Electromagnetic-thermal dual-field surrogate modeling of permanent magnet synchronous motor based on knowledge transfer learning

Abstract

Accurate and efficient electromagnetic and temperature performance prediction is an important basis for the design optimization of permanent magnet synchronous motors. To address the high computational cost of generating high-fidelity finite-element samples and the susceptibility of conventional DNNs to overfitting in small-sample regimes, this paper proposes a surrogate modeling method based on mechanism knowledge transfer. Firstly, a large number of low-fidelity source domain data are generated by using electromagnetic analysis and equivalent thermal circuit model respectively, and the target domain is constructed by using finite element data. Then, the DNN model is pre-trained on the source domain dataset, and the underlying parameters are frozen in the migration stage. Only a small number of target domain samples are used to fine-tune the high-level parameters to obtain the TL-DNN model. The experimental results show that the R2, MAE and RMSE of the electromagnetic field TL-DNN are 0.9951, 0.0799 and 0.1058, respectively, when only 5% of the target domain training samples are used. The corresponding R2, MAE, and RMSE values of the thermal TL-DNN are 0.9984, 0.3123, and 0.4131, respectively, Compared with the baseline DNN, the proposed model reduces the MAE and RMSE by 59.1% and 59.5%, respectively. This method effectively reduces the dependence of the surrogate model on a large number of high-fidelity finite element samples.