A frequency mask and decoupling max-logit based XAI method to explain DNN for fault diagnosis

Abstract

Recently, various deep neural network (DNN) models have been proposed for fault diagnosis. Owing to the black-box nature of the DNN, Diagnosis results are unexplainable. Therefore, explainable artificial intelligence (XAI) methods are required. However, it is difficult for existing XAI methods to separate fault and irrelevant features because the fault features are instantaneous. To address this issue, a frequency mask and decoupling max-logit-based XAI method (FM-Explainer) is proposed to explain the DNN for fault diagnosis. Because the fault features can be well represented in the frequency domain, the proposed method optimizes a mask on the frequency domain of the input to identify the fault features. In addition, to avoid unreliable explanations caused by out-of-distribution (OoD) data, a regularization is designed based on decoupling max-logit, and the spatial penalty is used, which ensures that no irrelevant features remain in the explanation. Extensive experiments are carried out to verify the effectiveness of the proposed method using five quantitative evaluation metrics: Insertion/Deletion, Sensitivity-N, and Degradation. The results show that the FM-Explainer outperforms existing methods, and explanations by the FM-Explainer are consistent with the fault characteristic frequency. This indicates that the FM-Explainer is effective in precisely identifying fault features.