Advanced Design Optimization of Switched Reluctance Motors for Torque Improvement Using Supervised Learning Algorithm

Existing research on geometry optimization of switched reluctance motor (SRM) using machine learning algorithms has focused only on the machine's static characteristics. The dynamic characteristics, however, are critical to improve the SRM performance, particularly at high speeds. This paper introduces an advanced optimization method utilizing a supervised learning algorithm to act as a surrogate model for both static and dynamic characteristics of the SRM. In this work, back-propagation neural network (BPNN) is applied to map out the SRM geometrical parameters, stator and rotor pole arc angles and their dynamic performance metrics such as average torque and torque ripples. To capture the training data, finite element analysis (FEA) and MATLAB Simulink models are implemented to study the static and dynamic characteristics of the considered 6/14 SRM. Levenberg-Marquardt is applied to train the BPNN. The results of the proposed optimal design candidates are verified using FEA and MATLAB simulations, confirming the effectiveness of the optimal design. The optimal design improves the average torque by around 2% and reduces the torque ripples by around 24%. Moreover, the proposed method significantly decreases the computational overhead.