Optimization Design Of Microwave Devices Based On Prior Knowledge Gaussian Process Modeling

The design of the structure and parameters of microwave devices depends on the electromagnetic software simulation.Although its performance parameters can be obtained by full wave electromagnetic simulation,this method is not only complex but also expensive to calculate.Therefore,researchers put forward a Machine Learning(ML)method to solve the problem of microwave device analysis.The common ML methods include Support Vector Machine(SVM),Gaussian Process(GP)and Artificial Neural Network(ANN),etc.ANN has the functions of self-learning,associative storage and searching for the optimal solution.However,it has no rigorous reasoning basis and reasoning process.Its accuracy depends on a large number of training data.The SVM is calculated using the hinge loss function and regularization of empirical risk and structural risk classification ML method,well solved the problems such as nonlinear,high dimension and local minimum points.However,because it does not involve probability measure,the experimental results have no probability significance.In addition,it is easy to overfit and difficult to solve multi-classification problems.GP is an emerging ML method,which is based on bayesian theory.Compared with ANN,its output is probabilistic and does not require a large number of training samples.Compared with SVM,its kernel function is easier to be determined.In addition,the output of GP is probabilistic and different kernel functions can be selected.Therefore,when designing microwave devices,GP model can be used as a substitute for the accurate model established by electromagnetic software,which can reduce the modeling time without reducing the prediction accuracy of the model.In this thesis,a new GP modeling method is proposed.Priori knowledge is introduced into the model.The improved model has an ideal optimization effect on microwave devices.(1)This thesis introduces the basic principle of GP,the model structure of GP and the evaluation method of the model,as well as the principle and basic flow of particle swarm optimization(PSO),and relevant theories of prior knowledge.(2)Empirical formula of microwave device is introduced the process of GP modeling.The resonance frequency and input impedance of the antenna is calculated the formula,and they are as prior knowledge and roughness value,precise values are obtained from the simulation software.Using GP model alternative HFSS,we can effectively reduce the modeling time.This modeling method is applied to cylindrical conformal microstrip antenna(MSA).In addition,the method is also applied to the prediction of resonant frequency of planar waveguide dual-frequency MSA and SIW back-cavity gap antenna.(3)The GP modeling method based on rough grid is studied.The rough value is calculated by the model with fewer HFSS analysis times,and the accurate value is from by the model with more HFSS analysis times.This method is simple and convenient to operate,which can not only guarantee the accuracy of the model but also reduce the running times of the simulation software and effectively reduce the calculation cost of the experiment.This method is applied to the modeling of the dual-frequency MSA,the new coplanar waveguide feeding S-type antenna and the new aperture coupled MSA.(4)The GP modeling method based on ADS is studied.The experimental results of ADS are taken as the prior knowledge and the GP model is combined for modeling.The model is applied to the coupled line bandpass filter,and the particle swarm optimization algorithm is used to optimize the filter.