Physics-Informed Machine Learning Based Permanent Magnet Synchronous Motor Temperature Estimation

Permanent magnet synchronous motor (PMSM) temperature estimation is essential to ensure motor works safely, and the temperature at critical locations needs to be estimated accurately and quickly. To this end, this paper combines data-driven method with physical knowledge to realize an end-to-end trainable temperature calculation framework. Firstly, a lumped parameter thermal network (LPTN) of a PMSM is established by an electromagnetic simulation software, and the simulated thermal parameters like capacitance, resistance, and loss are used as the initial values. Three independent neural networks are incorporated into the framework, and then the simulated parameters are corrected in real time according to working condition and internal temperature distribution. Finally, the temperature is updated through the ordinary differential equations of LPTN. Many physical constraints are considered, and a physics-informed loss term is added into loss function. The effectiveness of the proposed algorithm is verified with measured temperature data, and the final mean absolute error is 1.71°C, showing that this physics-informed framework can achieve high accuracy, excellent generalization, and robustness even with small dataset that lacks diversity.