Explainable AI in Wind Turbine Fault Detection

This study presents an explainable data-driven framework for wind turbine predictive maintenance, with Explainable Artificial Intelligence (XAI) as its core contribution to improving transparency, trust, and decision support in condition monitoring. An XGBoost model is developed to predict generator non-drive-end (NDE) bearing temperature using Supervisory Control and Data Acquisition (SCADA) data, with anomalies and early failures identified by analysing model residuals. An expert-guided preprocessing strategy was applied, using maintenance logbook records to identify and remove outliers from the SCADA data while maintaining a high data retention rate. Detected anomalies are aggregated into a daily Anomaly Operation Index (AOI), whose smoothed cumulative trend provides a sensitive indicator of early failure development. The generalisability of the framework is demonstrated by successfully applying a model trained on a healthy turbine to a faulty turbine without retraining, achieving reliable anomaly detection. To enhance transparency and support maintenance decision-making, post-hoc Explainable Artificial Intelligence (XAI) techniques are systematically integrated into the framework. Permutation Feature Importance (PFI) and SHapley Additive exPlanations (SHAP) are employed to interpret model predictions and identify physically meaningful failure indicators. The results confirm that the proposed framework enables accurate early fault detection, transforming predictive outputs into an explainable, actionable health-monitoring tool that supports condition-based maintenance planning.