Fault Feature Assessment Method for High-Voltage Circuit Breakers Based on Explainable Image Recognition

The current deep learning (DL) model for fault diagnosis of high-voltage circuit breakers (HVCBs) lacks explainability. It is difficult to further analyze the cause and mechanism of faults, which could provide little help for the maintenance and optimization design of HVCBs. To address this problem, a fault feature assessment method of HVCBs based on explainable image recognition is proposed to realize a quantitative analysis of faults. First, the vibration signals of HVCBs are preprocessed by continuous wavelet transform (CWT). The time-frequency diagrams of CWT are segmented by the travel curve to obtain the action sequence of the HVCB. Then, Shapley additive explanations (SHAPs) explain the deep residual network ResNet to obtain the feature importance distribution maps. Through the feature importance distribution map, accurate fault location and time traceability can be realized, and the frequency-domain features of the fault can be directly visualized from the distribution degree of the feature importance. The fault evaluation factor (FEF) is proposed to quantitatively study the time-frequency–amplitude comprehensive difference between the fault state and the normal state of the circuit breaker.