Study on Method of Magnetization Estimation of Permanent Magnet using DNN utilizing Heisenberg Model for Generation of Training Data

Permanent magnet (PM), which properties are maintained semi-permanently, is one of the essential magnetic materials for the electromagnetic devices. The performance of those devices substantially depends on the magnetization distribution of the permanent magnets. Therefore, it is essential to clarify the magnetization distribution of PMs in advance. To estimate a physical magnetization distribution in PM, a deep neural network (DNN) based magnetization estimation method using the measurement data of magnetic flux density has been proposed. However, when the estimation of 3-D magnetization distribution is carried out using DNN, the estimation accuracy is not high because an effective method for generating training data has not been proposed. In this paper, an effective procedure for the generation of training data using Heisenberg model, which can realize many distributions of PM magnetization, is proposed. As a result, it is shown that the DNN with the proposed data generation method is able to be effective for the accurate and fast estimation of 3D PM magnetization distribution.