Neural Network Meta-Modeling and Optimization of Flux Switching Machines

Abstract

This paper presents the application of neural networks (NN)in the design and optimization of the flux switching machine. A finite element (FE)model of the flux switching machine is used to create data for the training of NNs. The trained NN meta-models are used to predict the properties of machine designs. Subsequently, a preselection from these predictions for further FE calculations is employed. Based on this approach, a generational NN based optimization is implemented. Furthermore, during the optimization many NN topologies and output variable combinations are investigated. The NN predictions and their quality are evaluated according to output variables, generations, and NN layouts. In addition, the improvements in the machine's capabilities are presented using a Pareto-domination based approach.