Modeling dynamic hysteresis through Fully Connected Cascade neural networks

This work documents the research towards the development of a neural approach to represent ferromagnetic materials under dynamic excitation. The proposed approach is based on a Neural System (NS) composed by advanced neural networks featuring the novel Fully Connected Cascade (FCC) architecture. This architecture is particularly suited to solve complex problems. For the training of the network an ad-hoc second order algorithm, known as Neuron-by-Neuron, was used. The neural system was trained on experimental hysteresis curves obtained by measurements performed at different frequencies. Validation was performed both against a numerical model (the dynamic Jiles-Atherton model), and against measurements on a non-oriented Fe-Si device.