Gradient-Based Interpretable Graph Convolutional Network for Bearing Fault Diagnosis

With the advent of the era of big data, intelligent algorithms such as deep learning have played an increasingly important role in the operation and maintenance of manufacturing equipment, but the promotion of deep learning in many fields has been hindered due to its “black box” nature. To deal with such a problem, a gradient-based interpretable graph convolutional network (GIGCN) is proposed for bearing fault diagnosis, which analyzes the interpretability of the fault diagnosis model from the perspectives of the time domain and frequency domain. The proposed GIGCN was used to achieve high accuracy in bearing fault diagnosis tasks and support visualize interpretable signals analysis process. The visualization signal interpretability analysis indicates that the GCN network can judge the fault type by the signal characteristics at the faulty frequency and rotation frequency. The experiment results show that the proposed method can learn features and find the location of the characteristic frequencies with interpretable meaning from vibration signals under different work conditions, and provides a reliable explanation for the application of graph convolutional network.