Signal classification using novel pattern recognition methods and wavelet transforms

Abstract

A complete pattern recognition system consists of a sensor that gathers the observations to be classified or described; a feature extraction mechanism that computes numeric or symbolic information from the observations; and a classification or description scheme that does the actual job of classifying or describing observations, relying on the extracted features. A pattern recognition example, in this dissertation, is the Ballistocardiogram (BCG). The BCG measurement, recording systems, and signal pre-processing were studied as part of the work. The thesis reviews various BCG measurement techniques and devices, noise removal from the measurements and segmentation methods of the BCG signal.

Different types of wavelet transforms (WTs), as feature extraction methods, were studied and applied for the classification of BCG. A novel feature extraction method called Time-frequency moments singular value decomposition (TFMSVD) was also developed yielding results similar to the WT.

The development of machine learning algorithms is essential in developing intelligent systems such as autonomous robots. Artificial neural networks (ANNs) are one of the technologies in learning systems. Usually the learning process is based on training ANNs with a representative set of real world examples and then the trained network is embedded into a system. There are, however, a number of problems with most existing ANN structures. These include time consuming training, large amounts of training data and the fact that complicated structures are difficult to implement in embedded systems and integrated circuits, in particular.

The aim of the study was to address the above problems by developing novel methods for well-known pattern classification test data sets such as IRIS and Vowel data as well as for BCG. The developed learning algorithms (QuickLearn, CombilNet and its example SF-ART) performed equally well in pattern classification performance with conventional ANNs although SF-ART required less than ten training cycles. The QuickLearn algorithm classifies data almost as well as the traditional ANNs although it requires only one learning cycle.