A cross-domain fault diagnosis method for high-speed train axlebox bearings based on time-frequency mapping and adversarial networks

Abstract The health condition of axlebox bearings in high-speed trains is directly associated with operational safety. However, variations in rotational speed, load, and operating environment induce pronounced domain shift in vibration signals, which undermines the generalization capability of intelligent diagnosis models trained under a single operating condition.To address this issue, this paper proposes a cross condition bearing diagnosis method integrating a CNN BiLSTM hybrid network and multi constraint domain adaptation for cross condition diagnosis of high speed train axle box bearings. The method is verified using the public CWRU and JNU datasets as well as data collected from a 1:2 metro bogie test rig.First, constant Q transform (CQT) is adopted to implement time frequency mapping. Median absolute deviation (MAD) robust normalization and adaptive histogram equalization are applied to improve representation quality. Next, a CNN BiLSTM hybrid network is designed to extract multiscale spatial features and long range temporal dependencies separately. A residual fusion module is used to realize feature complementarity. On this basis, a multi constraint domain adaptive loss function is constructed. The function combines domain adversarial alignment, local maximum mean discrepancy (LMMD), center loss and supervised contrastive loss (SupCon Loss). It reduces cross domain distribution discrepancy and enhances intra class compactness simultaneously. The proposed method achieves high diagnosis accuracy across all eighteen cross domain tasks in experiments.The average accuracy reaches 98.86% and 98.91% for CWRU and JNU respectively. The accuracy of all cross domain tasks on CWRU and JNU remains above 97.50%. The average accuracy on the BJTU RAO dataset from the 1:2 scaled metro bogie test rig is 98.53%. The proposed method achieves obvious performance improvement compared with mainstream approaches including DANN, DAN, DDC and CDAN.