Research on a wind turbine health assessment model integrating mechanistic knowledge and deep learning

To enhance the accuracy and physical consistency of wind turbine operational condition assessment, a health evaluation model integrating mechanism knowledge and deep learning has been developed. The research design employs a CNNBiLSTM backbone network for feature extraction, integrating a physical information embedding mechanism, a physical constraint loss function, and a simulation-based pre-training strategy to achieve collaborative modeling of temporal features and structural mechanisms. By incorporating physical variables such as main shaft torque and aerodynamic power coefficient, a multi-layer perceptron output structure oriented toward health scoring is constructed, with multithreshold classification labels set to enable hierarchical state recognition. Experimental results on measured data from the Danish NREL wind farm and coastal wind farms demonstrate that this method significantly outperforms traditional machine learning and mechanism-agnostic deep learning models in metrics such as MSE, MDR, and physical consistency scores. It exhibits strong predictive capabilities and engineering deployment value.