Transfer Learning for Efficiency Map Prediction

Abstract

This paper explores methods to extend a trained deep neural network for predicting efficiency maps to work on different motor drive topologies. This procedure reduces the computation cost associated with training deep networks by transferring knowledge over similar tasks handled by the deep networks. Two types of synchronous AC machines, including a flat-type interior and a surface-mounted permanent magnet motor are tested over their entire torque-speed profiles to validate the applicability of the proposed methodology. The obtained results demonstrate an improvement in both computation time required for training and a reduction in the size of the required dataset for transfer learning. Also, the performance is compared with conventional supervised learning on the same data and the same neural network architecture.