Early Anomaly Detection in Wind Turbines by Causality-based Graph Attention Networks

Early anomaly detection is a key issue for ensuring reliability, availability and efficiency in Wind Turbines (WTs) operation. Traditional condition monitoring techniques struggle to identify anomalies because of: i) the strong interconnections among the subsystems of a WT, which induce complex spatio-temporal dependencies in the measured signals and can enable the cascading propagation of the effects of anomalies, ii) the variability of the operating conditions, which can affect the anomaly signatures, and iii) the presence of control systems that can mask the anomalies signatures. To address these challenges, this work proposes a novel anomaly detection method employing a Graph Neural Network (GNN) autoencoder. GNN is selected for its capability in modeling the interconnected nature of subsystems and capturing the spatio-temporal dependencies in operational signals. A key contribution is the use of the Granger causality test to generate the GNN functional graph. By leveraging the causal relationships among turbine signals, the obtained graph ensures that the learned topology reflects the underlying physical dependencies rather than spurious correlations, as it may occur in traditional graphs. To handle variability of the operating conditions, graph attention layers adaptively weight sensor relationships, while gated recurrent units process the temporal dynamics. The proposed method is validated on synthetic data simulated using a high-fidelity WT model and applied to SCADA data collected from real WTs. Results demonstrate a 4% increase in recall and a 45% reduction in detection delay compared to state-of-the-art methods.