Feature Extraction of Waveform Signals for Uncertain Dynamic Processes Using Neural Networks

Abstract

This paper presents a novel and simple feature extraction approach for drawing out the signal characteristics of uncertain dynamic processes by the feature neurons. Kohonen network is used to construct the feature neurons to represent its respective local features of a waveform signal. For a class of waveform signals, groups of feature neurons can be obtained. Incorporating with the ellipsoidal calculus, this approach can extract the process drifts and abnormal deviations in the process characteristics by limit checking. Moreover, it is robust even for the process with different process time durations. For the system with oscillatory transient response, this approach can be iteratively used to augment the amount of feature neurons to analyze the characteristics of any portion of the signal of interest in detail. With the merit of unsophisticatedness, this approach can be implemented for the determination of preventive maintenance and fault detection in the semiconductor manufacturing.