An interpretable anti-noise network for rolling bearing fault diagnosis based on FSWT

Rolling bearing fault diagnosis based on deep learning has low accuracy under strong noise conditions and weak interpretation of the diagnosis results, which reduces the trust in its industrial applications. An Efficient Multi-Scale Convolutional Neural Network (EMSCNN) with anti-noise based on visualization methods of interpretability is proposed to solve the questions. First, an improved visualization method—Smooth Global Gradient Class Activation Mapping (SGG-CAM) is proposed to analyze the anti-noise ability of modules. Then, SGG-CAM is used to analyze the anti-noise mechanism of the Multi-Scale Dilate (MS-D) module and Residual Channel Attention (RCA) module from the perspective of interpretability. Meanwhile, the EMSCNN network based on the MS-D module and RCA module is established. Besides, Frequency Slice Wavelet Transform (FSWT) is used to generate time–frequency images for enriching the sample information. Experimental results on the bearing datasets show that the presented method is more accurate than other methods under strong noise.