Design Optimization of Switched Reluctance Motors Using Deep Learning

Abstract

Switched Reluctance Motors (SRMs) are known for their low manufacturing costs, simple structure, high torque-to-inertia ratio, and minimal maintenance. However, they have high torque ripples due to the salient nature of the stator and rotor poles. To address this problem, this thesis applies a Deep Learning optimization approach to minimize the torque ripple and maximize the average torque by changing the stator and rotor pole arc angles, 𝛽𝑠 and 𝛽𝑟 respectively, using Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs) implemented in Python. The training data consists of cross-section images of a 6/14 SRM and static and dynamic characteristics were captured using the Finite Element Method and MATLAB Simulink models. The CNN model takes the cross-section image and predicts the average torque and torque ripple. The results of the CNN model are compared to previous papers that applied a similar method to predict various parameters of different motors. A variation of the GAN model called FastGAN was used to generate cross section motor images. The model takes a noise vector and generates a cross-section quarter model image. The combined FastGAN-CNN model produced an optimal design that increases the average torque by 2% and decreases the torque ripple by 24% while being 17 times faster than the traditional FEM.