Real-time Rotor Temperature Estimation of Induction Machines Using a Differentiated Input Thermal Neural Network

Monitoring the rotor temperature of the driving machine is crucial for the safety and performance of electric vehicles. The thermal neural network (TNN) integrating deep learning with the thermal resistance model is a state-of-the-art method for rotor temperature estimation. However, TNN does not actively select the influencing factors in the thermal parameter identification process, which limits both the accuracy and generalization performance of the model. Meanwhile, the multi-dimensional network input makes it difficult to be directly deployed in the embedded control system with limited computing power. Therefore, this article proposes a differentiated input thermal neural network (DI-TNN) for real-time rotor temperature estimation of induction machines. Firstly, a simplified yet concise lumped-parameter thermal network is established according to the structure of the oil-cooled induction machine, and the actual influencing factors of each thermal network parameter are obtained based on physical mechanisms. Subsequently, a DI-TNN model consisting of multiple differentiated input networks is proposed, and a feasible loss function is constructed for model training. Furthermore, the thermal parameters corresponding to various operating conditions are traversed using the trained model, and a look-up table based real-time temperature estimation method is deployed into an embedded system. Finally, a rotor temperature dataset covering a wide range of operating conditions is collected through bench tests, and experiments are conducted to compare and validate the DI-TNN against the TNN. The results demonstrate that the DI-TNN offers substantial improvements in the interpretability of thermal parameter identification, the accuracy of rotor temperature estimation, and its applicability in embedded deployment.