Semi-supervised Learning Gaussian Process And Its Application In Antenna Optimization

The optimization design of traditional antennas mostly relies on electromagnetic simulation software such as High Frequency Structure Simulator(HFSS),Computer Simulation Technology(CST)et al.If the antenna device has a complex structure and large size,it needs to call HFSS for many times,and the calculation cost is high and the time is long.Therefore,using modeling method instead of calling HFSS to conduct modeling analysis can effectively save time,which is a hot topic in electromagnetic optimization design.Gaussian process(GP)is a machine learning(ML)method developed based on Bayesian theory and statistical theory.The mapping relationship between the training input data and the training output data has the advantages of easy realization,adaptive acquisition of super-parameters and probability significance of the output,which is suitable for dealing with complex regression problems such as high dimension,small sample and nonlinear.Semi-supervised Learning(SSL)is the focus of pattern recognition(PR)and ML research,which mainly considers how to use a small number of labeled samples and a large number of unlabeled samples for training.The SSL method is innovatively combined with GP modeling to further mine the effective information for improving modeling accuracy in unlabeled samples and reduce the number of labeled samples required for training.It can save the time of data preparation for electromagnetic optimization,and provide a new idea for solving the fast design of complex electromagnetic structure and promoting the fast optimization of antenna structure.Based on the existing research of GP method,this thesis breaks through the limitation of using only labeled samples for training,and proposes a GP modeling method based on SSL,which saves the usage of labeled samples and achieves the purpose of rapid optimization design on the basis of ensuring good accuracy of output model.The main work is as follows:(1)The thesis briefly introduces the basic principles of SSL method and GP modeling method,and briefly explains the evaluation index of GP modeling,the way of sample selection,and the method of using MATLAB to call HFSS software.(2)The thesis introduces a method of resonant frequency prediction based on semi-supervised GP model.The self-training algorithm of GP models with the same kernel function and the co-training algorithm of GP models with different kernel functions are included.Three basic shapes of Microstrip Antenna(MSA),namely rectangular,circular and triangular,have been used to verify the effectiveness of the self-training algorithm.Annular Ring Microstrip Antenna(ARMSA)and Planar Inverted F-shaped Antenna(PIFA)have been used to verify that of the co-training algorithm with different kernel functions as the resonant frequency prediction method for antennas.(3)The thesis introduces an antenna optimization design method based on co-training algorithm of GP and Support Vector Machine(SVM).Based on the effectiveness of GP modeling and SVM modeling for small sample regression problems,combined with unlabeled samples,the differences between these two models have been used to update each other.On the basis of using fewer labeled samples,output the SSL model with high accuracy.The effectiveness of the proposed algorithm is verified by using single-frequency,dual-frequency and triple-frequency antenna,namely Yagi MSA,GPS-Beidou dual-mode MSA and Triple-frequency vertex-fed antenna,and the significance of the proposed co-training algorithm for rapid optimization design is verified.(4)The thesis introduces an antenna optimization design method based on tri-training of GP,SVM and Kernel Extreme Machine Learning(KELM).Extending from two models to three models,which improves the labeling confidence of pseudo-label samples.Combined with unlabeled samples,the differences between the three models have been used to learn from each other.On the basis of using fewer labeled samples,output the SSL model with higher accuracy.The feasibility of this method as an electromagnetic software replacement method is further verified by experiments with single-frequency,dual-frequency and wideband-frequency antenna,namely snaked PCB antenna,E-shaped dual-frequency MSA and single band-notch Ultra Wide-Band(UWB)antenna.In this study,SSL method is combined with GP modeling,SVM modeling,and KELM modeling.A new learning model is proposed,which can obtain satisfactory modeling accuracy in a relatively short time,further promoting the rapid optimization design of complex electromagnetic structures including antennas.