PT-LGBM: an end-to-end interpretable cross-domain model for high-speed train bearing fault diagnosis under class imbalance
Year: 2026
Type: article
Abstract: Generalizing the high diagnostic accuracy achieved under laboratory conditions to high-speed train bearings operating in real-world environments is critical. Existing methods remain limited in their ability to address fault-class imbalance, domain shift, and insufficient interpretability simultaneously, leaving the reliability of cross-domain diagnosis inadequately validated. This study proposes PT-LGBM, an interpretable cross-domain fault diagnosis model for high-speed train bearings, comprising four components: PICFS, a geometry-partition-based intelligent sampling method; TFRO-Align, a cross-domain feature-alignment method; an NSGA-II-optimized LightGBM classifier; and an interpretability module integrating t-SNE and SHAP. PICFS reconstructs minority-class samples near decision boundaries to reduce classification bias and boundary blurring caused by conventional interpolation. TFRO-Align combines feature reweighting, orthogonal mapping, and optimal-transport constraints to reduce discrepancies between the CWRU laboratory source domain and the in-service train target domain. The aligned features are classified using the optimized LightGBM model, while t-SNE and SHAP examine the transfer process and diagnostic decisions. On the labeled CWRU source-domain dataset, PT-LGBM achieved 94.12% accuracy, 93.45% Macro-F1, and 100% recall for the normal class. On 16 in-service target-domain records whose labels were blinded during the modeling stage, PT-LGBM achieved an Accuracy of 0.750 (12/16), providing a traceable analytical pathway from laboratory diagnosis to the in-service train environment.
Source: Scientific Reports
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7204857260
