Research On Modeling Method Of Complex Magnetic Properties Of Electrical Steel Sheet Based On Convolutional Neural Networks

With the continuous upgrading of electrical equipment,it is an important indicator to search for high power density and high efficiency in electrical engineering.The electrical equipment such as motors and transformers widely take electrical steel sheets as the magnetic material in the core.There exists multi-physics coupling of magnetic,temperature,and structural stress field during design,manufacturing and operation of electrical equipment.,where the magnetic properties of electrical equipment will be affected by temperature,stress and magnetization method.Therefore,the accurate measurement and modeling of the complex magnetic properties of electrical steel sheets under multiple physical factors is a prerequisite for fine design and optimization of electrical equipment.In order to numerically modeling the magnetic properties of electrical steel sheets,scholars at home and abroad have established many different magnetic hysteresis model.The neural network hysteresis model can avoid the complicated physical mechanism and mathematical operations required in traditional hysteresis mathematical models.It also can be combined with electromagnetic field numerical analysis to reduce calculation time.However,the back-propagation(BP)neural network magnetic hysteresis model owns slow learning rate and poor simulation accuracy.With the development of deep learning,the convolutional neural network has better ability to extract and learn features when dealing with high-dimensional complex problems.Therefore,a novel modeling method based on convolutional neural network was put forward for modeling of the vector magnetic hysteresis properties of electrical steel sheets.The main research contents of this thesis are summarized as follows:Firstly,the magnetic flux density B and magnetic field intensity H were measured by applying the self-developed vector magnetic characteristic measuring device for electrical steel sheets under temperature and stress conditions.The experimental data was preprocessed as a training sample for the model.Secondly,in order to improve the accuracy of modeling vector magnetic properties of electrical steel sheets,this thesis applies residual module to increase width and depth of the model,and establishes a deep convolutional neural network hysteresis model based on residual connection,However the characteristics information of the model appears deviation after training of the multi-layer network,which reduces training speed.Therefore,on the basis of the deep convolutional network,a batch normalization layer is introduced,and corresponding weights and biases in the network are batch normalized.Finally,under the coupling of temperature and stress,the vector magnetic properties of electrical steel sheets by different hysteresis models are compared.The feasibility and effectiveness of the suggested deep convolutional neural network magnetic hysteresis model are verified.Based on the analysis of final modeling results,it is found that the magnetic hysteresis model based on the deep convolution neural network can not only reduce the number of iterations,but also ensure the precision of the simulation of hysteresis properties.