Attention Mechanism-Enhanced Multi-scale Separable Atrous Convolution Method for Motor Bearing Fault Diagnosis

To address the issue of inadequate feature information extraction by single convolutional neural networks in high-noise environments, which leads to suboptimal diagnostic performance, this study proposes an attention mechanism-enhanced Multi-scale Atrous Separable Convolution (MASC) method for motor bearing fault diagnosis. The proposed method begins by performing overlapping sampling on motor bearing data, using the original time-series signals as inputs to the neural network to preserve the integrity of the original signals. Subsequently, a MASC module is designed for feature extraction; the MASC includes an efficient convolutional module with a residual structure to extract features from the input data while maximizing the retention of valuable information. Next, a channel-wise attention mechanism is constructed to perform feature selection on the extracted data. Finally, after mapping through a fully connected layer, these features are input into a softmax layer for fault identification. Testing and validation were conducted using the Case Western Reserve University (CWRU) dataset. At a signal-to-noise ratio (SNR) of 0 dB, the proposed method achieved a fault diagnosis accuracy of 98.50%, further demonstrating its strong diagnostic capability in high-noise environments.