Nondestructive Magnetisation Estimation for Permanent Magnets Using Autoencoder-Based Dimensionality Reduction

A magnetisation estimation method that uses an autoencoder (AE) model with a convolutional neural network (CNN) is proposed. The proposed method achieved high accuracy even for complex magnetisation distributions, including defective regions. The proposed model improved estimation accuracy by over 50.62% compared to the conventional multi-layer perceptron method. Additionally, the proposed model is effective for multi-output regression problems involving multidimensional vectors. This method enables non-destructive and rapid estimation of internal magnetisation in permanent magnets from external magnetic flux density.