Shape Generation of IPM Motor Rotor Using Conditional WGAN-gp

Abstract

In this study, we utilize a deep generative model called Conditional Wasserstein Generative Adversarial Networks with gradient penalty to generate the rotor geometry of an interior permanent magnet motor, based on given performance requirements. We introduce a distortion degree as a normalization term in the loss function to achieve smooth shapes for intuitive understanding and easy manufacturing. Our results show that the generated rotor shape combines multiple features from the training data. We also demonstrate that the latent variables affect various aspects of the rotor shape simultaneously. Furthermore, we determine the validity and appropriate coefficients of distortion degree for ensuring smoothness in the generated shapes. As future work, we identify the need to prevent intersections between the magnet and steel plate regions, incorporate an error term for the input-analytical value difference, and address mode collapse in the generated shape.