【作者】 张宇伟；

【导师】 徐金平；

【作者基本信息】 东南大学 ， 电磁场与微波技术， 2025， 博士

【摘要】 随着宽带通信、高分辨率雷达和复杂环境电子对抗等现代电子系统的迅速发展,微波滤波器作为微波射频系统中的重要部件,面临着日益增长的低插损、高选择性、宽频带、小型化、低成本和快速设计等多方面的技术挑战。基于经典综合理论和电磁仿真软件的传统设计方法过程繁琐且效率低,难以满足复杂系统中多种多样的高性能滤波器的设计需求。因此,开展基于人工智能的微波滤波器高效、智能化的设计方法研究,对于器件性能的提升乃至系统层面关键技术的突破具有重要的现实意义。随着人工智能领域深度学习技术的迅速发展以及向各个应用领域的广泛渗透,人工神经网络(ANN)在辅助微波元器件设计中呈现出越来越强大的发展潜力,它能够自动完成参数优化,显著降低试错成本,为各种复杂滤波器的高效设计提供了新的技术路径。近年来,基于生成式对抗网络(GAN)的逆向设计方法在微波与光器件研究领域取得了重要进展,为超表面、新型天线单元等设计提供了全新的思路。针对现有基于ANN的逆设计方法中存在的解的非唯一性问题,本文探索了基于GAN的平面微波滤波器逆向设计方法,并重点研究了结合归一化的数据处理方法以及基于迁移学习的模型训练策略,构建了三种用于平面微波滤波器逆向设计的生成式模型,有效拓展了基于GAN的生成式模型在微波滤波器逆设计中的应用范围。另一方面,针对射频与微波系统中宽带信号源在功率平坦度和大动态范围控制方面的技术难题,本文深入研究了无反射式幅度均衡器、方向性补偿的宽带定向耦合器和数字/模拟反馈环路等宽带电平控制中的关键技术。本文主要研究进展包括以下几个方面:1、基于GAN中神经网络的对抗博弈机制,提出了一种基于条件深度卷积(CDC)GAN的平面微波滤波器逆向设计方法,有效地解决了逆向设计神经网络中解的非唯一性问题。将此方法应用于构建一种基于方形贴片谐振器的双频微带滤波器生成式模型,其中平面微带滤波器的电路原型被表示为像素化图案,并在GAN中引入了条件变量和卷积神经网络,构建了电路原型和S参数的映射关系,使得基于CDC-GAN的生成式模型在输入期望的S参数后,快速生成对应滤波器的电路原型。通过四个通带中心频率分别位于S/C和C/X波段的双频微带滤波器逆设计实例,验证了该模型的有效性,为双频平面微波滤波器的设计提供了新的设计思路。2、提出了一种基于归一化数据处理方式的生成式模型构建方法,有效拓宽了生成式模型的适用频段并实现了多种尺寸电路原型的逆向设计。通过对不同分辨率的像素图案及其外围电路开发不同的自动化建模程序,并采用归一化图案像素和S参数向量格式,将原来需要多个模拟器协同预测的神经网络架构简化为仅需嵌入单一模拟器的架构,从而有效优化了复杂逆设计问题中的生成式模型结构。基于改进的生成式模型架构和互补分裂环谐振器(CSRR)的双通带产生机理,提出了一种双频基片集成波导(SIW)滤波器的生成式模型,据此设计了通带位于5～18GHz的多种双频SIW滤波器实验样品,并成功应用于频率源系统中的谐杂波抑制。3、提出了一种嵌入多模拟器神经网络的生成式模型架构,并应用于可调微波滤波器的逆向设计。该架构解决了生成式模型在逆向设计平面电路时,单一模拟器无法预测电调参数等额外变量维度引起的电磁特性变化的问题。此外,在模型训练过程中引入了迁移学习算法,采用冻结模拟器部分卷积层权重,并重新训练全连接层权重的策略,大大减少了训练新增模拟器神经网络所需的数据量和计算资源。应用这两种改进方案,提出了一种基于半哑铃型谐振单元的可调低通滤波器(TLPF)生成式模型。该模型借助多个模拟器神经网络,能够有效预测变容二极管容值变化引起的S参数变化。利用该模型设计了截止频率在5～10GHz范围内相对调谐带宽为15%～38%的四种TLPF,并通过结合两个紧凑微带谐振单元,在覆盖12～28GHz的超宽阻带范围内实现了 17dB以上的阻带抑制度。4、提出了一种具有宽带无反射特性的幅度均衡器的设计方法,并结合基于数字/模拟反馈环路的自动电平控制(ALC)系统,有效改善了宽带系统中普遍存在的功率不平坦问题,并实现了大动态范围的电平精确调控。所提出的宽带幅度均衡器采用固定衰减器与开/短路枝节线相结合的电路结构,具备无反射、宽均衡范围和小型化等优点。将该宽带无反射均衡器应用于ALC系统,更有效地利用了检波器的电平识别范围,从而扩大了反馈环路的调控范围。采用基于比例-积分-微分(PID)的反馈环路控制方法并结合环外大动态范围程控衰减器,在0.1～26.5GHz超宽频带内实现了-90～-10dBm的输出功率动态范围和±1.5dB的功率平坦度。5、针对宽带微带定向耦合器难以实现高方向性的技术问题,提出了一种基于容性加载贴片和无反射幅度均衡器相结合的方向性补偿的宽带微带定向耦合器技术方案。在微带耦合线之间加载的容性贴片有效改善了奇偶模相速的不平衡度,从而增强了定向耦合器的方向性;通过在耦合端和隔离端引入宽带无反射式幅度均衡器,将传统单节微带定向耦合器约50%的1dB分数带宽拓展至130%以上。结合这两种改进技术,设计并制作了 L-C和C-K波段的20dB微带定向耦合器。测试结果表明,这些定向耦合器的1dB分数带宽分别达到137%和132%,带内方向性均超过16.6dB。应用这两种耦合器实现了一种基于差频接收架构的宽带自动电平控制系统,在1～26.5GHz宽频带范围内实现了-50～-10dBm的功率动态范围和±1dB的功率平坦度。更多还原

【Abstract】 With the rapid development of modern electronic systems such as broadband communication,high-resolution radar,and complex electronic countermeasures in complex environments,microwave filters,as important components in microwave RF systems,are facing increasing technical challenges in terms of miniaturization,low insertion loss,high selectivity,low cost,and rapid design.Traditional design methods based on classical synthesis theory and electromagnetic software simulations are cumbersome and inefficient,making it difficult to meet the design requirements of a wide variety of high-performance filters in complex systems.Therefore,research on efficient and intelligent design methods for microwave filters based on artificial intelligence is of great practical significance for enhancing device performance and even achieving breakthroughs in key technologies at the system level.With the rapid development of deep learning technology in the field of artificial intelligence and its wide penetration into various application areas,artificial neural networks are demonstrating increasingly powerful potential in assisting the design of microwave components.This design method can automatically optimize parameters,significantly reduce trial-and-error time costs,and provide new technological paths for meeting the complex filter design needs.In recent years,inverse design methods based on Generative Adversarial Networks(GANs)have made significant progress in the research field of microwave and optical devices,providing new approaches for the design of meta-surfaces,novel antenna units,etc.To address the issue of non-uniqueness of solutions in existing ANN-based inverse design methods,this dissertation explores a GAN-based approach for the inverse design of planar microwave filters.It focuses on data processing methods combined with normalization and model training strategies based on transfer learning.Three generative models for the inverse design of planar microwave filters are developed,effectively expanding the application scope of GAN-based generative models in microwave filter inverse design.On the other hand,to address the technical challenges of broadband signal sources in RF and microwave systems regarding power flatness and large dynamic range control,this dissertation conducts in-depth research on key technologies in broadband level control,such as reflectionless amplitude equalizers,directional compensation in broadband directional couplers,and digital/analog feedback loops.The main research progress of this dissertation includes the following aspects:1.Based on the adversarial game mechanism of neural networks in GANs,a planar microwave filter inverse design method based on Conditional Deep Convolutional(CDC)GAN is proposed,effectively solving the problem of non-uniqueness in solutions within inverse design neural networks.This method is applied to construct a generative model for a dual-band microstrip filter based on a square patch resonator,where the circuit prototype of the planar microstrip filter is represented as a pixelated pattern.Conditional variables and convolutional neural networks are introduced in the GAN to establish the mapping relationship between the circuit prototype and the S-parameters.As a result,the CDC-GAN-based generative model can quickly generate the corresponding circuit prototype of the filter after inputting the desired Sparameters.The effectiveness of the model is validated through four examples of dual-band microstrip filter inverse design,with passband center frequencies located in the S/C and C/X bands.This provides a new design approach for dual-band planar microwave filters.2.A generative model construction method based on normalized data processing is proposed,which effectively expands the applicable frequency range of generative models and enables the inverse design of multi-size circuit prototypes.By developing different automated modeling programs for pixel patterns of varying resolutions and their surrounding circuits,and using normalized pattern pixels and S-parameter vector formats,the neural network architecture that originally required multiple simulators to work together for prediction is simplified to one that only needs a single simulator.This effectively optimizes the structure of the generative model in complex inverse design problems.Based on the improved generative model architecture and the dual-band generation mechanism of Complementary Split-Ring Resonators(CSRR),a generative model for dual-band substrate integrated waveguide(SIW)filters is proposed.Various experimental samples of dual-frequency SIW filters with passbands ranging from 5 to 18 GHz are designed,and these have been successfully applied in frequency source systems for harmonic distortion suppression.3.A generative model architecture incorporating multiple simulator neural networks is proposed and applied to the inverse design of tunable microwave filters.This architecture addresses the issue in generative models where a single simulator is unable to predict electromagnetic property variations caused by additional variable dimensions,such as tunable parameters,when performing inverse design of planar circuits.Additionally,a transfer learning algorithm is introduced during the model training process,employing a strategy of freezing the weights of certain convolutional layers of the simulators while retraining the weights of the fully connected layers.This approach significantly reduces the amount of data and computational resources required to train the new simulator neural networks.Utilizing these two improvement strategies,a generative model for a tunable low-pass filter(TLPF)based on a semi-dumbbell-shaped resonator unit is proposed.This model,assisted by multiple simulator neural networks,can effectively predict changes in S-parameters caused by variations in varactor diode capacitance.Using this model,four types of TLPFs were designed with cutoff frequencies ranging from 5 to 10 GHz and relative tuning bandwidths of 15%to 38%.By combining two compact microstrip resonator units,a stopband suppression of over 17 dB within the 28 GHz was achieved.4.A design method for an amplitude equalizer with broadband reflectionless characteristics is proposed,and combined with an automatic level control(ALC)system based on digital/analog feedback loops,it effectively improves the common power flatness issue in broadband systems and achieves precise level control with a large dynamic range.The proposed broadband amplitude equalizer employs a circuit structure that combines fixed attenuators with open/short-circuit branch lines,offering advantages such as reflectionless performance,wide equalization range,and miniaturization.By integrating the broadband non-reflective equalizer into the ALC system,the detector’s level detection range is utilized more effectively,thereby expanding the control range of the feedback loop.By employing a proportional-integralderivative(PID)based feedback loop control method together with an externally positioned programmable attenuator with a large dynamic range,the system achieves an output power dynamic range of-90 to-10 dBm and a power flatness of ± 1.5 dB across the frequency band of 0.1 to 26.5 GHz.5.To address the technical issue of achieving high directionality in broadband microstrip directional couplers,a new approach is proposed based on the combination of capacitive-loaded patches and a reflectionless amplitude equalizer for directionality compensation.The capacitive patches loaded between the microstrip coupling lines effectively improve the imbalance of oddeven mode phase velocity,thereby enhancing the directionality of the directional coupler.A broadband reflectionless amplitude equalizer introduced at the coupling and isolation ports extends the traditional single-section microstrip directional coupler’s 1dB fractional bandwidth from approximately 50%to over 130%.Combining these two improvements,20 dB microstrip directional couplers for the L-C and C-K bands were designed and fabricated.Test results show that the 1dB fractional bandwidths of these directional couplers are 137%and 132%,respectively,with in-band directionality exceeding 16.6dB.A broadband automatic level control system based on a difference frequency receiving architecture is implemented using these two couplers,achieving a power dynamic range of-50 to-10 dBm and a power flatness of ± 1 dB over a wide frequency range of 1 to 26.5 GHz.更多还原