Spectrum-aware feature disentanglement for domain-generalized fault diagnosis in rail transit systems

Rail transit electromechanical equipment operates extensively under frequently switching, non-stationary variable conditions, which poses a severe generalization challenge for deep learning-based fault diagnosis models. Existing cross-domain diagnostic methods primarily focus on the global alignment of feature distributions, struggling to resolve the feature entanglement between essential fault signatures and operational interferences. To address this issue, this paper proposes a spectrum-aware disentangled generalization network (SDGN). Based on the concept of orthogonal separation in the feature space, this framework constructs a dual-stream disentanglement architecture constrained by statistical independence, aiming to effectively filter out redundant domain-related interferences and isolate genuine fault-relevant features. Specifically, SDGN introduces a feature enhancement module that integrates frequency-domain priors with residual aggregation to adaptively distill critical frequency band information guided by physical mechanisms. Furthermore, to enhance the model’s robustness against unseen operating conditions, a multi-scale style perturbation mechanism is designed within the feature space. By simulating non-stationary domain shifts to expand decision boundaries, it forces the network to learn robust representations that are insensitive to variations in operational styles. Given that rolling bearings are the core transmission components and their fault mechanisms are highly correlated with the entire system, they are selected for proof-of-concept validation. We comprehensively evaluate SDGN on the public PU dataset and a laboratory-collected Mechanical comprehensive diagnostic simulation platform dataset. The results demonstrate that SDGN significantly outperforms state-of-the-art methods across diverse cross-domain tasks. Specifically, our model achieves average accuracies of 93.10%–94.26% on the two datasets. These metrics quantitatively confirm its exceptional reliability for practical rail transit applications.