Multiscale module with attention for bearing fault diagnosis under cross working conditions

The proper functioning of aerospace bearings is essential for the safe and stable performance of modern aircraft. In practical engineering applications, bearing vibration data are often distributed differently due to different working conditions. There fore, the model built by deep learning in one condition (source domain) cannot be directly applied to another condition (target domain). To overcome the above challenges, a novel model is introduced for bearing fault diagnosis. This model uses the share multiscale module with attention mechanism to collect the rich fault feature information directly from the original vibration signals. T hen the domain adversarial strategy is used to help the network to extract the invariant and discrimination features of the source and target domain data. Finally, local maximum mean discrepancy metric is constructed to reduce the subdomain distributional differences between the two domains. The proposed method is verified on Jiangnan University dataset and contrasted with other transfer learning methods. The experimental results indicate that the average diagnostic accuracy of the proposed method is more than 99.5%, thus substantiating its effectiveness.