Deep Learning-Based Multiswitch Open-Circuit Fault Diagnosis for Active Front-End Rectifiers Using Multisensor Signals

Abstract

Open-circuit switch faults (OCSFs) in power semiconductor switches are caused by wire bonding failures, gate driver malfunction, surge voltage/current, electromagnetic interference, and cosmic radiation. Under OCSFs, the signal characteristics are not excessively high, but prolonged OCSFs risk cascading system failures. This letter presents a comprehensive analysis of various deep neural network (DNN)-based architectures, such as long short-term memory (LSTM) and convolutional neural network (CNN), to diagnose multiclass OCSFs in three-phase active front-end rectifiers (TP-AFRs). A novel multisensor time-series sequence (MTSS) dataset is acquired at 500 Hz, comprising 624 observations from 19 sensor signals for single, double, and triple-switch OCSFs. The intertwining issue in the MTSS dataset is visualized using t-SNE, and the initial experiments with support vector machine (SVM) rendered the highest test accuracy of 93% against k-nearest neighbor, artificial neural network, and decision tree classifiers. Further, our investigations revealed that an architecture with two-layer CNN, one-layer LSTM, and one fully connected layer achieves a competitive testing accuracy of 95.03%, showing an improvement of 2.03% from the SVM classifier, and 7.03% from the one-layer LSTM network. These findings demonstrate the potential of this approach for enhancing reliability of TP-AFRs with the direct application of downsampled raw electrical signals.