Fault diagnosis of high-speed train bogie bearings under domain shifts via an encoding-matching-interactive framework

Fault diagnosis of bogie bearings in high-speed trains is challenged by limited labelled data and complex domain shifts caused by varying operating conditions, which severely degrade cross-domain generalization. An encoding–matching–interaction (EMI) transfer framework is proposed to address cross-condition and cross-platform domain shifts in high-speed train bogie-bearing diagnosis. In the encoding stage, a shared convolutional encoder learns hierarchical fault representations, while class-discriminative regularization and scale-aware feature fusion enhance feature separability and transferability. In the matching stage, multilayer higher-order moment matching (mUoMM) and multiclassifier-guided adversarial domain alignment (mGADA) perform complementary marginal and conditional alignment, respectively. These adaptation objectives are further coordinated through an interaction mechanism that dynamically balances transfer objectives according to transfer states. EMI is validated on multiple public machinery benchmarks and an engineering-scale high-speed train bogie-bearing dataset, covering cross-condition, cross-device, and cross-scenario transfers. Compared with eight representative baselines, EMI achieves consistently superior performance, outperforming the best baseline by 3.97–5.43 percentage points across cross-device and engineering-scale HST transfer tasks. Robustness under additive noise during testing (SNR = 12, 8, and 4 dB) is demonstrated, with the average accuracy degradation limited to 0.9 pp. Extensive evaluations, including ablation studies, sensitivity analyses, and statistical significance analysis, further validate the effectiveness and reliability of EMI.