A Globally Interpretable Convolutional Neural Network Combining Bearing Semantics for Bearing Fault Diagnosis

Abstract

Bearing fault diagnosis is crucial for maintaining the safety of industrial systems. With the massive data collected by the Industrial Internet-of-Things technology, deep learning (DL)-based end-to-end models have been extensively utilized in bearing fault diagnosis. However, their limited interpretability poses challenges to their reliability, hindering further advancements in the field. To address this interpretability issue, we propose a globally interpretable convolutional neural network (CNN) combining bearing semantics for bearing fault diagnosis. Specifically, the physical semantics of bearing signals are first constructed based on the fault characteristic frequency (FCF). Based on this bearing semantics, a novel bearing semantic embedding method is proposed to enhance the interpretability of convolutional layers. Moreover, a globally interpretable network (GINet) structure is crafted to ensure that the bearing semantics are visible throughout the entire network. Experimental results on two datasets demonstrate that the network’s performance remains comparable to the benchmark method while achieving global interpretability. This network also exhibits improved noise robustness, proving the effectiveness of semantic embedding. In addition, since this network is an interpretable modification of the basic CNN, it is not limited to bearing fault diagnosis. Theoretically, with the appropriate semantics, it can also be applied to other signal-based fault diagnosis tasks.