Paper FS08 1 Abstract—This paper presents a novel technique for the identification of the Preisach density function which is based on a Neural Network approach and which requires a relatively limited amount of experimental parameters. The fundamental idea of this method is to identify Preisach function of the material by training a neural network with a set of loops whose identification function is known. In the final section of the paper the method is verified on several cases. Index Terms—Hysteresis, measurements, modeling, soft magnetic materials. I. INTRODUCTION REISACH model is a powerful tool for the description of hysteresis[1]-[3]. It describes the hysteresis loop by a superposition of the magnetization values of Preisach particles. Hysterons, which are a mathematical function that can have two values, describes the magnetization state of each of these particles. The switching from one value to the other is triggered when the external field reaches specific values. The number and the characteristics of the Preisach particles necessary for describing the hysteresis properties of a given magnetic material are described through a distribution function generally called Preisach function[4]. Several procedures have been developed in order to identify Preisach function and some of them implies the solution of set of equations and the experimental knowledge of several parameters. [5]-[6] This paper presents a methodology for identifying a Preisach function by using neural networks (see appendix for further details on Preisach theory). The neural network used is a multiplayer perceptron trained with the Levenberg-Marquadt training algorithm. The network is trained by some hysteresis data, which are generated from a set of Preisach example functions. More precisely, the set of Manuscript received February 14, 2002. This work was supported by MIUR and CNR. Maurizio Cirrincione is with C.N.R. Ce.Ri.S.E.P c/o Palermo University, Palermo, I-90128 Palermo (e-mail: nimzo@cerisep.pa.cnr.it). Rosario Miceli is with the Department of Electrical Engineering, Palermo University, Palermo, I-90128 Palermo (e-mail: miceli@diepa.unipa.it). Giuseppe Ricco Galluzzo is with the Department of Electrical Engineering, Palermo University, Palermo, I-90128 Palermo (e-mail: ricco@diepa.unipa.it). Marco Trapanese is with the Department of Electrical Engineering, Palermo University, Palermo, I-90128 Palermo (e-mail: mtrap@diepa.unipa.it). training data consists of the applied field points, the corresponding level of magnetization and the used parameters in the Preisach function. Each function is one Gaussian function. Changing the average and/or the standard deviation of each function allows to have a different shape of the hysteresis loop. Several hysteresis loop data are submitted to the trained network. The network yields the average and the standard deviations that must be used in order to reproduce, by using Preisach model, the hysteresis loop that has been submitted. The results are verified by comparing the input hysteresis loop with the one obtained with the proposed approach. The fundamental idea of this method is to identify the Preisach function of a material through the knowledge of a collection of hysteresis curves. As a matter of fact, the neural networks provides the parameters necessary for describing a given hysteresis loop, under the assumptions that the type of the Preisach function is known. In this paper the type of the Preisach function, is assumed to be a Gaussian function, but this is not a limitation and the method can be extended to other functions, as already mentioned. Moreover, the method is not linked to a particular set of measurements and can be used in a wide range of cases. II. THE IDENTIFICATION PROCEDURE The identification procedure can be divided into four steps: A) network training; B) submission of test set ; C) output of the results; D) experimental verification. A. Network Training The objective of the training phase is to build a network, that is able to output some parameters that can be used to set a Preisach function. This allows to construct, by using Preisach model, the hysteresis loop which had been submitted to the network. The neural network that has been chosen is a three layer neural network: the first layer consists of 5 linear neurons, the second layer of 25 neurons with hyperbolic tangent sigmoid function, and the output layer of 2 linear neurons. The input is made up of 100 units. The first layer makes a linear reduction from the high dimension input space to the lower 5-dimension space: the output of the first layer is therefore the feature vector that sums up the most relevant characteristics of the input space and is strictly connected to five relevant properties Preisach Function Identification by Neural Networks Maurizio CIRRINCIONE, Rosario MICELI, Giuseppe RICCO GALLUZZO, and Marco TRAPANESE, Member, IEEE P