NEURAL NETWORK MODEL FOR SCALAR AND VECTOR HYSTERESIS

Miklós Kuczmann - Amália Iványi

   The classical Preisach model is one of the most generally applied techniques to simulate the behaviour of magnetic materials, to describe hysteresis phenomena. According to the theory of Weiss, the classical Preisach model assumes that ferromagnetic materials consist of many elementary interacting domains, and each of them can be represented by a rectangular elementary hysteresis loop. The fundamental concepts of the Preisach model is that different domains have some probability, which can be described by a distribution function, also called the Preisach kernel. On the basis of the so-called Kolmogorov-Arnold theory the feedforward type artificial neural networks are able to approximate any kind of non-linear, continuous functions represented by their discrete set of measurements. A neural network (NN) based scalar hysteresis model has been constructed on the function approximation ability of NNs and if-then type rules about hysteresis phenomena. Vectorial generalization to describe isotropic and anisotropic magnetic materials in two and three dimensions with an original identification method has been introduced in this paper. Good agreement is found between simulated and experimental data and the results are illustrated in figures.

Keywords: hysteresis characteristics, Everett surface, vector hysteresis, feedforward type neural networks, backpropagation training method

[full-paper]