Support Vector Machine–Artificial Neural Network Hybrid Model for Condition Monitoring of Electrical Machines

Abstract

Condition monitoring of electrical machines has garnered continuous study attention for more than thirty years, especially concerning spinning electrical machinery. Previously, operators meticulously monitored machine performance; however, this practice has diminished with the introduction of rapid-response digital protection mechanisms. This study, grounded in the author’s expertise and current literature, concentrates on online monitoring techniques, with minimal attention to variable speed drives and a preference for conventional machines over contemporary topologies. Data preparation utilises the DWT to efficiently identify and extract unique signal patterns. SVM-ANN are employed for classification, exemplified by a binary decision problem of fault versus no fault classification in early-stage virtual screening. The findings validate that DWT is an effective instrument for feature extraction in condition monitoring. Moreover, SVM exhibits superior performance with diminished standard error in comparison to ANN. SVM-ANN model consistently surpasses alternative methods with an accuracy of 96.06% regardless of training data volume, neural network techniques, or descriptor types. The amalgamation of wavelet-based methodologies with machine learning models demonstrates significant efficacy in the Condition Monitoring of Electrical Machines, facilitating dependable fault detection and enhanced predictive capabilities, while underscoring the necessity for further research on contemporary machine designs and variable speed applications.