Selective knowledge distillation-based domain adaptation framework towards edge computing fault Diagnosis for high-speed train bogie

Edge deployment of deep learning models for high-speed train bogie fault diagnosis is challenged by computational constraints and cross-domain diagnostic requirements under varying operational conditions. This paper proposes a selective knowledge distillation-based domain adaptation framework (SKDA) that simultaneously achieves model compression and cross-domain diagnosis. The proposed selective knowledge distillation combines Monte Carlo Dropout (MCD) with Kullback-Leibler (KL) divergence, selectively transferring high-quality diagnostic knowledge from the complex teacher to the lightweight student model. A three-branch multi-scale attention module (TMAM) is designed as the teacher network to capture multi-scale fault features and long-range dependencies. Experiments on two bogie bearing datasets show that the proposed method, with a model size of only 28.5kB, improves cross-domain diagnostic accuracy by at least 2.1% compared to existing methods. This provides an effective solution for edge deployment in high-speed train bogie fault diagnosis. • Selective knowledge distillation with Monte Carlo Dropout and KL divergence. • Three-branch multi-scale attention teacher model for fault feature extraction. • 2.1% improvement in cross-domain diagnosis with only 27kB model size.