Unsupervised dynamic domain symmetric adversarial networks for cross-domain fault diagnosis

Unsupervised domain adaptation seeks to develop classifiers for unlabelled target-domain samples. While existing methods primarily reduce domain divergence through domain-level distribution alignment, they often overlook the critical balance between domain invariance and category discriminability. To address this limitation, we propose the Dynamic Domain Symmetric Adversarial Network (DDSAN) – a dual-classifier paradigm that integrates two synergistic innovations: (1) category-level and domain-level confusion losses that jointly align feature distributions while preserving class separability and (2) an adaptive triplet loss with uncertainty-based pseudo-label refinement for robust cross-domain matching. Furthermore, the entire framework employs a unified optimisation strategy featuring: a reinforcement attention mechanism optimised via policy gradient to maximise diagnostic accuracy, coupled with an adversarial domain alignment loss that minimises feature distribution discrepancies. Experimental validation on bearing and gearbox datasets demonstrates DDSAN’s superior performance, achieving state-of-the-art accuracy across variable operational conditions.