| In electromagnetic optimization problems,complex electromagnetic devices need to be designed and optimized,which is solved by electromagnetic simulation software and numerical optimization algorithms in traditional solutions.However,the application of machine learning(ML)combined with optimization algorithm can greatly reduce computational cost and obtain optimal solution much more quickly.And it has been widely studied both abroad and at home.Student’s T Process(STP)is a small sample model in ML with flexible posterior variance and strong robustness,what can well establish the mapping relationship between input and output samples,and handle missing data as well as abnormal data.At the same time,it has advantages in the ability of strong generalization and adaptively obtaining parameters for model optimization.Bayesian Optimization(BO)is a fast global optimization algorithm,it can solve engineering optimization design problems effectively.Probabilistic agent model and sampling function are the cores of BO.By constructing efficient probabilistic agent model and designing sampling function with excellent mining ability,convergence speed of the algorithm can be accelerated and a better solution can be found.In this thesis,STP and BO algorithm are innovatively combined to optimize the model of STP,then to improve performance of model.As surrogate model of BO algorithm,new sampling function is designed and applied to optimization design of electromagnetic problem by the thesis,which provides a new idea for solving the rapid design of complex electromagnetic structures and promoting the rapid optimization of antenna structures.Based on existing studies,the thesis proposes a new algorithm combined STP and BO,what saves iteration time and achieves the goal of rapid antenna optimization design under the premise of ensuring accuracy of output model.The main work as follows:(1)The basic principles of STP model are briefly introduced,including mathematical process of single-output and multi-output;followed by the principle,implementation and basic classification of BO;then,rationale for selecting training data and test data and basic principle in software calling HFSS are introduced as well.(2)In order to verify the effectiveness of the STP model in antenna modeling,an adaptive Bayesian optimization algorithm is used to optimize the hyperparameter of the STP.Furthermore,an improved student’s T-process model was used to model and predict the resonant frequencies of rectangular and circular microstrip antennas,verifying the effectiveness and accuracy of the improved model in predicting antenna resonant frequencies.(3)Furthermore,an improved Bayesian optimization algorithm based on STP was studied.Firstly,the structure of the proposed improved algorithm was explained,and the designed collection function was introduced;Next,the performance of the improved algorithm was verified on the minimum value problem of multi extremum functions;Finally,the algorithm was applied to the size optimization problem of the antenna,and resonant frequency modeling was performed on the printed dipole antenna and the E-shaped antenna,respectively.The optimal size results of the antenna were obtained,and the results were verified on the full wave simulation software HFSS,verifying the effectiveness of the improved algorithm.(4)On the basis of completing the optimization problem of antenna single design objective in Part(3),a model based on multi-output student’s T process combined with improved Bayesian optimization algorithm is studied to achieve multi-objective optimization of antennas.Firstly,the implementation process of MOSTP in multi-objective optimization problems is introduced,and validation experiments are conducted on multi-objective function problems to verify the excellent ability of multi output student T process in multi-objective problems;Next,this thesis elaborate on the implementation process of combining the multi output student T process with the improved Bayesian optimization algorithm.This thesis improved and proposed a two-stage algorithm model.The improved algorithm will be experimentally validated on the multi-objective optimization problem of antennas.The algorithm will obtain the optimization results of complex planar multi band antennas and mountain shaped defect antennas based on predetermined design indicators,The experimental results obtained will be validated on full wave simulation software to demonstrate the effectiveness of the improved algorithm in multi-objective antenna optimization design.The methodology in this research combines BO with single-output and multi-output Student’s T Process modeling,designing a new sampling function,which can obtain satisfactory modeling accuracy within a short time,and promote multi-objective rapid optimization design of complex electromagnetic structures included antennas further. |
