Contribution of Artificial Intelligence to Predictive Maintenance of Wind Turbines: A Hybrid Framework Combining Multiphysics Modelling and Explainable Deep Learning

Abstract. Wind energy is a pillar of the energy transition. However, WT subassemblies are exposed to mechanical and electrical failures, resulting in high costs and reduced availability. This paper presents a hybrid framework combining Multiphysics modelling using COMSOL Multiphysics, frequency-domain signal analysis, and a comprehensive CNN-LSTM architecture. First, we produce consistent vibrational and electromagnetic torque signatures across varying conditions. Second, we train deep learning models on transformed signals from FFT and STFT. Finally, we use explainable AI tools, such as Grad-CAM and SHAP, to clarify the model’s decision-making process. The suggested architecture achieves an average classification accuracy of 95.6%, which exceeds traditional ML techniques (SVM and Random Forests). These results show that physics-based modelling and XAI work well together for fault detection in WT, establishing a robust baseline for sustainable maintenance strategies aligned with Net Zero 2050 goals.