Research On Modeling And Design Of Electromagnetic Interference Filter Based On Neural Network

In power electronic systems,high-speed switching of power semiconductor devices can cause serious electromagnetic interference(EMI)problems,which can affect the reliability of power electronics systems.In order to solve the EMI problems caused by high-speed switching of power semiconductor devices,EMI filters are widely used.However,the determination of EMI filter device parameters requires a large amount of experiments,therefore,the design process of the EMI filter is time consuming and costly.Replacing the experiments with modeling can greatly improve the design efficiency of the EMI filter.In addition,modeling and can also reduce the design cost of the EMI filter.The EMI filter modeling methods commonly used in the past mainly include equivalent circuit models and electromagnetic models.The parameters of the equivalent circuit model need to be continuously corrected by experiments,and it is difficult to find a completely suitable set of impedance parameters in a wide frequency range.The electromagnetic model depends on the acquisition of the physical parameters of the filter,and the calculation conditions are not the same as the actual situation.So the modeling results are often not applicable to the actual project.In order to develop a simple,high-precision,high-efficiency EMI filter modeling method,neural network(NN)is a feasible technique.This paper proposes an EMI filter modeling method based on recurrent neural network,which can accurately and efficiently simulate the insertion loss under conditions of relatively small amount of experimental data.The relative positions between the components are represented through the recurrent calculation method of recurrent neural network.Therefore,the topology-related information of EMI filter can be involved in the model.In addition,this paper proposes to use an encoding layer to encode filter candidate components into vectors.Thus,the encoded vectors instead of the parasitic impedance parameters of components in the EMI filter are taken as the input variables of the neural network model.This allows the model to learn more effective information under conditions of relatively small amount of experimental data,which reduces the training time and improving the accuracy of the model.Finally,using the simulation results of the neural network as reference,the genetic algorithm is used to quickly select the components of filter.The accuracy of the method is verified through comparing simulated and measured results of insertion loss of EMI filter.This method can make the design process of the EMI filter more efficient,and reduce the design cost of the EMI filter.