Incremental deep-learning approach for the optimal design of a class of IPM motor for electric vehicles

Abstract

In the paper, the optimal shape design of a class of Interior Permanent Magnet (IPM) motor for electric vehicles is performed considering both running and cogging torque as design criteria. The genetic algorithm NSGA-II is applied, based on surrogate models for the objective function evaluations. Specifically, feed-forward, Deep Neural Networks (DNNs) are utilized. The DNN training is performed with a database of solutions obtained with a Finite Element (FE) model of the IPM. In order to improve the training accuracy, transfer learning techniques are applied to DNNs, used as surrogate model for the evaluation of an IPM motor performance. They showed good results in terms of reduction of computational costs and accuracy of the trained DNN in predicting the cogging torque of the motor. Moreover, incremental deep learning surrogate models are used: this approach makes it possible a dynamic procedure of training. In fact, DNNs are trained within the optimization loop of the bi-objective problem, solved by means of NSGA-II. This approach allows to obtain a reduced computational burden, without a significant loss of accuracy of the optimal Pareto front identification.