A pseudo-labeling training-based progressive multi-scale adaptive network for fault localization of high-speed train bogies

The high-speed train (HST) bogie, as the sole mechanical connection between the train body and the track, endures diverse operational loads and forces, making it crucial to minimize abnormal vibrations and ensure safety. Existing intelligent diagnostic models, typically trained on single-speed datasets, often fail to generalize across varying speeds due to data distribution shifts. To address this, this paper proposes a progressive multi-scale adaptive network (PMAN) for fault localization under varying speeds with limited labeled data. PMAN’s innovations include a multi-scale convolutional adaptive network (MCAN) that extracts cross-scale features and uses dynamic adversarial attention network (DAAN) for cross-speed knowledge transfer, and a progressive pseudo-labeling strategy that integrates unsupervised and semi-supervised learning to screen reliable labels and refine training. Experimental results on SIMPACK-simulated HST datasets show PMAN achieves an average accuracy of 93.7%, significantly outperforming existing methods. This highlights its potential as a robust solution for HST bogie fault localization under varying operational speeds.