Rolling Bearing Fault Diagnosis Using Deep Transfer Learning Based on Joint Generalized Sliced Wasserstein Distance

The big data of rolling bearings for on-site monitoring usually contains very few failure samples and easily affected by noise and monitoring errors, so it is difficult to extract and identify useful fault information in normal samples. In addition, the rolling bearing samples of field test are un-labeled dataset of unknown fault types. If the existing fault diagnosis approaches are directly used for extraction and identification, it is easy to cause misjudgment or missing judgment. To solve this problem, a novel intelligent fault diagnosis approach using deep transfer learning based on joint generalized sliced Wasserstein distances (JGSWD) deep transfer learning is proposed. Firstly, the joint discrepancy between the data from real-case scenarios (DRS) and the data from laboratory equipment (DLE) is minimized by calculating the generalized sliced Wasserstein distances. Following, the marginal and conditional dataset distribution between source domain and target domain is balanced by using the dynamic domain alignment. Then, the top K correlated pseudo labels are calculated for reducing the conditional distribution and improving better transfer capability. Finally, the deep transfer learning from laboratory bearing dataset to field bearing dataset is carried out. The result shows that the proposed JGSWD method can achieve 97.56% fault diagnosis accuracy, which is higher than the other methods. Therefore, it is a practical semi-supervised learning approach for bearing fault diagnosis with small samples.