Continual Unsupervised Domain Adaptation for Bearing Fault Diagnosis Under Variable Working Conditions

Unsupervised domain adaptation (UDA)-based fault diagnosis models have been intensively studied. UDA focuses on establishing a model that can transfer knowledge from one or more source domains with labeled data to help learn a target domain with unlabeled data. However, in the field of fault diagnosis, an obvious drawback is that UDA ignores the case of multiple target domains. In real scenarios, fault data are distributed in a continuous flow of constantly generated information due to variable working conditions. The collected unlabeled data may be in different domains, resulting in an increase in the number of target domains, which is called domain increments. Continual unsupervised domain adaptation (CUDA), which combines continual learning (CL) and UDA, is proposed to address the problem of domain increments. CL aims to create a deep learning (DL) model that can learn in dynamic environments, similar to humans. The core of CL is to prevent DL from catastrophic forgetting. In this paper, a CUDA fault diagnosis (CUDAFD) method is proposed for bearing fault diagnosis under variable working conditions. In CUDAFD, the maximum mean discrepancy is used to continually adapt the model to new domains, and the idea of prototype learning is introduced to overcome the catastrophic forgetting. Finally, CUDAFD is applied to a diagnosis case of domain increments to verify its effect.