Synthesis And Optimization Design Of Microwave Filters Based On Neural Network

With the development of wireless communication technology,microwave components are required towards wideband,high frequency,low cost,etc.,but the available time is less.Facing the rapid development of science and technology,a set of high-efficiency design and high-precision methods for microwave wideband filters are particularly important.In this dissertation,the bandpass filter with wideband based on multiple-mode resonator realized by parallel-coupled line is selected as the research object,a synthesis model for the filter is constructed by utilizing BP neural network,and the synthesis and optimization of non-equiripple response are analyzed and studied.The prime contents of this thesis are summarized as:(1)A synthesis method of wideband filter based on BP neural network is proposed.Physical dimensions of the desired filter can be directly synthesized from the design requirements by this method.A complete theoretical synthesis method for arbitrary symmetric non-equiripple responses is derived,based on the synthesis theory of Chebyshev responses.In addition,a data collection scheme is proposed based on this complete synthesis process to collect sufficient training data for neural network.A BP neural network that is good at function approximation is used to express the mapping relationship of synthesis.Data collection,training and performance evaluation of the network model are studied.Finally,a neural network synthesis model with simple structure,high precision,high training speed and response speed is completed.(2)Furthermore,two methods are presented to optimize non-equiripple responses of multiple-mode filters with parallel-coupled line,which can solve the frequency shift and reflection lobe distribution problems in the initial simulation.Based on the analysis of multi-mode filters' frequency response characteristics and trend,a modification formula for the lower cut-off frequency is proposed,and the synthesis model is used to achieve the optimal design,which solves the frequency shift problem in the initial simulation.In addition,modification method is studied and specific modification formulas are proposed to optimize physical dimensions and the lower cut-off frequency of the desired filter by analyzing the causes of differences between the initial simulation and the target one.The optimized results show a great agreement with the desired one.Effectiveness of the designed non-equiripple response to improve measured performance is verified by compared with measured data.Compared with the traditional method,the proposed design method can avoid a lot of repeated calculation of intermediate processes,and one neural network can be used to achieve synthesis and optimization while ensuring high accuracy.After the neural network is finished,physical parameters are converted into pure mathematical data,which reduces the requirement of learning knowledge bfore design.The proposed optimization methods can greatly enhance the controllability of designing non-equiripple waveforms.