Transferable dynamic enhanced cost-sensitive network for cross-domain intelligent diagnosis of rotating machinery under imbalanced datasets

The imbalance is an inevitable problem in mechanical fault diagnostics, as most of the monitored samples for mechanical devices are normal, which results in the decision boundary of the classifier being heavily driven by the dominant class and ignoring the minority class. To settle this problem, a transferable dynamic enhanced cost-sensitive network (TDECN) is proposed in this study. Within this framework, the maximum classifier discrepancy approach is utilized as the backbone, in which sliced 1-Wasserstein discrepancy is exploited to measure the distance between two outputs and detect outlier targets. By doing this, the relationship between the task-specific decision border and the target features during distribution matching is considered. Simultaneously, a dynamic enhanced focal loss (DEL) is devised and embedded into the network. which makes the model pay more attention to error-prone and minor class instances compared with routine focal loss. Finally, extensive diagnostic experiments were implemented to evaluate the effectiveness of the proposed TDECN. The noticeable accuracy improvement demonstrates that our method is superior in resolving imbalanced cross-domain fault diagnosis problems over other approaches.