Optimal Design of IPM Rotor Shape Using Generative Adversarial Networks

In this study, we propose optimal rotor design of IPM motors using CNN (Convolutional Neural Networks) and DCGAN (Deep Convolutional Generative Adversarial networks), which are widely used as image classification and virtual image generation deep learning techniques. 56,400 images obtained from topology optimization analysis using the genetic algorithm within commercial finite element analysis software JMAG are classified by Back-EMF value and limited under THD (Total Harmonic Distortion) constraint. After confirming the applicability of deep learning through CNN classification using the acquired images, new fake images were generated using DCGAN. After generating new images, validation was carried out using JMAG. Through this, it was confirmed that rotor shape design using deep learning is possible. General optimization techniques were only able to obtain limited local optima, but by increasing solutions similar to the local optima using deep learning techniques, it is confirmed that the selection range of the optimal solution can be expanded.