Abstract:

In this paper, we propose a surrogate model based on neural networks, for the rapid evaluation of the performance of permanent magnet synchronous motor designs, especially the detailed torque waveform. In the training phase of our proposed method, motor design parameters are taken as input and gap flux density information is taken as output, both fed to train neural networks. In the test phase, gap flux density is predicted with the trained neural networks, the torque waveform is subsequently reconstructed, and a peak-to-peak cogging torque amplitude is estimated from the waveform. We compare the proposed method with conventional neural network based surrogate models, in which torque information is directly used for training, and confirm that the proposed method shows higher accuracy than conventional approaches, especially when the training data size is small.