Time series pattern recognition using s-transform and soft computing

Abstract

This paper presents a new approach to time series pattern classification using a modified wavelet transform for feature extraction of non-stationary time series data and a fuzzy multilayer perceptron network to generate the rules and classify the patterns. Also simple rule based event detection systems are used in a hybrid manner to classify all the categories of the non-stationary data that occur in a power distribution network during faults, sudden switching operations, and transient disturbances. The choice of modified wavelet transform known as multiresolution S-transform is essential for transient time series data of very short duration as they can not be handled by conventional Fourier and other transform methods for extraction of relevant features pertinent for temporal pattern recognition applications. The trained network infers the output class membership value of an input pattern and a certainty measure is also presented to facilitate rule generation. Several simulated data patterns along with the classification scores are presented in this paper.