Analysis and classification of signals with time-varying characteristics
Ravi Tammana, Lalit Gupta · 1994
This dissertation focuses on identifying the problems involved in analyzing and classifying signals with time-varying characteristics and proposes solutions to the problems. In particular, the goal is to develop estimation, clustering, artifact rejection of signals in a signal set generated by repetition of an experiment. The signals considered include evoked response potentials which are the responses of the brain to external stimuli and target boundary representations which are generated by a multisegmentation technique. These signals generally tend to be noisy and also experience inconsistent time spacings of events in the signals in the signal set. Additionally, the data collected in evoked response potential experiments is contaminated with artifacts which are atypical signals. Noise, inconsistent time spacings and artifacts significantly complicate the analysis and classification of signals. Therefore, special attention has to be focused on developing techniques capable of accommodating such aberrations in the signal set. The problem of estimating a signal by directly averaging the signals in the signal set is identified and in order to improve the estimate of the evoked response potential by signal averaging, a non-linear alignment algorithm is developed to optimally align the events in the signals prior to the averaging operation. The algorithm makes no prior assumptions about the characteristics of the signal. Two efficient and systematic non-linear alignment-averaging methods are developed to estimate the signal. Results from a series of experiments conducted show that the non-linear alignment algorithms preserve the time-dependent events in the signals and the estimate of the signal obtained through non-linear alignment averaging is quite robust. A discrepancy measure is proposed to improve the clustering of patterns which experience time-varying characteristics. The discrepancy measure is an outcome of a non-linear alignment procedure which optimally aligns the events in the signals in order to minimize the dissimilarity between the signals. The K-means clustering algorithm is modified to use the discrepancy measure to compute the similarity between signals and the cluster centers. A series of clustering experiments conducted on identical data consisting of target representations using the modified and standard K-means algorithm show that clustering performance of the modified algorithm is significantly superior to that of the standard algorithm. The modified K-means clustering algorithm is used to demonstrate that the match and mismatch evoked response potentials fall into two distinct classes and is also used to remove artifacts in the signal set. For the classification of evoked response potentials, a three-layer backpropagation neural network model is selected to formulate a global and a localized classification approach. The backpropagation network in the global approach is designed to operate on a single evoked response potential which is the average of the evoked response potentials generated at all electrode sites used in an experiment. The localized classification system consists of several identical three-layer backpropagation networks. Each network is designed to operate on the evoked response potentials generated at a single electrode site. Several experiments involving match and mismatch tasks were designed and the results obtained show that the neural network classifiers are able to discriminate with a high degree of accuracy between match and mismatch conditions in evoked response potentials.