An Explainable AI Approach Using SHapley Additive exPlanations for Feature Selection in Vibration-Based Fault Diagnostics of Wind Turbine Gearbox

Vibration analysis is a proven technique for detecting faults in wind turbine gearboxes. However, the accuracy and efficiency of vibration-based fault diagnostics depend on effective feature selection. This study introduces an explainable AI approach, utilizing SHapley Additive exPlanations (SHAP) to improve the transparency and effectiveness of feature selection. Vibration signals from a turbine gearbox were segmented into equal-width windows from each of which features were extracted. We applied SHAP for feature importance analysis and selected the dominant features to train a support vector classifier. The results show significant performance improvements, with accuracy rising by 2.5%, as well as notable gains in precision, recall, and F1 score. Additionally, training time decreased from 227 to 21 milliseconds. These results highlight the effectiveness of SHAP-based feature selection and recommend its use in vibration-based condition monitoring of wind turbine gearboxes.