Adaptive spatio-temporal graph neural networks for collaborative fault diagnosis of wind turbine drivetrain systems using multi-source monitoring data

Abstract Wind turbine drivetrain systems operate under complex conditions, and their fault diagnosis faces challenges in multi-source heterogeneous data fusion and spatio-temporal dynamic coupling modeling. Existing methods struggle to effectively capture the dynamic evolution of sensor network topology as the system state changes. This research proposes a collaborative fault diagnosis framework based on adaptive spatio-temporal graph neural networks (ASTGNNs). The framework contains two core modules: an attention-based adaptive graph construction module that dynamically learns sensor network topology from raw monitoring data, and a spatio temporal collaborative learning module that simultaneously captures spatial propagation patterns and temporal evolution of fault signals on this dynamic graph. Experiments on a public bearing fault dataset show that ASTGNN achieved 95.64% diagnostic accuracy, improving 1.77% points over Graph WaveNet. In early fault detection (0.007 inch defect), ASTGNN achieved 91.82% accuracy, 4.48% points higher than comparative methods. The model maintained performance fluctuation within 0.27% points across different load conditions and retained 87.68% accuracy at 5 dB signal-to-noise ratio. These results preliminarily validate the feasibility of the proposed framework.