Explainable Deep Neural Network for Design of Electric Motors

This study presents a novel two-step optimization method that incorporates explainable neural networks into topology optimization. The deep neural network (DNN) is trained to infer the torque performance from the input image of the motor cross section. The sensitive region that has a significant influence on the average torque is extracted using gradient-weighted class activation mapping (Grad-CAM) constructed from the DNN. Then, the optimization with respect to the torque ripple is performed only in the incentive region with little influence on the average torque. The proposed method is shown to increase the average torque of an interior permanent magnet (IPM) motor by 14% and reduce the torque ripple by 79% compared with the original model.