Skin Antenna Design Based On Machine Learning

The skin antenna has a decisive effect on the communication and stealth performance of a plane.It is a key problem to design an antenna with low radar cross section(RCS)while it maintains the radiation performance.It takes a lot of computation resources to analyze the skin antenna with traditional methods because there are plenty of variables in the skin antenna.The development of the skin antenna is restricted by the modeling and optimization methods.Recently,with the development of computer techniques and artificial intelligence(AI),machine learning(ML)has been introduced into the antenna modeling,and the design convenience has been improved effectively.The modeling and design of skin antennas,however,are still of low efficiency,inaccuracy and low degree of design freedom.The modeling and design of antennas with low-RCS and good radiation performance based on machine learning are important.This dissertation focuses on the modeling and design of the skin antennas based on machine learning.With the machine learning based modeling method to improve the antenna modeling performance,the antennas with low RCS and good radiation performance are designed.The main work and creativity of this dissertation include:1.A feature selection method based on relationship and sensitivity is proposed to reduce the complexity of antenna modeling.In the modeling of antennas,the input of the artificial neural network(ANN)is the geometrical parameters and the output is the electromagnetic(EM)response.ANN maps the nonlinear relationship between the input and output.It is an important precondition to measure the relationship between the input and output and select the impressive geometrical parameters as the input based on the measurement.The relation and sensitivity between geometrical parameters and EM response are studied.A feature selection method is proposed based on the maximal information coefficient(MIC)and sensitivity.MIC,as an exploratory data mining tool,is used to identify the linear and non-linear correlations between parameters and responses.The response range of the input parameter is used to measure the sensitivity.A wide response range implies a high-sensitive input while a narrow one implies a lowsensitive one.By selecting the high-sensitive and correlative parameters as the input of ANN,the sampling space is reduced effectively and the relationship between the input and output is simplified.Thus,the modeling efficiency of ANN is improved effectively.2.A new transfer function(TF)based on Fourier series is proposed to improve the accuracy of ANN modeling.In antenna modeling,the output of ANN is EM response.When the wideband EM response is considered,the dimensions of the output and input of ANN are unmatched and the accuracy is deteriorated.Usually,TF based on the poleresidue function is used to fit the wideband response.The coefficients of TF are applied to the output of ANN and its accuracy with the wideband output is improved.However,the orders of TF are different when the response changes,and the accuracy decreases when the coefficient dimension is fluctuant.We analyze the problem of order change and propose the transfer function based on Fourier series in this dissertation.By tuning the scaling factor in FS,the order of TF is determined.Then the accuracy and efficiency of ANN are improved.We also analyze the sensitivity and continuity of the coefficient,and the results show that the coefficient of the proposed TF is insensitive and continuous.3.A design method based on ANN for low-monostatic-RCS arrays is proposed.When a plane wave illuminates an array with a specific incident angle,the backward scattering of the array deteriorates the stealth performance.The mechanism of the backward scattering is analyzed and a method to suppress the scattering wave is proposed.An array with good scanning performance and low monostatic-RCS is designed based on ANN.The wideband low-RCS array based on the diffusion mechanism is also analyzed.By covering different metasurfaces(MTS)with different reflection phases in the array,the scattering waves are diffused and the RCS is reduced.A wideband array with lowRCS is designed based on ANN.4.A design method for low-bistatic-RCS array based on the generative adversarial network(GAN)and equivalent circuit is proposed.The theory of the low-RCS array based on lossy MTS is analyzed and the equivalent circuit model is studied.By using the pre-knowledge,the modeling and design of the array are broken up into the modeling and design of the lossy MTS and antenna.The equivalent circuit is also applied to collect the samples of GAN.The topology is determined when the equivalent circuit of the MTS is fixed and the dimension and time for data collection are suppressed effectively.The ability of image synthesis and editing of GAN is used to generate a satisfying MTS.Two scanning arrays with low RCS are designed based on the proposed method.