DACL-IDA: a dynamic alignment and compactness learning framework for imbalanced domain adaptation in bearing fault diagnosis

Abstract

Existing unsupervised domain adaptation methods for bearing fault diagnosis under varying operating conditions generally assume identical class priors across domains, overlooking the label shift that arises in practice, where healthy-state samples far outnumber fault samples and fault-type frequencies vary with speed, load, and operating history. When feature shift and label shift coexist, prematurely introduced high-intensity alignment can severely degrade minority-class discriminability. We propose DACL-IDA (Dynamic alignment and compactness learning for imbalanced domain adaptation), built on an alignment-scheduling principle rather than a new alignment loss: global adversarial alignment is delayed until classification warmup and black-box shift estimation (BBSE) have stabilized, and is kept restrained thereafter. The BBSE estimate both re-weights the source loss and drives the scheduler, supported by intra-class compactness regularization and long-tailed optimization. A diagnostic experiment separating global adversarial from local class-conditional alignment shows that sustained class-conditional matching (local maximum mean discrepancy) progressively degrades the classifier under severe (reversed-prior) label shift; the final configuration therefore retains only the scheduled global adversarial alignment. On the Paderborn University dataset, under a strictly leakage-free protocol (fixed last-epoch reporting, no target-label model selection; 4 protocols × 12 transfer tasks × 5 seeds), DACL-IDA attains an overall mean balanced accuracy of 89.63% versus 81.30% for the strongest published competitor, DIDA (a deep imbalanced-domain-adaptation framework), winning three of the four protocols by more than 13 percentage points each (paired-t p ⩽ 5.23 × 10−4); on the remaining protocol (I2B), where scheduled alignment is counter-productive, it underperforms both DIDA and the source-only baseline—a failure reported and analyzed rather than omitted. DACL-IDA is also the most convergence-stable method in the stability comparison: outside I2B, its gap between best transient and final-epoch accuracy stays within 2.4 pp, versus 8.2–11.2 pp for DIDA. Cross-dataset experiments on the Case Western Reserve University dataset, a second backbone, imbalance ratios up to ρ = 50, and injected measurement noise support the reported behavior.