Performance analysis of wavelet scattering transform-based feature matrix for power system disturbances classification

Abstract

Recently, the wavelet scattering transform (WST) was introduced as a powerful feature extraction tool for classification processes. It provides good performance in applications involving audio signals, images, medical data, and quadcopters for structural health diagnosis. It is also employed in several electrical engineering applications, such as the classification of induction motor bearing failures, electrical loads, and industrial robot faults. Despite its development, the performance of the wavelet scattering (WS) network constructed in the MATLAB environment to compute WST coefficients has not been highlighted in the literature so far. In this paper, the properties of the WST feature matrix are examined, and the parameters that have a significant impact on coefficient magnitudes and matrix dimensions are defined. With minimal configuration, a WS network could extract low-variance features from real-valued time series for use in machine learning and deep learning applications. The feature matrix, which contains zero, first, and second-level WST coefficients derived from various power system signal configurations, is constructed to be trained using long short-term memory (LSTM) networks. The simulation results demonstrate the efficacy of the proposed classifier with an accuracy approach of 100%. The MATLAB toolbox has been used to create different signals for the WS and LSTM networks. WST has proven to be a powerful tool for power system disturbance classification.