Research On The Design Method Of The Microwave Filter Based On Neural Network

In modern wireless communication technology,microwave filter as a critical frequency selection device,its performance directly affects the quality of the communication system.As the filter structure becomes more and more complex,the traditional design method based on electromagnetic simulation software is relatively accurate,but the calculation cost is high and time-consuming,which undoubtedly increases the difficulty and cost of filter design and restricts the development of filters.In this dissertation,the problem of how to achieve fast and accurate filter design will be studied in depth in combination with artificial neural networks,and results are as follows:(1)A filter design method based on coupling matrix extraction.To address the problems of slow and low accuracy of coupling matrix extraction in the filter design process,the one-dimensional convolutional neural networks(1D-CNN)is proposed to apply in the extraction process of coupling matrix.First,this paper adopts the relational formula of S-parameters and coupling matrix for data collection,which greatly reduces the time consumed in data collection.Then the mapping relationship between S-parameters and coupling matrix is learned by designing the structure of1D-CNN model and network parameters for accurate and fast extraction of coupling matrix.Finally,the fast design of the filter is assisted by the extracted coupling matrix.The experimental results show that the proposed method can extract the coupling matrix quickly and accurately,which is generally better than the traditional extraction methods.(2)One-dimensional convolutional autoencoder(1D-CAE)based surrogate modeling techniques are researched.Surrogate models are often used as an effective tool to replace full-wave electromagnetic simulations in the design optimization of microwave devices.The 1D-CAE model proposed in this dissertation can establish the mapping relationship between filter design geometric parameters and electromagnetic response,replacing the traditional electromagnetic simulation.1D-CAE model realizes the high-dimensional modeling of filters by encoding and decoding the data for dimensionality reduction and reconstruction.Compared to other neural network structures,the 1D-CAE model is able to build highly accurate and computationally fast surrogate models with fewer data sets to assist in filter optimization.(3)In the optimal design of surrogate model-based filters,the optimization-seeking strategy is critical to reduce the number of surrogate model calls and search for the global optimal solution.In this dissertation,two optimization-seeking methods are introduced: particle swarm algorithm(PSO)and multi-objective evolutionary algorithm based on decomposition(MOEA/D).First,the1D-CAE model replaces the direct call to the electromagnetic simulation software in the PSO algorithm with a cheap surrogate model of low computational complexity obtained by training.Meanwhile,the PSO algorithm updates the 1D-CAE model by resampling the optimal solution of the latest iteration as the initial value every few iterations,and improves the 1D-CAE model prediction accuracy during the iterative process.Second,for the limitations of PSO algorithm in dealing with multi-parameter and multi-objective problems,this paper also proposes MOEA/D combined with1D-CAE model for microwave filter optimization design method.Finally,the effectiveness of the above methods is verified by filter design examples,respectively.