Two-stage adversarial collaborative diagnosis framework via transferable perturbation mechanism for HST bogies with long-tailed data

Abstract

Deep learning models for high-speed train (HST) bogie transmission components fault diagnosis frequently suffer from severe performance degradation under unseen operating conditions due to domain shift and long-tailed data distribution. This paper addresses the challenging single-domain generalization (SDG) scenario by proposing a novel transferable diagnosis framework, whose core is a dual-stream teacher-student model (2s-ACNet). The primary stream incorporates a learnable perturbation mechanism (LPM) module, which proactively simulates potential domain variations by perturbing shallow-layer feature statistics, thereby expanding the feature distribution. The secondary stream integrates the superiorities of convolution and RelPos-multi-head attention to enhance the extraction of robust semantic features. The training process is strategically divided into two phases. The first stage aligns the teacher-student models on the source domain using covariance matrix alignment and self-supervised contrastive losses. The second stage fine-tunes to adapt to target domain distribution variations according to perturbing feature statistics. Comprehensive experiments on HST long-tailed fault data across varying loads and speeds indicate that proposed framework outperforms the-state-of-art SDG methods in accuracy, recall, and F1score, demonstrating strong potential in industrial diagnosis applications.