Thermal Neural Networks for Electric Vehicle Traction Machines: A Transfer Learning Approach

In this paper, the application of transfer learning techniques in the domain of thermal neural networks within the context of electric vehicles is addressed. For the increasing integration of electrified transportation, efficient thermal management is crucial for optimizing the performance and longevity of electric vehicles. Traditional methods often struggle to adapt to diverse operating conditions and dynamic environments. To address this challenge, this work investigates a transfer learning approach that leverages knowledge gained from one simulated thermal scenario to enhance the performance in another real world scenario. Thereby, sparse real world data can be boosted with additional information, which enhances the generalization capability when predicting unseen data samples. A suitable thermal neural network model is trained using a simulated dataset, that is generated by a lumped-parameter thermal network, which is solely based on finite element method simulation data. Subsequently, the derived weights are utilized to investigate the feasibility of transfer learning by retraining the model on a dataset obtained from test bench measurements. The transfer learning methodologies are then compared to a randomly initialized model, and the results are subjected to detailed analysis. The findings demonstrate that, under appropriate conditions, pre-training can yield improved model performance and demonstrates the potential for virtual development in the present use-case.