Physics-Informed Neural Networks for Real-Time Thermal Monitoring of Electric Vehicle Propulsion Motors

Thermal analysis is critical for preventing component failures in electric motors, such as magnet demagnetisation and insulation degradation, by en-suring that operating temperatures remain within safe limits. Industrial ap-plications commonly rely on Lumped Parameter Thermal Network (LPTN) models; however, these models often lack sufficient accuracy to capture complex thermal dynamics. This paper investigates the potential of Phys-ics-Informed Neural Networks (PINNs) to improve dynamic thermal temper-ature prediction in Battery Electric Vehicle propulsion motors, with a focus on Permanent Magnet Synchronous Motors (PMSMs). The performance of a PINN is compared with a Feedforward Neural Network (FNN) and a tradition-al 4-node LPTN model. The models were trained and evaluated using data from a PMSM prototype under dynamic random-walk driving cycles to em-ulate realistic operating conditions. The dataset was enhanced with physi-cally meaningful features, including copper and residual losses, while Ex-ponentially Weighted Moving Averages were incorporated to capture tem-poral dependencies. The PINN integrates physical knowledge by embedding LPTN-derived Ordinary Differential Equations (ODEs) within its loss func-tion. Results show that both neural network models significantly outper-form the LPTN baseline, which achieved an average Root Mean Square Error (RMSE) of 2.074. The FNN achieved the highest accuracy, with an RMSE of 0.709, followed by the PINN, with an RMSE of 0.824. The study demon-strates the feasibility of PINN for real-time thermal monitoring of PMSMs. It highlights the advantages and limitations of incorporating ODE-based physical constraints in neural network models.