Imbalanced Source-Free Adaptation Diagnosis for Rotating Machinery

There have been some studies on fault diagnosis in source-free domain adaptation (SFDA) environments, but, currently, all studies assume that the fault types are uniform. When fault diagnosis under imbalanced fault categories is studied, negative migration is exhibited by all existing SFDA methods. In reality, the vast majority of the entire lifecycle of devices is accounted for by the duration of healthy operation, while the fault market accounts for a relatively small portion. To address the problem of both label set drift and domain drift, an imbalanced SFDA (ISFDA) method on bearing fault diagnosis is proposed. In short, a potent source diagnosis model is first generated as the learning foundation using source information. Then, label modification, intraclass aggregation, and interclass alienation are combined for unsupervised learning in the target. Finally, a brand-new category centroids separation loss function is created. The robustness and stability of the algorithm were validated on two public datasets, including fault diagnosis experiments for ordinary bearings and aviation high-speed bearings. The best results are achieved by our method, comprehensively surpassing existing unsupervised domain adaptation (UDA) and SFDA methods. Finally, the powerful diagnostic ability of our method in imbalanced fault data was demonstrated through various visual verification methods.