Classification of Infrasound Events

Classification of Infrasound Events Using Optimized Feature Vectors and a Bank of RBF Neural Networks

Fredric M. Ham, Ph.D. Harris Professor of Electrical Engineering Florida Institute of Technology 150 West University Boulevard Melbourne, Florida (USA) 32901-6975 [email protected]

ABSTRACT: Results are presented for a bank of Radial Basis Function (RBF) neural networks, to discriminate between different infrasonic events. Each module in the bank of RBF networks is responsible for classifying one of several events, and thus, is trained to identify only that particular event. However, each module is also trained to not classify all other events (i.e., negative reinforcement). Output thresholds of each module are set according to specific 3-D Receiver Operating Characteristic (ROC) curves. Preprocessing of the infrasound signals is carried out by extracting an optimal set of cepstral coefficients and their associated derivatives (i.e., successive differences) that form the feature vectors used to train and test the RBF networks. When the optimal feature vectors are used for training and testing, a significant increase in the correct classification rate, i.e., greater than 90%, can be achieved over the ad hoc method of selecting a sequential set of cepstral coefficients along with a subset of associated derivatives.

FREDRIC M. HAM, PH.D., B.S. Electrical Engineering, Iowa State University, 1976; M.S. Electrical Engineering, Iowa State University, 1979; and Ph.D. Electrical Engineering, Iowa State University, 1980; is Harris Professor of Electrical Engineering at Florida Institute of Technology in Melbourne. He is also the Director of the Information Processing Laboratory. During the 2004-2005 academic year he was the Interim Dean of the College of Engineering and from 2000-2005 he was the Associate Dean for Research in the College of Engineering. He worked 10 years in industry before coming to Florida Tech in 1988. He was with the Harris Corporation from 1980 to 1988 and the Shell Oil Company from 1976 to 1977. He is currently the President Elect of the International Neural Network Society (INNS) Board of Governors (he was also Secretary in 2004 and 2005). Dr. Ham is a Senior Member of IEEE and has been an Associate Editor for the IEEE Transactions on Neural Networks since 2001. He was recently an invited plenary speaker at the International Workshop on Systems, Signals and Image Processing held in Chalkida, Greece, September 22-24, 2005. He has published over 100 technical papers and reports, mostly in the area of neural networks, but also in the areas of signal processing, controls and biomedical engineering (specifically, biosenors). He holds 3 U.S. patents and is author of the textbook: Principles of Neurocomputing for Science and Engineering, McGraw-Hill, 2001. He has served on many program committees for several conferences including, IJCNN, WCCI and SPIE (to name a few), and was the Program Co-Chair for the International Joint Conference on Neural Networks (IJCNN) 2005 held in Montreal, Quebec, Canada. Dr. Ham's current research interests include: applications of neural networks, recently for classifying infrasound events, adaptive filtering, and development of non-invasive methods for monitoring blood glucose concentrations.