Research On Modeling And Optimization Of Passive Microstrip Circuits Based On Machine Learning

Microstrip circuit is a common circuit form in modern wireless communication system and has the features of small,light and compact.Traditional microstrip circuit design methods rely heavily on full-wave electromagnetic simulation software,which often requires huge manpower,computer resources and time costs in the optimisation process and is no longer adapted to the development of current electronic technology.In recent years,machine learning methods such as Gaussian process(GP)and artificial neural network(ANN)have been used by researchers in the analysis and modelling of microwave/RF circuits and are gradually gaining importance.The modelling methods with small sample and efficient optimization algorithms based on GP and ANN are studied in this thesis,which greatly improve the design efficiency.The main works of this thesis are as follows.First of all,a review of the main current machine learning methods is presented,and the basic theories of ANN and GP are introduced.A particle swarm gradient descent algorithm optimized ANN is proposed and use it to model a microstrip vertical transition structure.Predicted accuracy as well as training time of ANN and GP are compared for different sizes of training samples,laying the foundation for further research.In the study of small-sample modeling of passive microstrip circuits,the modeling method of knowledge-based Gaussian process(KBGP)is investigated from lowfrequency and simple circuit structures,including difference-based and priori knowledge injected ANN-GP models.These two models are compared and analyzed by using a microstrip unequal power divider example and a low-pass filter example,on the basis of which a synergistic KBGP modelling scheme combining the two models is proposed.When the operating frequency of microstrip circuits increases or the circuit structures become complex,the discrepancy between the electromagnetic simulation and the equivalent circuit will gradually increases.In order to further investigate the modeling method that still has a good performance in the high frequency and complex cases,based on the KBGP modeling method,we propose an input-output Gaussian process space mapping model that trained by a semi-supervised learning algorithm to model high frequency microstrip circuits with small samples,which combines the space mapping technique and semi-supervised training algorithm.The method is validated by two examples of X-band microstrip filters.In terms of efficient optimization problems for passive microstrip circuits,an efficient optimization algorithm for microstrip circuits named Gaussian process half-space mapping is proposed.By determining the optimization direction,the optimization space is halved,and the circuit structure that satisfying the design specifications can be optimized with only a few iterations.In addition,considering the practical designing problem of microstrip circuit with many design parameters,large optimization range and difficulty in obtaining the priori knowledge,a generalized particle swarm Gaussian process algorithm is proposed.The algorithm needs only a small number of samples to train the initial GP,followed by the particle swarm algorithm to optimize it during each iteration,and electromagnetic simulation software to verify the optimal results predicted by the model then retrain the GP until it finds a circuit structure that meets the requirement of the specifications.With the help of this algorithm,the E-type dual-band patch antenna and the ultra-wideband microstrip slit antenna are optimally designed.