PGDA-Net: Physics-Guided Domain Adaptation for Cross-Condition Bearing Fault Diagnosis via Order-Domain Envelope Descriptors

Rolling-element bearings in transportation drivetrains, such as railway axle-box and traction-motor bearings, are safety-critical, yet fault diagnosis models trained under one operating condition suffer severe performance degradation when deployed under different speeds or loads, due to the frequency migration of fault-characteristic impulses. Existing unsupervised domain adaptation methods align feature distributions in a purely data-driven manner, ignoring the deterministic relationship between bearing fault frequencies and shaft rotational speed, which limits their robustness under large speed gaps. To address this problem, we propose PGDA-Net, a physics-guided domain adaptation network that exploits order-domain envelope descriptors to guide, weight and verify adversarial alignment. Specifically, a fully differentiable physics path computes the resonance band-pass envelope spectrum and soft-pools it with per-sample-rpm Gaussian kernels centered at theoretical fault orders, producing a speed-invariant descriptor. This descriptor is fused with raw-waveform features and drives a physics-weighted MMD term, a data/physics consistency loss, physics-consistent pseudo-labels and an order-center regularizer, forming a multi-level physics-guided adaptation objective. Experiments on the Paderborn multimodal bearing benchmark (leave-one-condition-out, vibration and motor current) and the CWRU cross-load transfer benchmark demonstrate that PGDA-Net delivers its largest and most consistent gains under large speed gaps, where it clearly outperforms strong adversarial, moment-matching and subdomain-adaptation baselines, while also attaining the best mean accuracy among the compared methods on the near-saturated small-gap CWRU transfers. Interpretability analysis further confirms that the learned order attention concentrates on theoretical fault frequencies and that domain discrepancy is effectively reduced after alignment. These properties suggest potential applicability to AI-driven diagnostics and predictive maintenance for transportation systems under variable duty cycles.