An improved inverse Preisach model for hysteresis modeling of non-oriented silicon steel based on neural network

Accurate characterization of the hysteresis loops of electrical steel under non-sinusoidal excitation is essential for optimizing the design of electrical apparatus via finite element analysis. In this study, an extended inverse Preisach model is proposed, in which the magnetization process is decomposed into reversible and hysteretic components, and a fully connected neural network (FCNN) is employed to represent the hysteresis behavior in place of the traditional discretized Everett function. To improve training efficiency, a cosine-annealing learning-rate scheduling strategy is adopted, combining an initial warm-up stage with a periodic decay mechanism. This approach enhances the global convergence properties of the model while mitigating the risk of convergence to suboptimal local minima. Compared with conventional discretization-based approaches, the proposed inverse hysteresis model, implemented with an enhanced FCNN, exhibits increased robustness to experimental noise and maintains high fitting accuracy over a wide range of magnetic flux density levels. Furthermore, by applying a backward-difference discretization scheme, the model is formulated with the magnetic flux density B as input and the magnetic field strength H as output. This formulation yields a stable and reliable computational framework for the unified representation of both major and minor hysteresis loops, thereby effectively overcoming key limitations of the classical Preisach model in numerical simulations.