Dimension Reduction Gaussian Process And Its Electromagnetic Applications

In recent years,electromagnetic simulation software combined with global optimization algorithm is a mainstream method used in microwave devices design.However,due to the long simulation time of the simulation software,this method is very inefficient.Therefore,finding alternative simulation software models for optimization has become a research direction recently,such as kernel extreme learning machine(KELM),polynomial regression(PR),artificial neural networks(ANN),Gaussian process(GP)and other modeling method.The traditional training model requires a large number of training samples to obtain better training accuracy and generalization ability.However,in the process of electromagnetic device design,the cost of calculation and simulation is very high,which brings great obstacles for us to design electromagnetic devices through machine learning method.In order to solve this problem,this thesis studies the application of dimensionality reduction GP model in electromagnetic field.The method in this thesis can greatly reduce the features of training data,improve the computational efficiency and save the training time of the model.Especially in the establishment of inverse surrogate model,the dimensionality reduction GP shows its powerful performance.The main contents of this thesis are as follows:(1)The basic principles of Gaussian process(GP),differential evolution algorithm(DE),isometric mapping(Isomap),principal component analysis(PCA)and convolution pooling are introduced,and the joint call method of MATLAB and electromagnetic simulation software HFSS is expounded.(2)A new differential evolution algorithm,namely population competition based self-adaptive differential evolution(PCSADE),is studied,and better optimization ability and convergence speed than other intelligent optimization algorithms is proved based on six benchmark functions.(3)By introducing manifold learning module into GP training model,a new training model with the function of data dimension reduction is studied,named DE based main fold Gaussian process(DE-MGP).In this model,PCSADE algorithm is used to train manifold parameters.The proposed model can work under the optimal manifold parameters.In this part,DE-MGP is applied to the extraction of coupling coefficients of fourth-order and sixth-order coupled filters respectively,and the experimental results prove the superiority of the model.(4)A deep GP based on principal component analysis(Deep PCA-GP)model is studied.The feature extraction module composed of multi-layer zero phase component analysis whitening(ZCA)and principal component analysis is used to reduce the dimension of high-dimensional data,which greatly decreases the calculation of training model and the dependence of model on the number of samples.In this part,the inverse surrogate models of printed dipole antenna,WLAN dual band monopole antenna and printed multistubs loaded resonator rectangular monopole antenna are established respectively.The experimental results show that the inverse antenna surrogate model of the Deep PCA-GP has very high ability of parameter extraction.