An Analysis Method for Interpretability of Convolutional Neural Network in Bearing Fault Diagnosis
Liang Guo, Xi Gu, Yaoxiang Yu, Andongzhe Duan, Hongli Gao
Abstract
With the rapid development of deep learning techniques, bearing fault diagnosis has progressively shifted from knowledge-based methods to intelligent model-based methods. The convolutional neural network (CNN), due to its advanced feature extraction ability of vibrational signals, has achieved promising results in bearing fault diagnosis. However, the working mechanism of CNN and the learned high-order features is still difficult to comprehend. Despite some efforts have made to understand the mechanism of CNN, most of their attention is paid on machine vision instead of fault diagnosis. Due to insufficient understanding and validation for its working mechanism, how the CNN process bearing signals remains opaque to researchers. Therefore, this article develops a new method for interpreting CNN in bearing fault diagnosis from a time–frequency domain perspective. In the time domain, the focus locations of the model are obtained by the application of the gradient-based class activation mapping (Grad-CAM) technique. The working mechanism of CNN is well studied by the gradient-ascent based kernel visualization technique in the frequency domain. The proposed method is verified through a series of experiments on two different datasets. The experimental results are further concluded and discussed, which improves the interpretability of CNN in bearing fault diagnosis.