A Hybrid Physics-Informed and Deep Adversarial Transfer Learning Network for Cross-Domain Fault Diagnosis of High-Speed Train Bearings

Reliable fault diagnosis of high-speed train bearings remains challenging due to strong noise interference and pronounced domain discrepancies between laboratory test conditions and real in-service operating environments.To address the degradation of diagnostic performance under cross domain scenarios, this study proposes a physics informed adversarial transfer learning framework for intelligent bearing fault diagnosis with unlabeled target data.The proposed method integrates physically guided kinematic features with deep time frequency representations extracted from Short-Time Fourier Transform spectrograms through a dual path feature fusion architecture, enabling the model to capture both intrinsic fault mechanisms and discriminative spectral characteristics.A domain adversarial learning strategy is further employed to reduce feature distribution mismatch between the laboratory source domain and the real world high-speed train target domain, thereby enhancing domain invariance and diagnostic robustness.Experimental validation is conducted using 12,000 labeled laboratory samples and 3,000 unlabeled field samples collected from high-speed train axlebox bearings.The results demonstrate that the proposed approach achieves a diagnostic accuracy of 98.5% on the target domain, significantly outperforming conventional convolutional neural networks and standard transfer learning methods.The proposed framework effectively combines transfer learning with physically guided diagnosis, providing a reliable and interpretable solution for cross domain fault diagnosis of high-speed train bearings under complex operating conditions.