Wind turbine anomaly detection and identification based on graph neural networks with decision interpretability

Condition monitoring of wind turbines with supervisory control and data acquisition (SCADA) data has attracted increasing interest due to its great benefits in improving operation reliability and reducing unplanned downtimes of wind turbines. However, existing SCADA data-based studies focus primarily on anomaly detection, and few studies have attempted to identify the underlying causes of the anomaly. To this end, this paper proposes a new system-level wind turbine anomaly detection and identification method based on an emerging graph neural network with decision interpretability (DIGNN). A correlation-based GNN is first used to capture complex inter-sensor correlations in SCADA data. More importantly, a decision interpretability module is designed to further analyze anomaly causes and effects through a twp-step global and local anomaly decision process. It can greatly improve the interpretability of the model and provide more accurate and meaningful decision results for maintenance purposes. The effectiveness and robustness of the proposed model were verified by four fault cases with SCADA datasets from a real wind farm. The experimental results demonstrated that the proposed model can provide earlier warning of anomalies with lower false alarm rates and accurate anomaly identification with good interpretability, providing valuable help for field maintenance.