Bearing Diagnosis Using an Anti-Noise Neural Network Based on Selectable Branch Multiscale Modules and Attention Mechanisms

Aiming at the problem that rolling bearing vibration signals are susceptible to noise interference in the working environment, an anti-noise network (MCAMDN) depending on a selectable branching multiscale convolution and attention mechanism is proposed. Compared to single-branch convolution, the selectable multibranch module (MConv) used in MCAMDN can effectively choose to use branches at different scales due to the change of the feature map size, which ensures that the network can learn more features, and can appropriately reduce the parameters of the network model and the amount of computation. Second, a DConv module similar to the DW convolutional structure is used to further extract the information about the feature map while saving the network model parameters. In addition, a variety of attention mechanisms and regularization methods are used in MCAMDN, which allow the network to automatically focus on the important features of the image during the training process and effectively mitigate the overfitting problem. In evaluating the datasets shared by Case Western Reserve University (CWRU) and Xi’an Jiaotong University (XJTU), experimental findings demonstrate that MCAMDN exhibits strong classification performance and robustness, particularly noise resistance. Therefore, MCAMDN can be utilized for defect diagnosis of rolling bearings in scenarios involving complex noise interference.