A Digital-Twin-Enhanced Adversarial Transfer Learning Framework for Fault Diagnosis of Train Axle Box Bearings

Abstract

Intelligent fault diagnosis methods have become essential for ensuring the secure and dependable functioning of modern railway systems. In particular, digital twin technology offers a highly promising approach to overcome the challenge of limited fault data availability. However, current digital-twin-based fault diagnosis approaches for axle box bearings still exhibit several limitations. 1) They rely on a single digital twin model, which limits the coverage of fault types and working conditions of axle box bearings, reducing the generalization ability of the diagnostic model. 2) They rely on cross-entropy and domain adversarial training without explicitly constraining the feature space, resulting in insufficient inter-class feature discriminability and poor generalization across domains. To address the above limitations, a digital-twin-enhanced adversarial transfer learning framework is proposed for fault diagnosis in axle box bearings. First, a unified feature extractor captures shared feature representations from multiple digital twin models, while class classifiers and domain discriminators are adversarially trained to obtain domain-invariant features. Then, a contrastive learning strategy is applied to explicitly structure the representation space, enhancing inter-class separability and intra-class compactness. Finally, hard example sampling prioritizes the most confusing positive and negative pairs near decision boundaries, further improving the discriminability of the learned features. Ablation and comparison experiments are conducted using fault simulation data from train axle box bearings, and experimental results show the effectiveness and advantages in leveraging multiple digital twins and improving diagnostic accuracy.