Adversarial Domain Adaptation With Dual Auxiliary Classifiers for Cross-Domain Open-Set Intelligent Fault Diagnosis
Bo Wang, Meng Zhang, Hao Xu, Chao Wang, Wei-Yi Yang
Abstract
Although numerous studies on mechanical intelligent fault diagnosis based on the closed-set domain adaptation methods have achieved remarkable success, when there are private classes in the target domain, it is challenging for the model to effectively recognize the private classes. To tackle this issue, we propose an approach of adversarial domain adaptation with double auxiliary classifiers for cross-domain open-set intelligent fault diagnosis. Specifically, the private fault classes in the target domain are automatically identified by the private class classifier, and the shared class alignment is accomplished simultaneously through a weighted adversarial mechanism. Furthermore, the generation of target representations that match the feature distribution of the source domain is enhanced and the negative impact of abnormal samples is mitigated through reweighting and maximizing the discrepancies between the double auxiliary classifiers. Finally, an adaptive overall classification balancing mechanism is designed, and the generalization and accuracy of the model are effectively improved. A considerable number of experimental results reveal that in comparison to the majority of existing methods, the proposed method boasts a higher accuracy rate for fault diagnosis in the open-set scenario and is capable of effectively identifying unknown classes.