 million tons. In addition, approximately 15,000 people were employed in these plants (4). Hatay has an important position among the cities in Turkey with high wind energy potential and the 4th largest city with 364.50 MW of installed capacity of wind power plant (WPP) in Turkey. With the completion of the ongoing wind energy sources MWs construction, the existing installed power will be 410 MW, while 144.3 MW of this installed power is in Belen region (5).

Wind energy potentials in Belen and other regions of Turkey are reported by Mert et al. (6-7), Bilgili et al. (8-9), Tanç et al.(10).

When investments are completed, the area where the WPP will be installed is of great importance. Most of the costs of the WPP are paid during the installation phase. However, the climatic and topographic structures of the area where the WPP will be installed affect the repayment period of the investment costs. False investments negatively affect repayment periods and therefore profitability (4). In this study, four-year wind data were studied between 2013-2016 by using WAsP program. A detailed wind atlas of 2016 was prepared. The results are compared with the net energy produced by a wind turbine in the region for four years.

2. Material and Method

2.1. Wind Atlas Analysis and Application Program (WAsP)

The Wind Atlas Analysis and Application Program (WAsP), developed by the Danish Technical University (DTU) Risø National Laboratory, allows for many analyzes such as wind turbine evaluation, wind turbines and location detection for wind power stations, and energy efficiency at these locations. The program is able to perform energy efficiency analysis for both on-shore and off-shore WPP. In the program database, wind turbine models produced by the world's leading wind turbine manufacturers and the technical information of these models contain. The WAsp program needs the following data when modelling (11-12).

- Wind speed and direction - Region location information and vector map

210

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

- Area roughness - Near obstacles

In WAsP, point analysis of a single turbine can be performed or analysis can be done for large WPP. Gross and net electricity generated by turbines, power density of the turbine in the rotor center, power density maps of the region, average wind speed at different heights (10m - 25m - 50m - 100m - 200m), frequency of wind formation, prevailing wind directions many sectors such as speed, energy and power density can be obtained by dividing into sectors. The Wasp program performs wind frequency modelling in accordance with the two-parameter Weibull distribution. General expression of two-parameter Weibull distribution for wind speed,

푘 푣 푘−1 푣 푘 푓 (푣) = ( ) ( ) exp [− ( ) ] (Eq. 1) 푤 푐 푐 푐 can be given with. Where fw (v) is the probability function of the measured wind speed v. k and c are Weibull parameters. k is the shape parameter and c is the scale parameter. (1-13)

The wind power potential of a wind turbine with a wing sweep area A at v speed,

1 푃(푣) = 휌퐴푣3 (Eq. 2) 2 can be given as. Here, ρ is the density of the air (kg / m³). The WAsP program calculates by varying the density of air at different temperature and heights. The average power density for the Weibull distribution is calculated as follows.

1 3 푃(푣) = 휌푐3Γ (1 + ) (Eq. 3) 2 푘

2.2. Data

2.2.1. Wind Speed and Direction

The wind speed data used in this study was obtained from meteorological data information sales and presentation system (MEVBIS) of the General Directorate of Meteorology. The data are recorded on an hourly basis by the meteorological station located in the district of Belen, Hatay, between 2013-2016.

2.2.2. Region Location Information and Vector Map

The map of Hatay used in the study is the SRTM (Shuttle Radar Topography Mission) map created by NASA (National Aeronautics and Space Administration). Map of Hatay region are created using Global Mapper program and digitized so that WAsP program can be read. Map of Hatay Region is given in Figure 1. The altitude of the measuring station is 603 meters and the measuring pole is 10 m above the ground level (14). Location Information of Belen Meteorology Station is given in Table 1. The turbine center height of the wind turbine, located at an altitude of 724 m, is 80 m from the ground. Reference Wind Turbine Position is given in Table 2.

Table 1. Belen Meteorology Station Location Information (14) Station Nr Station Latitude Longitude Altitude Anemometer Height 18057 Belen 36.4889° 36.2172° 603 m 10 m

211

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 1. Map of Hatay Region

Table 2. Reference Wind Turbine Position Turbine Central Turbine Latitude Longitude Altitude Wing Turbine Height Nr (°) (°) (m) Diameter (m) (m) T01 Vestas V90 36.4781 36.2061 724 80 90

2.2.3. Roughness

Each region has a roughness value. The map is divided into areas according to these roughness values defined by z0. The roughness length gives information about the terrain structure, vegetation and building structures of the region. Roughness values are expressed in meters. The roughness value of the water is 0.0 m. A small value of roughness means a clear and flat land (snow, sand, bare earth, etc.). The roughness lengths of the lands are given in Table 3. These roughness values can also be taken from pre- digitized maps (CORINE, ESA CCI, etc.). The WAsP program cannot read these maps in raw form. Therefore, with the help of Global Mapper and similar programs, the fields and roughness values in the maps are digitized and processed on the elevation map so that the WAsP program can read (15-16).

Table 3. Roughness Length (15)

Roughness Type Length (m) City 1 Forest 0.8 Open Fields 0.01 Sea 0.0001

212

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

2.2.4. Obstacle

Anemometer, turbine or any obstacles around the point to be taken on the area and to be analyzed are processed into the WAsP program as it will change the wind flow.

2.3. Wind turbine

The turbine used as a reference for the theoretical study is the Vestas V90-3.0 MW VCS 50 Hz model. The rotor diameter of the turbine is 90 m and the hub height is 80 m. Reference turbine characteristics are given in Table 4. Vestas V90 the 3 MW VCS 50 Hz wind turbine has a minimum operating speed of 4 m/s and a maximum operating speed of 25 m/s. The power curve graph of the turbine is given in Figure 2.

Table 4. Reference Turbine Features (17)

General Model Vestas V90 3MW Capasity 3,000 kW Hub Height 80 m Working Speeds Minimum Speed 4 m/s Maksimum Speed 25 m/s Nominal Speed 15 m/s Rotor Diameter 90 m Sweeping Area 6,362 m2 Number of Wings 3 Wing Length 44 m Speed range 8.6-18.4 rpm Rated Speed 16.1 rpm

VESTAS V90 3 MW VCS 50 Hz 3000

2500

2000

1500 P P [kW] 1000

500

0 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 v (m/s)

Figure 2. Vestas V90 3 MW VCS 50 Hz wind turbine power curve graph

213

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

3. Results and Discussion

In the study, wind data from Belen Meteorology station were analyzed by the WAsP program as hourly, daily, monthly and yearly. The reference temperature of the station was 17.4 ° C, the air density was 1.129 kg / m³, the average pressure was 96,409 Pa and the relative humidity was 72.02%. Monthly average wind speeds according to the years measured in the station are given in Figure 3. The highest wind speed average was 3.93 m/s in July 2014 and the lowest wind speed average was 1.53 m/s in October 2015.

Belen Meteorology Station Monthly Average Wind Speeds 4,50 4,00 3,50

3,00 (m/s)

m 2,50 2,00 1,50 Speed, v Speed, 1,00 0,50 0,00 1 2 3 4 5 6 7 8 9 10 11 12 2013 2,85 2,61 2,56 2,01 2,20 3,43 3,68 3,31 2,42 1,67 2,09 2,94 2014 2,33 2,00 2,29 2,28 2,40 2,78 3,91 3,49 2,70 1,72 2,17 2,82 2015 2,55 2,51 2,36 2,16 2,35 3,32 3,39 3,24 2,19 1,53 2,47 2,40 2016 2,60 1,99 2,46 2,05 2,36 3,30 3,28 3,13 2,61 2,09 2,98 2,68 Months

2013 2014 2015 2016

Figure 3. Monthly wind speed averages over Belen region.

The wind speed and direction are measured by the wind climate analysis (Wasp Climate Analyst) program at 10 m were calculated as raw meteorology data. The wind direction is divided into 30° intervals and is composed of 12 sectors between 0° and 360° The program calculates the prevailing wind direction and frequency (f) separately for each year. The average speed and frequency of frequencies calculated in 2016 are given in Table 5. The dominant wind direction is the sector 11 and it is in the North-West direction. The wind-blown directions and the frequency-wind speed graph are given in Figure 4.

Table 5. Belen meteorology station average wind speed and frequency by sectors (Year 2016)

Sector 1 2 3 4 5 6 7 8 9 10 11 12

F (%) 0.80 0.20 0.80 6.50 18.20 13.70 1.00 1.00 2.20 8.10 27.00 20.40

U [m/s] 1.97 0.77 1.60 2.69 2.12 3.08 0.86 1.04 1.55 2.24 2.70 2.99

214

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 4. Belen meteorology station wind blow direction and wind speed frequency chart (2016 year)

The average wind speeds, frequencies, Weibull parameters, power densities of the four-year (2013- 2016) and sectors of the reference turbine are given in Table 6. Frequency value is the highest in the 12th sector with 20% when the working period is taken into consideration. According to the frequency of all sectors, the highest wind speed average was 8.12 m/s in 2013 and the lowest was 7.93 m/s in 2016. Average power density values were 517 W/m² for 2016 and 544 W/m² for 2013. In the classification made by the European Wind Energy Association, the energy potential of turbine hub heights is close to 100-300 W/m², good between 300-700 W/m² and very good if it is higher than 700 W/m² (1). The power density at the reference turbine position is in the range of 500-550 W/m² and the energy potential is considered to be good.

Table 6 . WAsP Annual wind characteristic analysis by sector a) Year 2013 (80 m) Sector Wind Statistics 2013 P (ρ=1,109) Angle Weibull- c Speed Power Nr Frequency [%] Weibull-K [°] [m/s] [m/s] [W/m²] 1 0 11.50 14.80 4.11 13.39 1,629 2 30 1.50 7.90 1.24 7.33 818 3 60 4.00 9.30 2.21 8.20 532 4 90 5.50 6.90 1.96 6.09 244 5 120 8.90 5.90 1.62 5.29 200 6 150 13.60 8.40 1.96 7.42 441 7 180 8.80 11.10 2.23 9.79 899 8 210 2.40 5.40 2.15 4.74 105 9 240 4.10 6.10 3.30 5.49 123 10 270 6.50 6.30 3.86 5.69 128 11 300 13.20 7.60 4.57 6.90 215 12 330 20.00 10.20 3.93 9.24 543 All sectors 100.00 8.12 544

215

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

b) Year 2014 (80 m) Sector Wind Statistics 2014 P (ρ=1,109) Angle Frequency Weibull-c Speed Power Nr Weibull-k [°] [%] [m/s] [m/s] [W/m²] 1 0 12.60 15.00 3.76 13.56 1,748 2 30 1.30 6.60 1.31 6.07 421 3 60 3.60 7.40 1.99 6.57 301 4 90 5.30 6.00 2.06 5.35 158 5 120 9.00 6.30 1.99 5.58 185 6 150 12.60 8.30 2.27 7.33 372 7 180 7.70 10.50 2.47 9.33 715 8 210 2.80 5.70 2.00 5.07 138 9 240 4.90 6.20 3.21 5.59 131 10 270 6.90 6.10 3.76 5.51 117 11 300 12.80 7.20 4.21 6.58 192 12 330 20.50 10.30 3.45 9.22 569 All Sectors 100.00 7.98 523 c) Year 2015 (80 m) P Sector Wind Statistics 2015 (ρ=1.109) Angle Frequency Weibull-c Power Nr Weibull-k Speed [m/s] [°] [%] [m/s] [W/m²] 1 0 13.70 14.60 3.83 13.23 1,613 2 30 1.40 6.40 1.80 5.69 218 3 60 3.60 7.20 2.12 6.35 256 4 90 5.40 5.90 1.99 5.21 151 5 120 9.40 6.00 1.71 5.36 194 6 150 13.50 8.40 2.01 7.43 431 7 180 8.50 11.00 2.27 9.72 868 8 210 2.80 5.90 2.10 5.25 146 9 240 5.00 6.60 3.59 5.96 151 10 270 6.40 6.10 4.01 5.53 115 11 300 10.80 7.00 4.10 6.37 176 12 330 19.50 10.10 3.39 9.06 545 All Sectors 100.00 7.99 536

d) Year 2016 (80 m) P Sector Wind Statistics 2016 (ρ=1.109) Angle Frequency Weibull-c Power Nr Weibull-k Speed [m/s] [°] [%] [m/s] [W/m²] 1 0 11.80 13.90 3.69 12.55 1,394 2 30 1.20 6.90 1.33 6.32 462 3 60 3.20 7.50 1.87 6.66 337 4 90 5.00 5.90 1.79 5.26 174 5 120 9.10 6.00 1.63 5.34 204 6 150 13.60 9.20 2.08 8.14 550 7 180 8.60 12.40 2.42 11.01 1.198 8 210 2.60 6.40 2.76 5.69 150 9 240 4.50 6.50 3.47 5.87 147 10 270 6.80 6.10 4.14 5.51 113 11 300 13.30 7.00 4.43 6.35 169 12 330 20.30 9.40 3.47 8.49 444 All sectors 100.00 7.93 517

216

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

In the Wasp program, wind atlas of each year were created separately. With this wind atlas, calculations can be made for the desired height and different surface roughness values. Table 7 shows the wind atlas data for the analysis region in 2016.

Table 7. Wind Atlas of 2016 2016 Height [m] 0.00 m 0.03 m 0.10 m 0.40 m 1.50 m Weibull-c[m/s] 6.52 4.71 4.13 3.26 2.18 10.0 Weibull-k 1.97 1.80 1.83 1.87 1.92 Wind [m/s] 5.78 4.19 3.67 2.90 1.94 Weibull-c[m/s] 7.15 5.64 5.09 4.29 3.30 25.0 Weibull-k 2.02 1.90 1.93 1.96 2.00 Wind [m/s] 6.34 5.00 4.52 3.81 2.93 Weibull-c[m/s] 7.70 6.51 5.96 5.18 4.24 50.0 Weibull-k 2.10 2.06 2.07 2.09 2.12 Wind [m/s] 6.82 5.77 5.28 4.59 3.75 Weibull-c[m/s] 8.35 7.67 7.07 6.25 5.31 100.0 Weibull-k 2.13 2.25 2.25 2.24 2.22 Wind [m/s] 7.39 6.80 6.26 5.54 4.70 Weibull-c[m/s] 9.16 9.32 8.58 7.65 6.63 200.0 Weibull-k 2.08 2.23 2.22 2.20 2.17 Wind [m/s] 8.11 8.25 7.60 6.77 5.87

After the wind atlas has been formed, the energy potential map of the region and the average wind speed map are extracted in the light of this information. These maps make it easier to decide which locations turbines can be placed on. Calculations are performed at the height of the selected wind turbine. In the study, the hub height of the reference turbine is 80 m. The average wind speed atlas of the Belen region in 2016 is given in Figure 5 and the wind energy potential atlas is given in Figure 6. The highest power density in the analyzed region is calculated as 1092 W/m² and the lowest power density is 126 W/m². Average wind speeds vary between 5.09 m/s and 10,12 m/s.

Figure 5. Wind speed atlas of Belen region in 2016 (80 m)

217

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 6. The wind energy potential atlas of Belen region in 2016 (80 m)

Wind power plants record unintentional stops (maintenance, malfunction, periodic checks, etc.). The ratio of the total time of these stops to the whole working time is called at involuntary waiting coefficient (IWC). The energy is produced yearly by the subject turbine that is divided into the IWC so this way prevents the effect of the unintentional stops on the theoretical study. The actual values of the turbine are divided into these values and are calculated by assuming that there is no involuntary stand. The region of the subject turbine was analyzed and the total annual energy produced in the turbine was calculated for four years. The actual energy produced in the turbine and the theoretical value calculated by the WAsP program were compared. The theoretical energy production values of the subject turbines calculated with WAsP program and the actual energy production values are given in Figure 7. The annual wind energy production values calculated in the WAsP program are considered to be very close to the actual values. The deviation between the real and estimated values was realized in 2013 by 7.11%. The closest value to real production was in 2016 with deviation of 0.82%. The differences between the values are given in Table 8.

218

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Annual Energy Produced in Turbine 8.750.000 8.500.000 8.250.000 8.000.000 7.750.000 7.500.000 7.250.000 P P (kWh) 7.000.000 6.750.000 6.500.000 6.250.000 6.000.000 2013 2014 2015 2016 T01 WAsP (kWh) 8.375.000 8.028.000 8.175.000 7.909.000 T01 Actual (kWh) 7.818.885 8.240.323 7.811.206 7.844.631 Years

T01 WAsP (kWh) T01 Actual (kWh)

Figure 7. Actual energy production values of the subject turbines and theoretical energy production values calculated with the help of WAsP program

Table 8. Comparison of actual power generation values and WAsP theoretical power generation values T01 WAsP T01 Actual Difference Difference T01 (kWh) (kWh) (kWh) (%) 2013 8,375,000 7,818,885 -556,115 -7.11% 2014 8,028,000 8,240,323 212,323 2.58% 2015 8,175,000 7,811,206 -363,794 -4.66% 2016 7,909,000 7,844,631 -64,369 -0.82%

4. Conclusions

In this study, wind energy potential of Belen in district of Hatay was investigated between 2013-2016. WAsP program and Vestas V90-3.0 MW VCS 50 Hz model values were used for the analysis. As a result of the calculations, the difference between the reference turbine and the theoretical study was the highest deviation in 2013 with a rate of 7.11%. The closest result to real production was obtained in 2016 with a deviation of 0.82%. However, according to the classification conditions of the European Wind Energy Association, it is assumed that the Belen region has good - very good rated areas.

References 1.Bilgili, M., Şahin, B., & Şimşek, E. (2010). Turkey, South, Southwest and Wind Energy Potential in the Western Region. Journal of Heat Science and Technology, 30 (1), 1-12.

2.World Wind Energy Association 2017 Statistic https://wwindea.org/blog/2018/02/12/2017-statistics/

3.Wind energy in Europe: Outlook to 2022. https://windeurope.org/members-area/files/protected/market- intelligence/reports/Wind-energy-in-Europe-Outlook-to-2022.pdf.

4.https://www.tureb.com.tr/files/yayinlar/tureb_ruzgar_enerjisi_ve_etkilesim_raporu.pdf

5.https://www.tureb.com.tr/files/tureb_sayfa/duyurular/2018/03/turkiye_ruzgar_enerjisi_istatistik_raporu_ocak_ 2018.pdf

219

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

6.Mert, İ., Karakuş, C., & Üneş, F. (2016). Estimating the energy production of the wind turbine using artificial neural network. Neural Computing and Applications, 27(5), 1231-1244.

7.Mert, İ., Karakuş, C., & Peker, F. (2014). Antakya bölgesi rüzgar karakteristiğinin incelenmesi. DÜ Mühendislik Dergisi, 5(1), 13-22.

8.Bilgili, M., Simsek, E., Sahin, B., Yasar, A., & Ozbek, A. (2015). Estimation of human heat loss in five Mediterranean regions. Physiology & behavior, 149, 61-68.

9.Bilgili, M., Şahin, B., & Kahraman, A. (2004). Wind energy potential in Antakya and Iskenderun regions, Turkey. Renewable Energy, 29(10), 1733-1745.

10.Tanç B., Mert, İ., Arat, H. T., Karakuş, C., & Baltacıoğlu, E. (2014). Estimation of wind energy potential using WAsP in Hatay Airport region. International Journal, 2(2), 2311-2484.

11.http://www.wasp.dk/wasp

12.https://www.mgm.gov.tr/genel/ruzgar-atlasi.aspx

13.Usta İ., Kantar Y.M., Estimation of wind power potential using different probability density functions, Dokuz Eylul University Faculty of Engineering Science And Engineering Journal 2016; 18(3): 362-380.

14.MEVBIS Data order information No. 201710025E41 of the General Directorate of Meteorology.

15.Pusat, Ş. (2017). A study of wind energy potential for Sakarya University. Pamukkale University Journal of Engineering Sciences, 23 (4), 352-357.

16.Mortensen, Gylling N., Wind resources assessment using the WAsP software (DTU Wind Energy E-0135)

17.Vestas V90 3MW 50 Hz Product Brochure. http://arquivo.pt/wayback/20090706121351/ http://www.vestas.com/Admin/Public/DWSDownload.aspx?File=%2fFiles%2fFiler%2fEN%2fBrochur es%2fProductbrochureV90_3_0_UK.pdf

220

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

The Effects of Sodium Nitrite on Corrosion Resistance Ofsteel Reinforcement in Concreta Güray Kılınççeker, Nida Yeşilyurt

Department of Chemistry, Faculty of Sciences and Letters, Çukurova University,

[email protected]

Abstract This study describes a laboratory investigation of the influence of 0.1 M nitrite ions on the corrosion of reinforcing steel and on the compressive strength of concrete. The effect of 0.1 M nitrite ions on the corrosion resistance of steel reinforced concrete was evaluated by electrochemical tests in distilled water, 3.5% NaCl and 3.5% NaCl + 0.1 M NaNO2 solutions for 90 days. In the presence of 0.1 M nitrite ions polarization resistance (Rp) values of reinforced concrete were higher than those without sodium nitrite. AC impedance spectra revealed the similar results with Rp measurements. The compressive strength of concrete specimens containing 0.1 M nitrite ions was measured and an increase of 15.2–28.6% was observed.

Keywords: Corrosion, Reinforcing steel, Concrete, Nitrite ions, Inhibition.

1. Introduction Corrosion is the deterioration of a metal, its alloys or steel rebars in concrete by reaction with the environment. Corrosion occurs by the loss of alkalinity of concrete in the form of carbonates and thus causes cracks in the concrete that provides a direct route for chlorides to approach the reinforcing steel and prevent re-passivation reaction that leads to pitting corrosion. 0.1 M nitrite ions as a anorganic corrosion inhibitor for metal. There are different views regarding the effect of 0.1 M nitrite ions on the corrosion of various metal. In general, the nitrite ion has no effect on the corrosion of ferrous materials in neutral or close to neutral condition. As Nitrite ions, form a protective barrier; have been frequently studied for prevention of corrosion of various metals, under various conditions. Especially for iron based materials, 0.1 M nitrite ions act as anodic inhibitor and that this ion continues to exhibit this behavior in environments which contains other aggressive ions [1-6]. The objective of this paper is to study the effectiveness of 0.1 M nitrite ions, in providing corrosion protection to reinforcing steel in 3.5% NaCl solution. The effect of the inhibitor on compressive strength of concrete was also determined. 2. Materials and Methods

The mild steel samples were used as a working electrode. The electrochemical measurement were carried out in a conventional three electrode system where the auxiliary electrode was a platinum sheet and Ag/AgCl electrode was used as the reference electrode. All the potentials given in this paper are referred to this electrode. The EIS measurements were obtained at open circuit potential in a frequency range of 105 Hz-10-3 Hz by amplitude of 5mV. The polarization curves were recorded with a scan rate of 4 mV/s after 90 days of immersion time.

221

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 1. Working electrode (test electrode)

3. Results and discussion

On the average, the Ecor of the rebar in specimen B was ~0.100 V more negative than that of rebar in specimen A as is shown in Fig. 2. This indicated that the rebar embedded in the concrete containing 0.1 M nitrite ions(specimen C) had become more positive than the rebar in concrete without 0.1 M nitrite ions (specimen A, B). In literature, it is reported that probability of corrosion would be greater than 95% if the observed corrosion potential is more negative than -0.270 V/SHE and corrosion potential falls well below -0.320 V/SHE, indicating the initiation of corrosion [1]. Hence, while the addition of 0.1 M nitrite ions in concrete has reduced the corrosion rate decrease the distilled water as external solution. On the other hand, the addition of 0.1 M nitrite ions in chloride solution could be effective on corrosion - resistance of rebar in concrete. Specimen C that contained NO2 ions than specimen B were found to be less corrosive just as expected. It was shown that, nitrite ions could inhibit anodic dissolution of reinforcing steel in aggressive chloride media [2-4].

The effects of 0.1 M nitrite ions addition on the corrosion behavior of rebar in the concrete were monitored and the results are shown in Fig. 2 and Table 1. For the specimens in all media have varied from -0.154 to -0.929 V (versus Ag/AgCl) for a period of 90 days (Fig. 3). Values in the concrete specimen without corrosion inhibitor (specimen A and B) was more negative than the corrosion potentials of the steel in the concrete specimen prepared with 0.1 M nitrite ions (specimen C).

The inhibition effect of nitrite ions was resulted from complex formations (between nitrite ions and corrosion products) which physically adsorbed on the surface.

The inhibition effects decreased with the chloride ions. Nitrite presence at stable potentials lowers the flow density, because nitrite presence lowers the speed of diffusion oxygen molecules which causes the cathodic reaction happened on surface [5, 6].

Table 1. The electrochemical data determined from polarization and impedance measurements for reinforcing steel at pH 11.5.

222

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

E E I-1.400 V I-0.250 V Time corr I =0 R Solution (V vs. Ag/AgCl) P (.)mAcm−2 (.)mAcm−2 (Day) ()ohm I%

7 -0.506 - 3981.1 - - 14 -0.500 - 4305.3 - - - Distilled 21 -0.517 - 4008.7 - Water 28 -0.507 - 4477.1 - - 60 -0.518 - 4698.9 - - 90 -0.578 -1.0469 4325.1 2.98×10-4 2.91×10-4 58.50 7 -0.928 - 1258.9 - - 14 -0.929 - 1374.0 - - 21 -0.926 - 1402.8 - - 3.5%NaCl 28 -0.902 - 1603.3 - - 60 -0.850 - 1527.6 - - 90 -0.803 -0.8853 1794.7 2.15×10-3 1.14×10-3 - 7 -0.401 - 9794.9 - - 14 -0.397 - 9484.2 - - 3.5%NaCl 21 -0.395 - 11614.5 - - + - 0.1 M 28 -0.383 - 12531.4 - NaNO2 60 -0.316 - 13583.1 - -

90 -0.154 -0.8955 12882.5 1.24×10-3 1.81×10-4 86.07

223

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 2. The corrosion potential versus reciprocal time for reinforcing steel in distilled water (A), 3.5% NaCl (B), 3.5% NaCl + 0.1 M NaNO2 (C).

(a) (b)

Figure 3. The Nyquist of reinforcing steel in 3.5% NaCl + 0.1 M NaNO2 (a) and 3.5% NaCl (b) solution at pH=11.5 (90th day).

224

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

4. Conclusions

The effect of 0.1 M nitrite ions on the corrosion of steel embedded in concrete was examined and the addition of 0.1 M nitrite ions to the mixing or external solution was found to have increased the Rct + Rs + Rd resistance and have minimized the corrosion of the rebar in concrete. Meanwhile, the increase in the strength varied from 20.5 MPa to 28.6 MPa in the concrete specimens admixed with 0.1 M nitrite ions together with chloride ion. This study shows that 0.1 M nitrite ions did not adversely affect the compressive strength of concrete.

Acknowledgement

The authors are greatly thankfull to Çukurova University, Art and Science Faculty, Chemistry Department for support.

References

[1] G. Kılınççeker, C. Menekşe, The effect of acetate ions on the corrosion of reinforcing steel in chloride environments, Protect. Met. Phys. Chem. Surf. 51/4, (2015), 659-666.

[2] G. Kılınççeker, M. Erbil, The effect of phosphate ions on the electrochemical behaviour of brass in sulphate solutions, Mater. Chem. Phys.119 (2010) 30-39

[3] G. Kılınççeker, H. Galip, Electrochemical Behaviour of Zinc in Chloride and Acetate Solutions, Protect. Met. Phys. Chem. Surf. 45 (2009) 232-240.

[4] G. Kılınççeker, The effects of 0.1 M acetate ionson electrochemical behaviour of brass in chloride solutions, Colloids Surf. A: Phys. Eng. Asp. 329 (2008) 112-118.

[5] T. Doğan, G. Kılınççeker, The effects of glucose, maltose and starch on electrochemical behaviour of copper in chloride solutions, Corrosion (TR). 14 (2006).

[6] H. Kahyaoğlu, M. Erbil, B. Yazıcı, A.B. Yılmaz, Corrosion of reinforcing steel in concrete immersed in chloride solution and the effects of detergent additions on diffusion and concrete porosity. Turk J Chem. 26 (2002).

225

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

The Evaluation and Comparison of Daily Reference Evapotranspiration with ANN and Empirical Methods Fatih Üneş1, Süreyya Doğan2, Bestami Taşar3 Yunus Ziya Kaya4, Mustafa Demirci5

1,2,3,5 Department of Civil Engineering, Iskenderun Technical University, Turkey [email protected] [email protected] [email protected] [email protected] 4Department of Civil Engineering, Osmaniye Korkut Ata University, Turkey [email protected] Abstract Evapotranspiration is an important parameter in hydrological and meteorological studies, and accurate estimation of evaporation is important for various purposes such as the development and management of water resources. In this study, daily reference evapotranspiration (ET0) is calculated by using Penman-Monteith equation, which is accepted as standard equation by FAO (Food and Agriculture Organization). ET0 is tried to be estimated by using Hargreaves-Samani and Turc traditional equations and results are compared with Artificial Neural Network (ANN) model performance. A station which is stated near to the Hartwell Lake (South Carolina, USA) was chosen as the study area. Average daily air temperature (T), highest (Tmax) and lowest daily air temperatures (Tmin), wind speed (U), solar radiation (SR) and relative humidity (RH) were used for daily average evapotranspiration estimation. Feed forward-back-propagation ANN method is used for model creation. Comparison between empirical equations and ANN model shows that ANN model performance for daily ET0 estimation is better than others.

Keywords: Evapotranspiration, Penman-Monteith equation, Hargreaves-Samani equation, Turc equation, Artificial Neural Network 1. Introduction Evapotranspiration is defined as the return of liquid or solid water as gaseous to the atmosphere by the effect of meteorological factors. The majority of the rain falling on the earth returns to the atmosphere by evaporation and transpiration before direct runoff. Determination of these losses is of great importance especially in the dry season. ET0 is an important component of the hydrological cycle and knowledge of water losses due to ET0 under the hydrological cycle is an important consideration in terms of water management and planning. Accurate estimation of ET0 is required for good planning and management of water resources. It is also important to determine the ET0 in many hydrological problems such as water budget and irrigation. Direct or indirect methods may be used to determine ET0 on free water surfaces. As the direct method, the most widely used method in the world is the evaporation pan. Indirect methods, in terms of complexity and data requirements, are temperature-based formulas (Thornthwaite, 1948); radiation-based approaches (Turc, 1961); formulas based on humidity (Romanenko, 1961); formulas that combine temperature, humidity and wind speed (Penman, 1948). These and similar methods have been used and estimated by many researchers for ET0

226

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY estimation (Gümüş,V., et. all; Warnaka and Pochop, 1988; Choudhury, 1999; Abtew, 1996; McKenzie and Craig, 2001). Generally, these formulas also contradict each other and therefore it is very difficult to determine the best solution. The complexity, and uncertainty of the problem do not allow the casual modeling of classical methods. There are other methods that can be used more appropriately in these situations. In recent years, many researchers have used artificial intelligence methods as an alternative to classical methods in hydrology and water resources studies. (Üneş et al., 2013, 2015, 2017; Demirci et al., 2015,2016; Taşar et al., 2017) Nowadays, artificial neural networks (ANN) are being used widely because they can easily resolve complex and difficult relationships. ANN is applied to many fields of science. This approach is also used in hydraulics and hydrology as other fields of science to achieve good results. In the past years, researchers have estimated the use of artificial intelligence methods in predicting hydrological events such as evaporation or ET0 (Aytek et al., 2008; Fenga et al., 2016; Partal 2016). Doğan et al. (2007) was estimated the daily evaporation amount for Sapanca Lake by using feed-forward back-propagation (FFBP) and radial-based artificial neural network (RBNN) model and compared with the Penman- Monteith (PM) model. Kaya et al., (2016), studied to estimate the amount of evaporation. They used M5T and Turc methods to predict. They observed that the M5T method, which is one of the data mining methods, gave better estimation results than the Turc empirical method. Taşar et al. (2018) selected the study area as Massachusett, USA (Cambridge Reservoir and Basin) to determine the average daily ET0, by using wind speed (U), duration of sunshine (DS), relative humidity (RH) and then they compared the results of traditional Hargreaves-Samani, Ritchie and Turc methods. When the empirical methods and ANN model results were compared, it was observed that ANN model gave better results than empirical methods. The aim of this study is to investigate the feasibility and validity of artificial neural network (ANN) method and classical methods for a different study area. Daily data were taken from the meteorological station near the Hartwell lake in the Anderson region, South Carolina, USA.

2. Methodology In this study, Hargreaves-Samani, Turc equations which are two of the empirical (classical) methods, and Artificial Neural Networks (ANN) method which is one of the artificial intelligence methods were used for prediction of daily ET0. 2.1. Hargreaves-Samani Equation Necessary parameters for calculation of daily ET with Hargreaves-Samani equation are daily maximum temperature (Tmax), daily minimum temperature (Tmin) and extraterrestrial solar radiation (Rs) (Hargreaves & Samani, 1985). The equation which is used for calculation is given below;

ET=0.0135×0.408 Rs (T+17.8) (Eq. 1)

=0.16 . ) (Eq. 2)

227

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY where, “T” represents daily mean temperature and “Rs” extraterrestrial solar radiation in Hargreaves-Samani equation. 2.2. Penman FAO 56 Equation SR, AT, RH, U daily meteorological parameters are needed to calculate daily ET using Penman FAO 56 equation. Equation is given as below (Jensen et al., 1990),

= (Eq. 3) where γ is the psychometric constant, ∆ is the slope of the vapour pressure curve, Rn is the net radiation, u2 is the wind speed at 2 m height, ew is the saturation vapour pressure, ea is the actual vapour pressure and λ is the latent heat of vaporization in equation. 2.3. Turc Equation In the Turc method, the amount of evaporation depends on the parameters such as air temperature, relative humidity and amount of sunbathing. The Turc equation is given below (Turc, 1961).

If RH > %50 ; ET = 0.0133× × ( SR + 50 ) (Eq. 4)

If RH < %50; ET = 0.0133× × ( SR + 50 ) × (Eq. 5)

Where,ET is the daily evapotranspiration [mm day-1], SR is the solar radiation [MJ m-2 day-1], 0 Tm is the mean air temperature [ C] and RH is the relative humidity [%]. 2.4. Artificial neural networks (ANN) The concept of artificial neural networks first emerged with the idea of making computer simulations based on the working principle of the brain. According to Yurtoğlu (2005); ANN defines the relationship between input variables and target variables of previous examples by weight assignment method. In other words, they are trained. Once these relationships have been determined, the ANN can now run estimates with new data. The performance of a trained network is measured by the intended signal and error criterion. The output of the network is compared with the intended output to obtain a margin of error. An algorithm called Back Propagation is used to adjust the weights to reduce the margin of error. This process is repeated several times and the training of the network is completed. With this training process, it is tried to obtain the most suitable solution based on performance measurements.

228

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 1 - Artificial Neural Network Layers used in this study

ANN consists input layer, hidden layer and output layer as it is shown in Figure 1. Wij and Wjk are weights of connections between layers. 3. Study Area The station which is used in this study given in Figure 2, is located in South Carolina State in the southeastern part of the USA, where it is located in temperate subtropical climate zone and the summers are very hot and humid, while winters are mild and soft. The station is located on 34 ° 30'30 "latitude and 82 ° 51'19" longitude and is managed by the South Atlantic WSC Clemson Field Office. Data set which is taken from station contains 4 years daily records (2013- 2017).

Figure 2 - Location of station in South Carolina (USGS) 4. Results and Discussions

In this study, ET0 is estimated by using artificial neural networks (ANN), Hargreaves- Samani (HS) and Turc (T) equations. Obtained results are compared with each other. 1416 total daily meteorological data were used which include 2013-2017 time period. In the study, 80% of all data is reserved for training, 20% for testing. 1133 daily data was used for the training and 283 day measurement data were used for the testing. The data used was taken from the USGS. At first, ET0 values are calculated by using standard PM FAO 56 equation. Necessary parameters for this calculations are daily air temperature, daily maximum air temperature, daily minimum temperature, wind speed, relative humidity and sunshine duration. In the second part of the study ET is calculated by Hargreaves-Samani and Turc equations and ET estimations belonging to test set are done with ANN method. Mean square error (MSE), mean absolute error (MAE) and correlation coefficient (R) statistics were used to determine the success of the models used to estimate the daily ET0 value. 1 N MAE =  ETiobserved − ETiestimate N i=1 (Eq. 6)

229

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

2 1 N MSE = (ETiobserved − ETiestimate ) N i=1 (Eq. 7) where, N represents data numbers and ETi daily evapotranspiration data. Table 1 - Test set statistics

Hargreaves-Samani Turc ANN INPUTS T,SR T, SR, RH T, SR, RH,U MSE 1.524 1.077 0.119 MAE 1.061 0.878 0.290 R 0.686 0.613 0.976

All calculated MSE, MAE and correlation coefficient statistics for test set are given with Table 1.

Figure 3 - Hargreaves-Samani equation distrubition graph for test set

230

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 4 - Hargreaves-Samani equation scatter chart for test set Figure 3. and Figure 4. show the performance of Hargreaves-Samani equation against PM FAO 56 equation. Correlation coefficient for Hargreaves-Samani is calculated as 0.686.

Figure 5 - Turc equaiton distrubition graph for test set

231

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 6 - Turc equation scatter chart for test set

Figure 5. and Figure 6. show the performance of Turc equation against PM FAO 56 equation. Correlation coefficient for Turc is calculated as 0.613.

Figure 7 - ANN method distribution graph for test set

232

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 8 - ANN method scatter chart for test set Figure 7. and Figure 8. show the performance of ANN method. Correlation coefficient for ANN is calculated as 0.976 which is the highest correlation value of this study. The purpose of these graphs is to determine how the ET0 values calculated by conventional methods and ANN are different from actual ET0 observations. It is possible to see high performance of ANN method against conventional methods from drawn graphs and charts. As a result, it is seen that estimation of daily ET0 by using daily meteorological parameters can be done by empirical equations and ANN method both. But, as the results show that empirical equations errors values are higher than ANN method and also correlation between ANN method and PM FAO 56 equation is much more acceptable for the study area. Authors suggest that similar solutions must be done with ANN by using less input meteorological parameters to understand ANN performance better.

5. Conclusion In this paper, ET0 is tried to be estimated by using Hargreaves-Samani and Turc traditional equations and results are compared with Artificial Neural Network (ANN) model performance. A station which is stated near to the Hartwell Lake (South Carolina, USA) was chosen as the study area. Applicability of ANN models was investigated using meteorological variables such as mean daily air temperature, wind speed, solar radiation and, relative humidity for ET estimation. As a result of the present study, ANN methods give correct results in solving the problem. This method could provide low mean square error (MSE) and mean absolute error (MAE) values for estimating the amount of daily ET0. Hargreaves-Samani, and Turc methods have the worst results in all criteria. Nevertheless, ANN's performance has been better than empirical methods in ET predictions.

233

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

The presented work has shown that daily evaporation modeling is determined with the ANN method. As an alternative to the empirical methods of Hargreaves-Samani, and Turc, the ANN model of artificial intelligence methods can be presented in estimating the amount of daily ET0. In estimating the amount of evapotranspiration, ANN is more advantageous than the traditional methods because of the ANN structure’s nonlinear dynamics to the solution problem. 6. References

1. Abtew, W. (1996). Evapotranspiration measurement and modeling for three wetland systems in South Florida. Water Resour. Bull. 32, 465–473.

2. Aytek, A., Güven, A.,Yüce, M.İ., & Aksoy, H. (2008). An explicit neural network formulation for evapotranspiration. Hydrological Sciences Journal, 53, 4, 893-904, DOI: 10.1623/ hysj.53.4.893.

3. Choudhury, B.J. (1999). Evaluation of an empirical equation for annual evaporation using field observations and results from a biophysical model. Journal ofHydrology, 216 (12), 99–110. doi:10.1016/S0022-1694(98)00293-5.

4. Demirci, M., & Baltaci, A. (2013). Prediction of suspended sediment in river using fuzzy logic and multilinear regression approaches, Neural Computing Applications, 23, 145-151.

5. Demirci, M., Üneş, F., & Aköz, M.S. (2015). Prediction of cross-shore sandbar volumes using neural network approach. Journal of Marine Science and Technology 20(1), 171-179.

6. Demirci, M., Unes, F., & Akoz, M. S. (2016). Determination of nearshore sandbar crest depth using neural network approach. International Journal of Advanced Engineering Research and Science, 3(12).

7. Dogan, E., Isik, S., & Sandalci, M. (2007). Estimation of daily evaporation using artificial neural networks. Tek Dergi 18 (2), 4119–4131.

8. Fenga, Y., Cuib, N., Zhaob, L., Hud, X., & Gonga, D. (2016). Comparison of ELM, GANN, WNN and empirical models for estimating reference evapotranspiration in humid region of Southwest China. Journal of Hydrology, 536, 376–383.

9. Gümüş,V., Yenigün, K., Toprak, F., & Baçi, N. (2018). Şanlıurfa ve Diyarbakır istasyonlarında sıcaklık tabanlı buharlaşma tahmininde YSA, ANFIS ve GEP yöntemlerinin karşılaştırılması, DÜMF Mühendislik Dergisi 9, syf. 553 – 562

10. Hargreaves, G. H., & Samani, Z. A. (1985). Reference crop evapotranspiration from temperature. Appl. Engng. Agric. 1(2), 96–99.

11. Jensen, M. E., Burman, R. D., & Allen, R. G. (1990). Evapotranspiration and Irrigation Water Requirements. ASCE Manuals and Reports on Engineering Practices no. 70., ASCE, New York, USA.

12. Kaya, Y. Z., Mamak, M., & Unes, F. (2016). Evapotranspiration Prediction Using M5T Data Mining Method. International Journal of Advanced Engineering Research and Science (IJAERS), 3(12), 225- 229.

13. McKenzie RS., & Craig AR. (2001). Evaluation of river losses from the Orange River using hydraulic modelling. J Hydrol. 241, (12), 62–9.

14. Partal, T. (2016). Comparison of wavelet based hybrid models for daily evapotranspiration estimation using meteorological data. KSCE Journal of Civil Engineering, 20(5), 2050–2058

15. Penman, H.L. (1948). Natural evaporation from open water, bare soil and grass. Proc R Soc Lond 193:120–146

16. Romanenko, V.A. (1961). Computation of the Autumn Soil Moisture Using a Universal Relationship for a Large Area. Ukrainian Hydrometeorological Research Institute, Kiev, No. 3.

234

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

17. Taşar, B., Kaya, Y. Z., Varçin, H., Üneş, F., & Demirci, M. (2017). Forecasting of suspended sediment in rivers using artificial neural networks approach. International Journal of Advanced Engineering Research and Science, 4(12).

18. Taşar B., Üneş F., Demirci. M., & Kaya Y.Z. (2018). Forecasting of Daily Evaporation Amounts Using Artificial Neural Networks. Journal of Dicle University Engineering 9(1), 543.

19. Thornthwaite, C.W. (1948) An approach toward a rational classification of climate. Geograph. Rev., 38, 55-94

20. Turc, L. (1961). Evaluation des besoins en eau d’irrigation, évapotranspiration potentielle, formulation simplifié et mise à jour. Ann. Agronomiques 12, 13–49.

21. USGS.gov | Science for a changing world [WWW Document], n.d. URL https://www.usgs.gov/

22. Warnaka & Pochop. (1988). Analysis of Equations for Free Water Evaporation Estimates. Water Resources Research 24(7), 979-984. DOI: 10.1029/WR024i007p00979

23. Yurtoğlu, (2005). Yapay sinir ağlari metodolojisi ile öngörü modellemesi: bazı makroekonomik değişkenler için Türkiye örneği, DPT.

24. Ünes, F., Yildirim, S., Cigizoglu, H.K., & Coskun, H. (2013). Estimation of dam reservoir volume fluctuations using artificial neural network and support vector regression. Journal of Engineering Research, 1(3), 53-74.

25. Üneş, F., Demirci, M., & Kişi, Ö. (2015). Prediction of millers ferry dam reservoir level in usa using artificial neural network. Periodica Polytechnica Civil Engineering, 59(3), 309–318.

26. Unes, F., Gumuscan, F.G., &Demirci, M. (2017). Prediction of Dam Reservoir Volume Fluctuations Using Adaptive Neuro Fuzzy Approach. EJENS, 2 (1), 144-148.

235

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

The Importance of Business Intelligence Solutions in The Industry 4.0 Concept Melda Kokoç1, Süleyman Ersöz2, A. Kürşad Türker3

1 ECTS Coordination Office, Gazi University, Turkey [email protected] 2, 3 Department of Industrial Engineering, Kırıkkale University, Turkey [email protected] [email protected]

Abstract

Industry 4.0, as the name suggests, is the fourth phase of industrialization, which aims at high level of automation in the manufacturing industry by adopting information and communication technologies. Information systems have become more important at this fourth phase. In this study, the relationship between Industry 4.0 and information systems is reviewed and it is discussed how intelligent business systems can be effectively addressed through the Industry 4.0 revolution and how it can help to reveal the production potential.

Keywords: Business intelligence, industry 4.0, information system

1. Introduction

Significant changes in the industrial revolution and remarkable developments in the field of mechanical and microelectronics technology have led to formation of an information society. Together with the communication sectors, information revolution has emerged and computers have involved in this process (Bulut, 2017). With this process, many concepts such as internet of things, big data analysis and artificial intelligence have emerged. These technologies have penetrated into the manufacturing industry and have increasingly made special requirements such as better quality, shortening a period to place on market. A variety of sensors are used in the equipment to meet the specific requirements in advanced technologies and to enable the machines to self-sense, act individually and communicate with each other. With these technologies, real-time production data can be obtained and shared to facilitate fast and accurate decision making. Using advanced information analytics, the idea that networked machines can perform more efficiently, collaboratively and flexibly has transformed the manufacturing industry into the next generation, namely Industry 4.0. This process, named as Industry 4.0, was introduced in 2011 and it has been drawing considerable interest in business and academic world since then. Industry 4.0 consists of nine pillars in technology. These are virtual reality, artificial intelligence, industrial internet, industrial big data, industrial robot, 3D printing, cloud computing, information processing automation and industrial network security presented in Figure 1 (Cheng et al., 2016).

236

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 1-Nine pillars of Industry 4.0 (Rüßmann et al., 2015)

In this study, developments coming with Industry 4.0 and the place and importance of information systems in Industry 4.0 are discussed. In the second part, some studies on Industry 4.0 are included. In the third part, the importance of developments obtained with Industry 4.0 and knowledge integration are explained. In the fourth part, significance of intelligent business systems in Industry 4.0 is mentioned. In the conclusion part, general evaluation is given.

2. Literature Research

With increasing interest in Industry 4.0 in recent years, this concept has taken its place as a new research area in literature. Some of studies in literature are summarized below. In order to get a systematic literature review on the purpose of reviewing academic articles, Lu (2017), Liao et al. (2017), Ben-Daya et al. (2017) and Saucedo-Martinez et al. (2017)’s studies can be benefited. Jazdi (2014) defined the significance of internet of things and its important role in terms of future professional and daily life. In addition, it was shown that Industry 4.0 had already begun and would clearly affect our lives and future business model. For instance, significance of Industry 4.0 was explained by a project on industrial coffee machines in the Industrial Automation and Software Engineering Institute. Rüßmann et al. (2015) identified the nine technology trends that constitute basic units of Industry 4.0 and discussed their potential technical and economic benefits for producers and production equipment suppliers. Case studies from the German industry known as the world leader in industrial automation were also used to present and share information and findings. In Wang et al. (2016)’s study, a smart factory infrastructure including industrial network, cloud and supervisor control panels with intelligent workshop-based objects such as machines, conveyors and products was presented. Arkan (2018) examined a diaper factory as an example of Industry 4.0 transformation and a baby diaper production line was revised considering the new system in second half of 2014. With systematic approach generated in study, the level of compatibility with industry 4.0 in both systems was analyzed. Then, based on real production data and financial reports, the impact of industry 4.0 transformation on cost efficiency was evaluated. In Çevik (2018)’s study, the eras of industrial revolution were researched so as to analyze added values of Industry 4.0 from the point of today and future. In the study of Yıldız (2018), Industry 4.0 and basic paradigms were explained, information was given about intelligent factories and a general evaluation was made.

237

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

3. Industry 3.0 and Management Information Systems

Considering developments in the past century, the role of the customers in both production stage and consumption stage in terms of products or services has varied. In the beginning, the need for planning emerged and then it was realized that a plan was required at every stage of production. In parallel with these developments, the relations of national enterprises with foreign enterprises also improved and thus the concept of competition triggered the need for differentiation in enterprises as a global process. Businesses that cause these differences have had to create decision support mechanisms that enable them to take quick decisions to assess the opportunities that have competitive advantages and to achieve their long-term goals. Depending on this situation, enterprises have begun to investigate the functional software which can effectively use all functions in their internal operations, and then they have started to use the functional software they developed in their internal processes. As a result of the developments in the information management systems, it has been realized that the harmonious and integrated of these functional software can bring positive acceleration to productivity. Information management systems integrate supplier and customer relations functions except the basic operations of enterprises into the process structure of enterprises. In this way, enterprises have an opportunity to manage demands of the customers in electronic environment and to plan and follow the production in accordance with these demands and to reach the suppliers in electronic media by taking into consideration the material needs and stocks arising from the demands. In addition, it provides companies with an opportunity to select, evaluate and keep suppliers under control (Yenigül, 2004). The most important feature of the MIS is that it can share the resources of the factories in different regions, supplier companies and distribution centers in a coordinated way. In this context, the capacity and characteristics of distribution, production and supply sources of all affiliated companies are taken into consideration simultaneously in order to meet the demand of the customer in the shortest time, at the desired quality and cost (Beşkese, 2004). The aim of the implementation of information management systems is to develop external links and to support the firm's value chain activities as well as to provide process integration within the enterprise.

4. Industry 4.0 and Business Intelligence

Industry 4.0 can be defined as the fact that machine power is able to manage production processes by replacing manpower. The fact that machines can be coordinated through new developments in computers and internet technologies has emerged new industrial revolution. With this new system known as the Internet of Things (IoT) term, production technology has advanced and factories have gained self-manage ability (Bulut, 2017). One of the concepts that comes with the new industrial revolution is the Cyber-Physical system (Figure 2). The connection of physical production equipment and devices along with the large data analysis in the digital world over the Internet has led to the emergence of revolutionary production tools, namely cyber-physical production systems (CPSs) (Zheng et al., 2018).

238

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 2- Industrial Revolution Stage (1. https://www.fostec.com/en/competences/digitalisation-strategy/industry-4-0/)

CPSs are defined as transformative technologies to manage interconnected systems between physical assets and computing capabilities. With the latest developments resulting in higher availability of sensors, data acquisition systems and computer networks, the competitive nature of today's industries leads to more factories implementing high-tech methodologies. As a result, the increasing use of sensors and networked machines have led to the continuous production of high volume data known as Big Data. In such an environment, the CPS has been developed to manage Big Data and to take advantage of the interconnection of machines to reach interconnected, intelligent, flexible and self-adapting machines (Lee et al., 2015). These linked CPSs can interact with each other using standard internet-based protocols and analyze data to predict errors, configure themselves and adapt to changes. Industry 4.0 makes it possible to collect and analyze data between machines, and to have more flexible and more efficient processes to produce higher quality goods at lower costs. Thus, production efficiency, shift economy, industry growth will be supported, the profile of the workforce will be changed and the competitiveness of the companies will be changed (Rüßmann et al., 2015; Stock and Seliger, 2016). For these systems, the level of conversion of data to information is among the distinguishing features. Therefore, in the process of transition to Industry 4.0, to create national software, to train qualified programmers are the most important stages (Sener and Elevli, 2017). Information technologies, such as IoT, big data, and cloud computing, along with artificial intelligence technologies, help to implement the smart factory of Industry 4.0. Smart machines, conveyors and products communicate and negotiate with each other to reconfigure themselves for flexible production of multiple products. The industrial network collects large data from smart objects and transfers them to the clouds. This provides feedback and coordination for the large data analytics- based system to optimize system performance (Jazdi, 2014; Wang et al., 2016). The smart factory, which is one of the fundamental principles of Industry 4.0, deals with vertical integration and networked production systems for intelligent production. In this principle, the importance of information system integration and network connection is prominent. Intelligent objects must be combined with large data analyzes to implement intelligent factories. While intelligent objects can be dynamically reconfigured to achieve high flexibility, large data analysis can provide global feedback and coordination to achieve high efficiency. In this way, the smart factory can produce customized and small lot products efficiently and profitably (Lee et al., 2015; Wang et al., 2016). Furthermore, Industry 4.0 significantly affects the production environment and operations. Unlike traditional forecast-based production planning, Industry 4.0 enables real-time planning of production plans through integrated information systems (Tupa et al., 2017). In the Industry 4.0 manufacturing environment, it is possible to manage the production of management information systems, which are supported with instant data, in order to conduct the works

239

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY under ideal conditions. However, in case of unforeseen situations, intelligent support systems based on decision support, artificial intelligence, expert systems and data mining will be needed to enable the system to reactivate itself and produce alternatives. In parallel with the communication between the machines, machines need to produce intelligent decisions that can realize themselves in the structure of the information system. These systems are defined as follows: When there are problems in the processes; • Business intelligence systems that define alternative processes, • Business intelligence systems that produce overtime decisions, • Business intelligence systems that produce subsidiary industry decisions. When there are problems in the raw material; • Business intelligence systems that identify alternative raw materials, • Business intelligence systems that identify alternative suppliers, • Business intelligence systems that commission alternative production orders. When demand contraction problems, which is related to product, in the market are experienced; • Business intelligence systems that define alternative product work orders.

Intelligent systems to be created will vary depend on product range and characteristics of the production system.

5. Conclusion

Industry 4.0 is a revolution that emerges as a collective term including automation system, data exchange and production technologies. This revolution will enable more efficient business models to emerge as it enables each data to be collected and analyzed in a well-organized environment. In this study, some of the opportunities that Industry 4.0 brings to the production sector are discussed. It has been explained that increased digitalization and advanced cyber-physical intelligence at the factory will effectively address the many important concerns faced by manufacturers. Solutions to problems that may arise with respect to processes, raw materials and products are presented with intelligent business systems that can be applied with Industry 4.0.

References 17. Arkan, Ö. (2018). A Case Study on the Concept of Industry 4.0 and the Effect of Industry 4.0 Transformation on Production Costs: Diaper Production, Master Thesis, Istanbul Arel University, Institute of Social Sciences, Department of Accounting and Finance. 18. Ben-Daya, M., Hassini, E., & Bahroun, Z. (2017). Internet of things and supply chain management: a literature review. International Journal of Production Research, 1-24. 19. Beşkese, M. (2004). Techniques for Erp Software Selection. Istanbul Technical University Proceedings Book, 144-148. 20. Bulut, E., & Akçacı, T. (2017). Industry 4.0 And Within the Scope of Innovation Indicators Analysis of Turkey. ASSAM International Refereed Journal, 4(7), 55-77. 21. Çevik, G., Z.,(2018). With Respect to Industry 4.0 an Analysis on Turkey’s Current and Future State, Master Thesis, Nişantaşı University, Institute of Social Sciences, Department of Business Administration. 22. Cheng, G. J., Liu, L. T., Qiang, X. J., & Liu, Y. (2016, June). Industry 4.0 development and application of intelligent manufacturing. In Information System and Artificial Intelligence (ISAI), 2016 International Conference on (pp. 407-410). IEEE. 23. https://www.fostec.com/en/competences/digitalisation-strategy/industry-4-0/. 24. Jazdi, N. (2014). Cyber physical systems in the context of Industry 4.0. In Automation, Quality and Testing, Robotics, 2014 IEEE International Conference on (pp. 1-4). IEEE. 25. Lee, J., Bagheri, B., & Kao, H. A. (2015). A cyber-physical systems architecture for industry 4.0-based manufacturing systems. Manufacturing Letters, 3, 18-23.

240

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

26. Lee, J., Bagheri, B., & Kao, H. A. (2015). A cyber-physical systems architecture for industry 4.0-based manufacturing systems. Manufacturing Letters, 3, 18-23. 27. Liao, Y., Deschamps, F., Loures, E. D. F. R., & Ramos, L. F. P. (2017). Past, present and future of Industry 4.0-a systematic literature review and research agenda proposal. International journal of production research, 55(12), 3609-3629. 28. Lu, Y. (2017). Industry 4.0: A survey on technologies, applications and open research issues. Journal of Industrial Information Integration, 6, 1-10. 29. Rüßmann, M., Lorenz, M., Gerbert, P., Waldner, M., Justus, J., Engel, P., & Harnisch, M. (2015). Industry 4.0: The future of productivity and growth in manufacturing industries. Boston Consulting Group, 9. 30. Saucedo-Martínez, J. A., Pérez-Lara, M., Marmolejo-Saucedo, J. A., Salais-Fierro, T. E., & Vasant, P. (2017). Industry 4.0 framework for management and operations: a review. Journal of Ambient Intelligence and Humanized Computing, 1-13. 31. Sener, S., & Elevli, B. (2017). New Businesses and Higher Education in Industry 4.0. Journal of Engineer Braıns, 2(1), 25-37. 32. Stock, T., & Seliger, G. (2016). Opportunities of sustainable manufacturing in industry 4.0. Procedia Cirp, 40, 536-541. 33. Stock, T., & Seliger, G. (2016). Opportunities of sustainable manufacturing in industry 4.0. Procedia Cirp, 40, 536-541. 34. Tupa, J., Simota, J., & Steiner, F. (2017). Aspects of risk management implementation for Industry 4.0. Procedia Manufacturing, 11, 1223-1230. 35. Wang, S., Wan, J., Zhang, D., Li, D., & Zhang, C. (2016). Towards smart factory for industry 4.0: a self- organized multi-agent system with big data based feedback and coordination. Computer Networks, 101, 158- 168. 36. Yenigül, M.F., (2004). Enterprise Resource Planning and Implementation in Turkey, Gazi University Industrial Engineering Publications, Issue 1. 37. Yıldız, A. (2018). Industry 4.0 and smart factories. Sakarya University Journal of Science, 22(2), 546-556. 38. Zheng, P., Sang, Z., Zhong, R. Y., Liu, Y., Liu, C., Mubarok, K., ... & Xu, X. (2018). Smart manufacturing systems for Industry 4.0: Conceptual framework, scenarios, and future perspectives. Frontiers of Mechanical Engineering, 1-14.

241

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Turkish Abusive Message Detection with Methods of Classification Algorithms Habibe Karayiğit1, Çiğdem Aci2, Ali Akdağli3

Department of Electrical and Electronics Engineering, Mersin University, Turkey [email protected] Department of Computer Engineering, Mersin University, Turkey [email protected] [email protected]

Abstract

The era in which we live is the era of technology. With the use of Internet Access to the smart phone, people started to spend more time the internet. More data is available as social media is used more often. This huge data accumulated in social media is a very important resource for researchers and analysts.

Sentiment and Opinion Analysis examines whether the data contains emotions and basically examines their positive-negative-neutral states. In this study, the data obtained from Turkish social media were analyzed by using classification algorithms such as Linear SVC, Logistic Regression (LR) and Random Forest (RF). As a result of the analyzes, the results were observed and it was explained which classification method gave better results. Which method gives better results is indicated in the experimental results part. Recommendations for better results are indicated in the conclusion part.

Keywords: Turkısh Abusive Messages, Turkish Sentiment Analysis, Machine Learning, Linear SVC, Logistic Regression, Random Forest, Social Media, Instagram

1. Introduction

Nowadays; Since the forums, blogs and social media are used very intensively by all everyone, people have now started to share their opinions, ideas and feelings through these environments. People live a virtual social life in the virtual world with social media. In this virtual social life, users can share their feelings and thoughts and make comments to their friends. They can see not only the sharing of friends but also the shares of people who are famous in political, sporting and media terms.

Recently, unidentified or identifiable users send swearing messages, which we can see more often, especially in the comments made on the magazines. These users will not dare to do it in real life, but they can write racist or sexist comments in a virtual environment. Turkish abusive messages is analyzed with the method of emotion analysis and opinion mining.

Sentiment analysis and opinion mining is a field of study that analyzes people's views, evaluations, attitudes, and emotions from a written platform. Sentiment analysis is used to find what an existing text expresses emotionally.

The first study on sentiment analysis was conducted by Pang, Lee and Vaithyanatham in 2002 and movie comments existing in Internet Movie Database were taken and classified by various machine learning algorithms [1].

The first study in which the concepts of Sentiment Analysis and Opinion Mining are used together, semantic relations were established between emotional expressions and the subject rather than being classified as positive or negative.

242

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

In the first study in which the concepts of sentimental analysis and opinion mining were used together, semantic relations were established between emotional expressions and the subject rather than positive or negative classification [2].

In the first work under the name of opinion mining, a list of both product qualities (quality, characteristics) was created using the opinion mining tools, and opinions (weak, mixed, good) were collected about each one [3].

In the second chapter, Turkish data set, pre-processing, model creation and feature extraction stages are mentioned. In the third part, classification methods used in the data set are explained. In the fourth part, the success rate was measured by the classification methods. In the conclusion part, evaluation was made for Turkish abusive analysis and it has been tried to suggest solutions for the problems encountered in studies conducted in Turkish. In the last section, references are included.

2. Testing Ground

d. Data Set

Grammar rules vary for each language. Verb conjugation in a sentence differs between languages. The Turkish language is a structurally agglutinative language. Structural processing is more difficult than English. In Turkish, verb conjugation and lexical items are different from other languages as shown in Figure 1 [4].

Gender discrimination in languages such as Arabic, English, German does not exist in Turkish. A sentence spoken by multiple words in English can be explained in Turkish by a word. For example; ‘Gidemedik’ is expressed in English 'We were not going'. There are 8 vowels (a, e, i, o, ö, u, ü) and 21 consonants (b, c, ç, d, f, g, ğ, h, j, k , l, m, n, p, r, s, ş, t, v, y, z) in Turkish. The seven letters are unique to the Turkish language alphabet (ç, ı, ş, ö, ü, ğ, I).[zzz] When doing sentiment analysis in Turkish, you can not use language models like sentiment analysis in English but all of the methods used in the sentiment analysis literature can be used for the Turkish sentiment analysis [5].

Figure 1 – Sentence structures in Turkish and English [6]

The data set consists of 6077 messages from social media. These messages were obtained from the social media site called Instagram, which was written in Python. Messages in the data set are classified as 1, if not 0 by hand procedure. There are 927 data in abusive class and 5150 data in other class. 20% of the training data set is used as test data.

e. Data Set Features

Because of the data set is unsupervised data, it has been subjected to a certain pretreatment. The stages of the report are as shown in Figure 2.

243

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Figure 2 – Stage of Turkish abusive message detection

These preprocesses; removing some of the symbols and signs, removing stop-words, finding stem only on the words of abusive. Then the list of terms has been created. Using the Countvectorizer and Tf -Idf methods, the number of data in the data set and the maximum utilization rate were found and converted into vectors.

3. Classification and Feature Selection Algorithms

Python program was used in this report. This programme is preferred because there are too many libraries in the Python program and there are many examples in this programme. Random Forest, Linear SVC and Logistic Regression classification algorithms were tried on the data set, respectively. Text in the data set must be converted to numeric expressions. Conversion to vectors is essential for machine learning algorithms to work.

a. Logistic Regression Algorithms

In the multiclass case, the training algorithm uses the one-vs-rest (OvR) scheme if the ‘multiclass’ option is set to ‘ovr’, and uses the cross- entropy loss if the ‘multiclass’ option is set to ‘multinomial’. (Currently only the ‘lbfgs’, ‘sag’ and ‘newton-cg’ solvers support the ‘multinomial’ option.) [7].

This class implements regularized logistic regression using the ‘liblinear’ library, ‘newton-cg’, ‘sag’ and ‘lbfgs’ solvers. It can handle both dense and sparse input. Use C-ordered arrays or CSR matrices containing 64-bit floats for optimal performance; any other input format will be converted (and copied) [8].

b. Random Forest Classification Algorithms

Random Forest algorithm generates more than one decision tree for more successful results, it selects the best of the randomly received attributes on each node to divide all nodes. The decision tree uses N times the algorithm to make better results. The average of the results obtained by using N times is taken to increase the accuracy rate of the estimation. Decision trees are randomly received subsets from the train set.

c. Linear Support Vector Classification Algorithms

Similar to SVC with parameter kernel=’linear’, but implemented in terms of liblinear rather than libsvm, so it has more flexibility in the choice of penalties and loss functions and should scale better to large numbers of samples [9].

This class supports both dense and sparse input and the multiclass support is handled according to a one-vs-the-rest scheme [10].

244

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

4. Experimental Results

6077 social media messages were used in the study. Classification using linear SVC, Random Forest and Linear Regression classifiers was performed. The data set has been subjected to certain preprocessing steps, resulting in the removal of undesirable symbols and signs. The Turkish stop words are removed from the data set. A different method was applied as stemming process. A stemming process was applied only for abusive words and the success rate of the classification algorithm was increased.

The dataset was converted to vectors with the Countvectorizer operation. The reason for converting words into vectors is that classification algorithms can make classification on vectors and matrices. TF and IDF values were generated. 80% of the available data was used as training data and 20% as test data. . In linear SVC, a success rate of 80% was higher than other algorithms. 79% success was achieved by random forest classification and 73% success was achieved by logistic regression classification.

After the pre-processing of the data, removing the words of the stop and finding the root values, the most changing value was obtained in the Logistic Regression classification. The success rate, which was 50% at the beginning, reached to 73% at the end of the pre-processing phase.

Table 1 –Model performance results obtained by classification RANDOM precision recall f1-score support FOREST 0 0.96 0.97 0.96 1038 1 0.81 0.78 0.79 178 Avg/Total 0.94 0.94 0.94 1216 LINEAR SVC precision recall f1-score support 0 0.96 0.97 0.97 1038 1 0.82 0.78 0.80 178 Avg/Total 0.94 0.94 0.94 1216 LOGISTIC precision recall f1-score support REGRESSİON 0 0.94 0.98 0.96 1038 1 0.85 0.65 0.74 178 Avg/Total 0.93 0.93 0.93 1216

5. Conclusion

A text mining application based on emotion analysis was used by using various classification algorithms with the data set formed from the mixed messages obtained from social media. A data set has been created, all in Turkish. Three algorithms have been tried. The most successful of these was the classifier Linear SVC. However, the values of the other two classification algorithms were close to Linear SVC.

A study on this subject has not been found in the field of Turkish text mining and Turkish sentiment analysis. Such a study is considered to be the first in our country. Increasing the data in the data set, finding the stemming process of all words and using various selection algorithms are thought to increase success.

245

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

References 1. B. Pang, L. Lillian, V. Shivakumar, Thumbs up? Sentiment Classification using Machine Learning Techniques, Proceedings of the ACL-02 conference on Empirical methods in natural language Processing- Volume 10. Association for Computational Linguistics, 2002 2. T. Nasukawa, J. Yi, Sentiment analysis: Capturing favorability using natural language processing, Proceedings of the 2nd international conference on Knowledge capture, Sanibel Island, FL, USA, 2003. 3. K. Dave, S. Lawrence, D.M. Pennock, Mining the Peanut Gallery: Opinion Extraction and Semantic Classification of Product Reviews, WWW Conference, 2003 4. H. Karayiğit, Ç. Acı and A. Akdağlı, A Review of Turkish Sentiment Analysis and Opinion Mining, Balkan Journal of Electrical and Computer Engineering, pp. 26-30 5. H. Karayiğit, Ç. Acı and A. Akdağlı, A Review of Turkish Sentiment Analysis and Opinion Mining, Balkan Journal of Electrical and Computer Engineering, pp. 26-30 6. H. Karayiğit, Ç. Acı and A. Akdağlı, A Review of Turkish Sentiment Analysis and Opinion Mining, Balkan Journal of Electrical and Computer Engineering, pp. 26-30 7. Scikitlearn.20 Ekim 2018 tarihinde http://scikitlearn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html adresinden erişildi. 8. Scikitlearn.20 Ekim 2018 tarihinde http://scikitlearn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html adresinden erişildi. 9. Scikitlearn.25 Ekim 2018 tarihinde http://scikit-learn.org/stable/modules/generated/sklearn.svm.LinearSVC.html adresinden erişildi. 10. Scikitlearn.25 Ekim 2018 tarihinde http://scikit-learn.org/stable/modules/generated/sklearn.svm.LinearSVC.html adresinden erişildi.

246

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

A K Adil Baykasoğlu ...... 123 Kerem Elibal ...... 77 Adnan Aktepe ...... 55, 113 Kerim Melih Çimrin ...... 172 Ahmet Gökçen...... 164 M Ahmet Kürşad Türker ...... 55, 113, 236 Ali Akdağli ...... 242 Mehmet Kabak ...... 77 Ali Fırat İnal ...... 55, 113 Melda Kokoç ...... 236 Aptulgalip Karabulut ...... 89 Mete Celik ...... 140 Metin Dağdeviren ...... 77 B Muharrem Düğenci ...... 70 Bestami Taşar ...... 106, 226 Münire Sibel Çetin ...... 159 Bülent Sezen ...... 20 Murat Uçar ...... 65 Burak Erkayman ...... 197 Mustafa Demirci ...... 106, 226 C N Cansu Canbolat ...... 187 Nazmiye Çelik ...... 178 Çiğdem Aci...... 242 Nida Yeşilyurt ...... 221 Cihan Çetinkaya ...... 32, 178 O Cuma Karakuş ...... 124, 209 Oğuz Findik ...... 70 D P Damla Tunçbilek ...... 55 Devrim Kayali ...... 99 Pınar Kocabey Çiftçi...... 11 Domenico Belli ...... 48 S Duygu Erdem ...... 159 Sedat Tarakci ...... 89 E Selen Eligüzel Yenice ...... 20 Emine Uçar ...... 65 Selma Gülyeşil ...... 149 Ercan Öztemel ...... 37 Selya Açikel...... 164 Eren Özceylan ...... 32, 77 Serdar Yıldırım ...... 202 Serhan Ozdemir...... 89 F Süleyman Ersöz ...... 55, 113, 236 Fatih Peker ...... 209 Süleyman Serhan Narlı ...... 202 Fatih Üneş ...... 106, 226 Süreyya Doğan ...... 226 G T Gizem Karaca ...... 39 Tuğba Açıl ...... 133 Gokhan Altan...... 192 Turgut Özseven ...... 70 Güray Kılınççeker ...... 221 Türkay Dereli ...... 178 H U H. Gokay Bilic ...... 89 Ulus Çevik ...... 99 Habibe Karayiğit ...... 242 V Hakan Varçin ...... 106 Vahit Calisir ...... 48 I Y İbrahim Miraç Eligüzel ...... 32 İlker Mert ...... 209 Yakup Kutlu ...... 39, 133, 192 İlyas Özer ...... 70 Yaşar Daşdemir ...... 172, 202 İpek Abasıkeleş-Turgut ...... 187 Yunus Ziya Kaya ...... 106, 226 İrfan Yildirim ...... 140 Z İsmail Üstün ...... 124 Zeki Mertcan Bahadırlı ...... 106

247

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

Zeynep D. Unutmaz Durmuşoğlu ...... 11, 149

248

International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) Nov 15-17, 2018 Iskenderun, Hatay / TURKEY

249