A Structure-Guided Fault Localization Method for High-Speed Train Suspension Monitoring

The dynamic stability of high-speed trains is critically dependent on the health condition of the suspension system, which directly affects operational safety. This imposes strict requirements on the reliability and robustness of fault diagnosis methods. To address the challenge of reliable vibration-based monitoring under structural complexity and potential sensor failures, this work proposes a structure-aware framework for suspension fault analysis in high-speed trains. The framework first performs physically informed directional alignment of vibration sensors across different locations and employs a shared feature extraction mechanism to obtain consistent representations, thereby enhancing the normalization and comparability of structural features. Subsequently, a condition-dependent sparse structural graph is dynamically constructed from sensor observations to capture latent intersensor couplings. This graph is then used to guide a structure-informed fusion and dispatch mechanism based on a mixture-of-experts (MoE) network, enabling the model to adaptively respond to variations in structural dynamics and sensor conditions. The proposed method is capable of operating without prior knowledge of mechanical configurations or component parameters, while supporting interpretable modeling of dynamic structural behavior. Moreover, it enhances the robustness of the monitoring system under sensor faults and signal loss scenarios, enabling reliable decision support for condition-based maintenance. Experimental results on a high-speed train suspension fault dataset demonstrate that the proposed method achieves both high diagnostic accuracy and strong physical consistency. Compared to commonly used vibration-based diagnosis frameworks, it offers improved robustness and structural coherence under degraded sensing conditions, validating its effectiveness as a measurement-oriented approach for intelligent monitoring in railway systems.