Semi-Supervised Contrastive Domain Adaptation Network for Fault Diagnosis of Rotating Machinery Under Cross-Working Conditions

Existing domain adaptation (DA) methods, which focus on realizing the class-level alignment for cross-domain features to solve the problem of fuzzy classification boundaries in feature learning process, however, it is difficult to deal with samples located at the classification boundary in this way, resulting in diagnosis performance being limited. To solve the problem, this article proposes a semi-supervised contrastive DA network (SSCDAN) for realizing the cross-working condition rotating machinery fault diagnosis method. Specifically, an in-domain semi-supervised contrastive learning (SSCL) strategy is designed, with the supervision of class information, which guides the discriminative learning of different classes in each domain to eliminate the fuzzy classification boundaries, and facilitates the DA of cross-domain features which utilizes the local maximum mean discrepancy (LMMD). Meanwhile, the negative impact from poor-quality target domain pseudo-labels on SSCL and DA is mitigated by dynamically limiting the associated contrastive learning loss and DA loss gain, and introducing domain adversarial. Finally, the effectiveness of SSCDAN is validated by the ablation and comparison experiments on the Paderborn University (PU) and wind turbine gearbox (WTG) datasets. Compared with deep subdomain adaptation (DSAN), SSCDAN improves the overall average accuracy by 22.43% and 7.36% on the cross-working condition diagnosis tasks of PU and WTG datasets, respectively, and outperforms other popular DA methods.