106 / 2025-04-20 11:23:33

Domain Knowledge Embedded Heterogeneous Graph Learning for Wind Turbine Fault Diagnosis

wind turbine,knowledge-data fusion,Heterogeneous Graph Representation Learning,Fault Diagnosis

Condition monitoring of wind turbines (WTs) is critical for enhancing reliability and reducing maintenance costs. Existing data-driven approaches often neglect the multilevel heterogeneity among subsystems (e.g., rotor vs. generator systems) and sensor signals (e.g., linear vs. nonlinear correlations), leading to suboptimal fault diagnosis. To address this, we propose HGLKEWT, a novel framework integrating domain knowledge with heterogeneous graph learning. First, a multilevel heterogeneous graph is constructed by embedding prior knowledge of physical topology and functional dependencies, where nodes represent subsystems/sensors and edges encode hierarchical interactions (e.g., mechanical couplings and causal relationships). Second, we design the Heterogeneous Graph Representation Learning (HeteroGRL) module, which employs meta-relation-aware attention and edge-type-specific message passing to jointly model node- and edge-level heterogeneity. Experiments on four real-world WT datasets demonstrate HGLKEWT's superiority, achieving an average accuracy of 97.17%. Ablation studies validate that statistical feature engineering improves computational efficiency by 43.75%. The results highlight the importance of fusing domain knowledge with graph learning for interpretable and robust fault diagnosis in complex industrial systems.