Deep Learning Based Design Methodology for Electric Machines: Data Acquisition, Training and Optimization

Abstract

This paper presents an automated deep learning-based design methodology to facilitate the design and optimization processes in electro-mechanical energy conversion devices. To validate the generality of the model, a complex machine structure with hybrid Permanent Magnets (PMs) is selected as the case study. First, the machine’s geometrical topology and the respective design variables are described in the cylindrical coordinate system and programmed into a Finite Element (FE) software package. Next, the program sweeps through the predefined ranges of selected design variables and captures corresponding air-gap flux distribution through an automated FE-based parametric analysis. The air-gap flux density data is post-processed and fed into a deep neural network (DNN) training algorithm. In particular, 10,000 data sets are utilized for training the DNN model. The trained model successfully predicts the machine’s performance for any random set of parameters, as confirmed via FE. Finally, by leveraging the trained model, the structural parameters of the machine are optimized to limit higher-order spatial flux harmonics and the cogging torque.