DiagConTA: Continual Test-Time Domain Adaptation for Industrial Fault Diagnosis Under Dynamic Working Conditions

Dynamic working conditions in industrial systems cause continual distribution shifts in test data that degrade fault diagnosis performance, while vanilla domain adaptation for distribution shifts faces limitations like source data dependence and single-stage adaptation. Continual test-time domain adaptation, an emerging paradigm that integrates continual learning and test-time domain adaptation, has the potential to continually adapt a pretrained model to a series of domains during testing without using source data. In this article, we propose DiagConTA, a novelContinualTest-time DomainAdaptation framework for Industrial FaultDiagnosis under dynamic working conditions. DiagConTA consists of two stages: source domain pretraining and sequential target domain adaptation. In source domain pretraining, we adopt prototype contrastive learning to pretrain the model on the labeled source domain to learn the initial model parameters with better generalization ability. During sequential target domain adaptation, for effective adaptation, we introduce top-$k$pseudolabeling to provide more reliable labels for the unlabeled instances and maximize mutual information to align the distribution shifts without source data. Meanwhile, considering the catastrophic forgetting and the stability-plasticity dilemma in continual learning, we propose stochastic recovery and prototype-guided contrastive alignment, which explicitly and implicitly constrain the update of the model parameters, respectively. Extensive experiments demonstrate the superiority of DiagConTA in achieving robust fault diagnosis under dynamic working conditions.