Switched Reluctance Motor Design Optimization: A Framework for Effective Machine Learning Algorithm Selection and Evaluation

This study employs various machine learning algorithms (MLAs) to map out the stator and rotor pole arc angles of 6/14 switched reluctance motor (SRM) and their static and dynamic nonlinear characteristics. The MLAs under consideration include a back-propagation neural network, radial basis function neural network, generalized regression neural network, and conventional regression fitting algorithms. This work introduces an extensive analysis of these MLAs, including their structure, fundamentals, and learning process. Additionally, a comprehensive evaluation framework is established, encompassing assessments of training results, generalization capability, and computational time. It also addresses key challenges inherent in learning MLAs, specifically overfitting and underfitting issues. These evaluation criteria guide the selection of the optimal machine learning topology tailored for geometry optimization in SRMs. The chosen MLA is then applied to predict the optimal pole arc angles that enhance the average torque and decrease torque ripples of the considered SRM.