An Online Torque Prediction Method of PMSM Using PINNs

Torque prediction is crucial for the control and safe operation of electric motors. Compared to offline methods, online identification methods can adapt to dynamic changes and individual variations in motor characteristics in real time, but they face challenges such as rank deficiency and the failure of physical equations. Data-driven methods, which do not rely on specific mathematical models, are better suited to address these issues. However, the lack of torque feedback during actual motor operation can render supervised learning ineffective. This paper proposes a method that combines offline learning and online updates using Physics-Informed Neural Networks (PINNs). The method enables online learning of motor parameters through real-time data and constructs a torque output model to support more efficient and safe online systems. Compared to least-squares methods based on physical formulas, the PINN approach demonstrates greater accuracy and dynamic adaptability in real-time testing. Experimental results in a data set from a physical test bench with a torque output range of ±260 N·m demonstrate that the proposed PINN method achieves an MAE of 0.7745 N·m and an RMSE of 1.6393 N·m. The method also converges in seconds in new environments, highlighting its robustness and practical applicability.