Deep Transfer Learning Method for Bearing Fault Diagnosis Under Varying Operating Conditions

Bearing fault diagnosis based on deep learning have developed rapidly in recent years. End-to-end deep learning based fault diagnosis, which can automatically extract fault features from massive data, greatly reduced manual involvement. As bearings typically operate under varying conditions, which leads to distribution discrepancies in the collected data set, the accuracy and generalization performance of the deep learning network deteriorate. To address this issue, a Multi-Scale Residual Attention network combined with the Transformer encoder (MSRA-Transformer) is proposed which can learn the data feature automatically. Then an improved hydrid domain adaptation method is added to the proposed network to minimize the distribution discrepancy between the source and target domain data. Experimental results demonstrate that the proposed method achieves higher accuracy and better generalization under varying operating conditions, resulting in better fault diagnosis performance.