Application Of Gaussian Process Ensemble In Microstrip Antenna

Antenna design and optimization problems generally rely on electromagnetic simulation software combining with global optimization algorithm,but it takes a long time and requires high computer hardware.In order to solve this problem,domestic and foreign literatures have proposed the method of using surrogate model.For example,machine learning algorithms such as Support Vector Machines(SVM),Artificial Neural Network(ANN)and Gaussian Process(GP)are used to build surrogate models to analyze antenna problems.Because more and more complex structure of microwave device,sometimes it may lead to reduce model precision and can't meet the actual demand.Simultaneously,we can't get enough accurate samples to improve model accuracy because of the great time consuming.On the basis of the existing GP,this thesis puts forward a modeling named GP ensemble,which obtains more precise surrogate model.The main contents are as follows:(1)This thesis introduces the basic principles of Particle Swarm Optimization(PSO)and GP and proposes a selective GP Ensemble method based on PSO(PSO-based GP Ensemble,PSO-GPE).This method includes two models.One is the GP Ensemble based on Decimal PSO(DPSO-GPE),which uses decimal PSO to optimize the ensemble weight of GP according to the training error.The other is the Gaussian Ensemble Based on Binary PSO(BPSO-GPE).In this case,the dimension of particles is only two dimensions,and it can only be 1 or 0,where 1 means that the model participates in the ensemble,0 means that the model does not participate in the ensemble,and the final models participating in the ensemble are integrated by the average method.Finally,the S11 curve fitting of the C-type compact microstrip antenna and the compact F-type slot three-band antenna used in WLAN/Wi MAX applications is performed to verify the effectiveness of the proposed algorithm.(2)The basic principle of AdaBoost adaptive lifting algorithm is introduced,and an AdaBoost GP Ensemble algorithm(AdaBoost GP Ensemble,AGPE)is proposed.In this method,several GP with different kernel functions are used as the weak learning machine,and the AdaBoost algorithm is introduced as the ensemble framework to integrate the weak learning machine,and the final strong learning machine is obtained.The S11 curve fitting of CPW butterfly microstrip antenna and UWB stepped microstrip monopole antenna proves the effectiveness of the proposed method.(3)The basic principle of Stacking model ensemble algorithm is studied,and the GP ensemble model based on Stacking algorithm(SGPE)is established,in where SVM,BP ANN and Random forest(RF)are as secondary models.The data set is divided by 5 fold cross validation to train base models,and the output results are as a new feature set,which acts as the second layer of meta-model GP input.Through the two layers of training,the final model is fused based on Stacking algorithm.In this thesis,the model is applied to the resonant frequency modeling of wide-band circularly polarized question mark liked antennas and DCS/PCS/WLAN UWB half cut plane antennas,and the good prediction results show that the model can improve the model accuracy to a certain extent.