Transfer Learning-Based Modular Neural Network for Multi-Objective Optimization of Interior Permanent Magnet Synchronous Motors

Finite element analysis is frequently used to optimize the characteristics of interior permanent magnet synchronous motors throughout the design phase. The existing toolchains enable the full automation of simulating and optimizing a reference motor by manipulating the input design parameters within the feasible design space. However, for each motor design, a complete simulation is required, implying a high computational burden and time cost. Moreover, once the input design parameters undergo variations, it becomes necessary to initiate the simulation process from the beginning. The previously obtained simulation results are not helpful for the new task. In this paper, a new method using modular neural networks based on transfer learning (TL) under dimensionally varying input space conditions is presented. By transferring certain parts of the pretrained neural networks (NNs) of the old task to the new task's NN, the previously learned parameters can be applied as the initial weight for the new network. Finally, the optimization process is completed by combining this approach with multi-objective optimization. The results show that the learning of the new NN is promoted with the help of TL. In addition, highly flexible surrogate models are achieved, enabling accurate prediction capabilities and a fast optimization time of around 12 seconds.