Fault diagnosis of rolling bearing based on channel and spatial reconstruction networks

Since the fault vibration data collected in the real engineering may be accompanied by noise, traditional diagnostic models are difficult to identify fault categories. To address this problem, a rolling bearing fault diagnosis research method based on channel and spatial reconstruction and progressive convolutional neural networks (CSRP-CNN) was proposed. The model utilized channel and spatial reconstruction convolution （CSConv） to reduce the redundant information of channels and space in fault features, and reduced the complexity and computation to improve the performance; using the convolutional block attention module (CBAM), attention enhancement operation was carried out in the channel and spatial dimensions to make the model pay attention to the important fault feature information; and the progressive convolutional network structure was used in the shallow layer of the network, which would fuse the previous fault feature information with the current input to obtain the richer feature information. The performance of CSRP-CNN was evaluated by two different datasets of Case Western Reserve University (CWRU) and machinery fault simulator magnum (MFS-MG). After the noise and ablation tests, it is verified that CSRP-CNN has strong robustness and the effects of CSConv, CBAM and progressive convolutional neural network (PCNN) on the model noise immunity performance.