🤖 AI Summary
To address the challenge of cross-device fault diagnosis for wheelset bearings in heavy-haul freight trains under variable rotational speeds and scarce fault samples, this paper proposes an interpretable, label-free domain adaptation method. We innovatively design a modulation-aware differentiable short-time Fourier transform (STFT), integrate physical constraints of rotating machinery to construct a balanced spectral metric, and jointly leverage physics-informed neural networks (PINNs) and spectral-domain adversarial alignment to achieve rotation-speed-invariant feature learning. The method ensures both time-frequency interpretability and strong cross-device generalization. Experimental results demonstrate an average diagnostic accuracy improvement of 12.7% across multi-condition, multi-sensor datasets, and achieve a 96.3% F1-score under few-shot settings—substantially outperforming conventional STFT-based approaches and state-of-the-art deep transfer learning methods.