Fast Calculation Method Of Electromagnetic Field Numerical Simulation Results Based On Deep Learnin

The electromagnetic comprehensive performance analysis and optimization of modern electrical equipment has become an urgent need to solve the problem of calculation difficulty,multiple time and space scales,strong coupling and nonlinearity,making the efficient simulation analysis and service safety of modern electrical equipment and its systems face major scientific challenges.At present,the electromagnetic field analysis of electrical equipment is mainly simulated by finite element calculation.For a motor or transformer model with a slightly complicated structure,the finite element simulation time may take tens of hours,or even hundreds of hours,which may lead to computer system breakdown.Sometimes in order to improve the accuracy of the calculation results,it is necessary to finely divide the space of the model,and the occupied computer memory will increase geometrically.If there is a slight change in the geometric structure size or material properties,the numerical solution must be re-calculated.Therefore,finding a performance analysis method that can accurately establish the mapping relationship between structure and performance without requiring huge computational costs has become an important research topic.This paper uses finite element simulation software to obtain sample data to train deep convolutional neural networks,and uses deep learning to directly learn the effective features in the data,without the advantage of human intervention,to achieve accurate prediction of the electromagnetic field of electrical equipment.The research content of this article includes the following aspects:First,The basic structure of convolutional neural network is introduced,including convolutional layer,pooling layer,activation function and fully connected layer.The multi-scale feature extraction and feature fusion methods in image feature extraction are studied,and the working principles of convolutional neural networks,fully convolutional neural networks and generative confrontation networks based on the encoding-decoding structure are further introduced.Secondly,the physical models of the transformer and the motor were established by finite element analysis software,and the corresponding magnetic field distribution cloud image was obtained by changing the geometric structure parameters,materials and excitation conditions as the sample data set of the deep neural network.Trained three different network models,namely CNN,FCN and GAN based on the encoding-decoding structure.Due to the shallow network depth,the three models were not comprehensive enough to extract the deep feature information of the image,and could not accurately extract the magnetic field cloud image data set.The mapping relationship between the two,the prediction effect is relatively poor.Finally,in view of the lack of prediction accuracy of the above three models,further improvements are made,and an improved U-net model prediction method for magnetic field cloud images is proposed.The CNN,FCN,GAN,U-net and improved U-net models based on the encoding-decoding structure are trained on the transformer and motor data sets respectively,comparing and analyzing the prediction accuracy of magnetic field cloud images of several networks,the improved U-net model can learn the mapping relationship between magnetic field cloud image data sets,generate high-resolution images,and compare them with the finite element calculation results.It is superior to finite element calculation in terms of prediction accuracy and time efficiency,and can replace the image mapping process of finite element simulation calculation to a certain extent,and provides great practical reference for the design and optimization of modern electrical equipment.