Rolling Bearing Fault Diagnosis Method Based on Multiple Efficient Channel Attention Capsule Network

In the environment of strong noise, it is very difficult to extract bearing fault characteristics from vibration signals. To solve the problem, this paper proposes a fault diagnosis method based on Multiple Efficient Channel Attention Capsule Network (MECA-CapsNet). Due to diverse scales channel of attention mechanism, MECA-CapsNet can obtain multi-scale channels feature, enhance information interaction between different channels, and fuse key information of diverse scale receptive field. So, our model can effectively abstract the key information of bearing fault characters from noisy vibration signal. To verify the effectiveness of MECA-CapsNet, experiments are carried out on the bearing data set of CWRU. When the signal-to-noise ratio is from 4 dB to −4 dB, the accuracies of MECA-CapsNet are better than typical fault diagnosis methods. Then, T-SNE technology is used to visualize the features extraction process. The visualization result verifies that multiple ECA modules on different scales can effectively reduce noise interference and improve the accuracy of rolling bearing fault diagnosis.