Instance adaptive multisource transfer for fault diagnosis of rotating machinery under variable working conditions

Abstract

Recent years have witnessed the successful development of using knowledge transfer strategies to tackle cross-domain fault diagnosis problems. Compared with the traditional single-source transfer approach, the multisource transfer approach has shown its advantages and gradually become a research hotspot. However, most existing multisource transfer methods mainly focus on the elimination of domain distribution discrepancies and neglect the resulting degradation of diagnostic performance. In addition, more common scenarios (especially in cross-working condition diagnostics), namely the absence of domain labels and the presence of category shifts, are not considered, which leads to severe limitations in the application of existing methods. To this end, this paper proposes an instance adaptative multisource transfer approach. It improves the model's learning ability through instance-adapted feature extraction and training guidance. Specifically, an input-dependent dynamic feature encoder is elaborated to achieve statistically adapting domains by adapting instances, thereby mitigating domain conflicts and lifting the constraints that require domain labels. Furthermore, considering the adverse influence of negative instances on model adaptation, the instance distillation and weighted training strategy is designed to address it and tackle the category shift problem. Experimental results of extensive cross-domain diagnosis tasks built on two datasets and comparison with multiple state-of-the-art methods validate the proposed method's effectiveness and superiority.