Environment-Aware graph relational reasoning for interpretable and generalizable mechanical transmission system distributed fault diagnosis

In recent years, numerous fault diagnosis models have been developed to monitor the health status of mechanical transmission systems under dynamic environments. However, these models generally perform point-to-point monitoring of individual components, focusing on a local perspective while overlooking the coupling relationships among multiple components. Faults originating in a single component can propagate to adjacent components through vibration transmission, which may ultimately lead to misdiagnosis or misinterpretation. To address this issue, this paper proposes an environment-aware graph relational reasoning framework based on a discover-evaluate-refine paradigm, aiming to achieve comprehensive system-level health monitoring of mechanical transmission systems. The framework constructs stable relational subgraphs by identifying the significance of sensors in the diagnostic decision process and capturing collaborative signal variations among sensors. Samples from different source domains are then used to perturb each other, simulating variations in working conditions and evaluating the robustness of subgraph structures. This assessment provides feedback that guides the refinement of subgraph discovery, ensuring the model effectively captures environment-invariant correlations. Extensive experiments conducted on a self-built experimental platform, a high-speed train, and a metro train bogie demonstrate the superiority of the proposed method. Visualization of the relational subgraphs provides interpretability support for the diagnostic results of the model. Our code and dataset are publicly available at: https://github.com/CHAOZHAO-1/EAGRR.