Abstract: Fault diagnosis of rolling bearings is related to the safety and stability of equipment operation.Traditional convolutional neural network is used to diagnosis malfunctions,whose model arithmetic is too large and prone to overfitting phenomenon.It caused some problems such as poor diagnostic accuracy,end-to-end modeling with poor trustworthiness and so on. In view of this,it was supposed to propose a rolling bearing fault diagnosis method based on improved Symmetrized Dot Pattern (SDP) and FasterNet-GCAM network. Firstly,the one-dimensional vibration signals were processed by wavelet thresholding noise reduction,and then fed into the SDP method optimized by the Pearson′s image correlation coefficient method to generate the SDP images. FasterNet network by adding the idea of partial convolution was constructed into an improved SDP-FasterNet model for further feature extraction to complete the classification and diagnosis of different faults in the bearing. Lastly,in order to verify the credibility of the model′s decision-making in the process of image recognition,Gradient-weighted Class Activation Mapping (Grad-CAM) was combined with the FasterNet network,which can highlight the important parts of the SDP graph which are relevant to decision making. The experimental results show that the proposed method has faster convergence speed and stronger robustness than other methods. And the results′ diagnostic recognition accuracy had achieved 99.20 %. It improved the interpretability and credibility in the diagnostic process,which provides a diagnostic model with good feasibility and robustness in the field of fault diagnosis.

Key words: rolling bearing, fault diagnosis, FasterNet network, partial convolution, gradient-weighted class activation mapping, Symmetrized Dot Pattern (SDP)