Characteristics Estimation and Design of SPMSM Using Analytic Method-Based Transfer Learning

In the initial design stage of a surface-mounted permanent magnet synchronous motor (SPMSM), it is necessary to analyze characteristic parameters according to key geometric design variables using either analytical method (AM) or finite element analysis (FEA). The AM has the advantage of fast computation and the ability to generate large-scale data, but it has limitations in accurately reflecting nonlinear characteristics that vary with geometric design parameters. Therefore, to achieve more accurate predictions, FEA must be performed. However, it involves high computational cost and long analysis time. To address this trade-off, this study proposes a predictive framework based on transfer learning, in which a deep neural network (DNN) is pretrained using AM-generated data and fine-tuned with a limited amount of FEA data. The proposed method takes advantage of AM to rapidly generate a large dataset for initial training of the deep learning model, followed by transfer learning using a small number of FEA-labeled samples to improve prediction accuracy while minimizing computational burden. The proposed method can quickly and accurately estimate the electromagnetic characteristics according to the geometric and physical design variables of the SPMSM, and it has been confirmed that it effectively reduces motor property computation time by complementarily leveraging the strengths and limitations of both AM and FEA.