Optimal Design of Axial Flux Motors with Housing Loss Consideration Based on Transfer Learning Using Quasi-3D and 3D FEA

Abstract

Axial flux motors (AFMs) exhibit higher torque density and compact structure than radial flux motors (RFMs), especially in pancake structures with a ratio of outer diameter to axial length. However, unlike RFM, AFM cannot fully express the three-dimensional (3D) structure in a two-dimensional (2D) cross-section, which causes high errors in the 2D and 3D finite element analysis (FEA) results. To reduce the error while not substantially increasing the computational time, this study proposes to generate a surrogate model through a Deep Neural Network (DNN) with transfer learning. By using quasi-3D FEA as a large amount of pre-trained data and 3D FEA as a small amount of training data, we were able to create a surrogate model with a short computational time and accurately predict the performance of AFM. In addition, the housing eddy current loss due to the axial leakage flux is important in the AFM with a fixed housing. In quasi-3D FEA, the end cover loss tends to be overestimated, and the housing side cannot be modeled, which limits the accurate prediction of housing loss. To address this limitation, housing loss was more accurately estimated by applying transfer learning in this paper. Finally, an optimal design model was derived through an optimal design process that considered housing loss using the proposed method, and consistency with FEA and reduction in housing loss were confirmed through experiments.