Meta Bi-classifier Gradient Discrepancy for noisy and universal domain adaptation in intelligent fault diagnosis

Abstract

Unsupervised domain adaptation fuels knowledge transfer in fault diagnosis, but it depends entirely on a clean source domain and a common label set across domains. Nevertheless, due to the absent knowledge on the label set, the source domain may cover the utmost in classes and unknown classes also come in the target domain, making UDA utterly inapplicable and often impractical. Besides, the ubiquity of incorrect annotations is prone to occur in industrial dataset curations, but it is largely overlooked and left unsolved in UDA. To address these issues, towards a noisy and universal domain adaptation, a Meta Bi-classifier Gradient Discrepancy (MBGD) method is proposed. Specifically, a reliable dual-stage sample selection approach is first initiated to select clean and reusable source samples. Then, to handle private samples on both sides, a coupled divergence criterion is presented to separate target private samples, and a residual bi-classifier with learning constraints is constructed to weaken source private samples. Finally, with a bi-classifier gradient discrepancy, a meta-adversarial domain adaptation is proposed to empower a safe domain alignment on common samples. Comprehensive experiments on various noisy and universal tasks validate the efficacy of MBGD.