Vibration-based multi-class fault severity classification of wind turbine gearboxes using explainable machine learning

• Proposes a novel framework for automatic fault severity classification in wind turbine gearboxes. • Uses unsupervised K-means clustering to identify early, moderate, severe faults without prior levels. • Achieves high performance, with Random Forest reaching 99.95% accuracy for ring gear and bearing faults. • Employs SHAP-based explainable AI to interpret feature contributions and enhance model transparency. • Reduces preprocessing needs and supports real-time, data-driven maintenance. This study presents a comprehensive fault severity classification framework for wind turbine (WT) gearboxes using vibration data from the National Renewable Energy Laboratory (NREL). The objective is to identify the extent of damage, categorized as early, moderate, or severe, based on analysis of vibration signal windows from sensors mounted on the gearbox. The methodology focuses on two fault locations, ring gear and bearing, where specific sensors (AN4 and AN8) were selected after examining variability within the damaged dataset. K-means clustering was employed to group the damaged data into three distinct severity levels, followed by cluster interpretation using frequency domain characteristics. Supervised machine learning models were subsequently trained using the labeled data to automate severity classification. Among the models evaluated, the Random Forest classifier achieved approximately 99.95% accuracy for both ring gear and bearing severity classification. However, this performance should be interpreted with caution, as it is based on a laboratory dataset with limited operational variability. To ensure transparency and interpretability, robustness and the SHAP (SHapley Additive exPlanations) techniques were applied to quantify and visualize the impact of individual features on the model’s predictions. The proposed approach enables automated multiclass fault severity identification in multiple fault locations supporting SHAP and improved system performance. The proposed approach enables automated multiclass fault severity identification in multiple fault locations supporting SHAP and improved system performance.