Data driven parameter identification of magnetic properties in steel sheets

Abstract As simulations play a crucial role for the development of modern electrical machines, it is very important to have good material models used in these simulations. Material models are dependent on certain material parameters which often cannot be measured directly and usually require a lot of computational resources to be determined. This paper investigates the application of neural networks and Gaussian processes for the identification of the magnetic permeability in electrical steel sheets. Through the manufacturing process of such steel sheets, different cutting techniques produce different material behaviour in the vicinity of the cutting edge. Therefore, the method requires the generation of datasets dependent on the degradation profile of the cut steel sheets. This is achieved through simulation and the constructed models can be reused without further simulation runs. This paper also uses an ensemble method to mitigate the issue of measurement noise. For the whole training and testing only simulation data is used as actual measurement data is not yet available.