Research Of Multi-output Gaussian Process And Application In The Design Of Complex Antennas

The development of modern communication technology is inseparable from the development of antennas.Antennas have been widely used in the wireless communication industry.When designing and optimizing antennas,traditional full-wave electromagnetic(EM)simulation software is costly and computationally expensive.In order to improve the efficiency of antenna design,the current novel method is based on the optimization of the Surrogate-based optimization(SBO)model,Such as artificial neural network(ANN).However,it needs a lot of electromagnetic simulation data.Therefore,it is time-consuming,difficult to obtain.Also,the structure is difficult to determine,and the generalization ability is not enough.Based on the traditional Gaussian process model,this thesis mainly optimizes the design of complex antennas,and the main research of this paper includes:1.This thesis introduces the principle of GP,and uses particle swarm optimization(PSO)to instead of the conjugate gradient algorithm and improves the problem that the conjugate gradient algorithm easily falls into a local optimal in process of solving hyperparametric.Taking the prediction of microstrip antenna resonance frequency as an example to model,the experimental results show that this method greatly improves the prediction accuracy of the model.2.The thesis studies progressive Gaussian process(PGP)method.The characteristic is that a large variance in the prediction result of GP is selected,and then put into the EM software during the model process as a new Training samples.This method not only increases trusted training data but also improves the generalization ability of the model.The proposed method is used to optimize two monopole antennas.The results show that the antenna design index is met on the premise that the optimization time is greatly reduced,and the effectiveness of the method is proved.3.The principle of the multiple-output Gaussian process(MOGP)is studied.For the first time,the multiple-output Gaussian process model is applied to the antennas.The MOGP model uses the convolution of the smooth kernel function and the Gaussian process to model each output,and captures the correlation between multiple outputs by sharing the common base Gaussian processes.Using High Frequency Structure Simulator to parameterize circularly polarized antennas,the standing wave ratio,reflection coefficient and radiation gain of the antenna are studied.Compared with single output Gaussian process(SOGP),MOGP can effectively improve the performance of antenna performance prediction and evaluation.