Optimization of IPMSM Barrier Shape Based on Neural Network

Abstract

This paper presents an Interior Permanent Magnet Machine(IPMSM) barrier shape optimization method using a neural network. To express the detailed shape of the barrier, the region of the barrier is divided into small parts. And the material property of these parts is set as a variable. As the number of the variable is increased, the neural network is applied to express the complex relationship of variables an performance of the machine. The neural network is trained with random datasets generated with Finite Element Analysis (FEA). These random datasets are generated by assigning the materials of each part randomly between 'Core' and 'Air'. In this paper, the average torque of IPMSM is set as optimizing value. A neural network is trained with input data of materials and output of average torque of IPMSM model. Structure of the neural network is decided by considering the fitness of test dataset. Using this trained neural network, the average torque of every possible barrier shape can be guessed. By examining the shape with high torque output, shape characteristics are identified,x and optimized model is suggested. © 2019 IEEE.