【作者】 苏涛；

【导师】 保铮；

【作者基本信息】 西安电子科技大学 ， 信号与信息处理， 1999， 博士

【摘要】 大运算量的科学计算以及高速实时信号处理靠传统的单处理机系统已无法完成，必须通过并行处理技术实现，特别是新一代的雷达信号处理系统功能强、结构和信号处理方式都很复杂，对实时性、运算精度、动态范围和数据吞吐量提出了更高要求，采用每秒几十亿至几百亿次浮点运算速度的大规模实时并行处理系统势在必行。本文主要讨论了针对雷达信号处理如何设计和构成高性能的实时并行系统，并对两种雷达信号处理应用作了实用性的研究和实现。 并行处理机的设计是一门综合性很强的应用领域，涉及到算法研究、VLSI设计理论、系统结构、网络拓扑等多个方面，第二章中，我们首先建立了一个能全面反映并行处理机性能的时间模型，在此模型的基础上，得到了影响并行处理机两个基本性能—加速比和并行效率的三个重要因素：处理单元性能，并行处理机网络结构，并行算法和任务分配方法；然后从雷达信号处理的角度分别对这三个因素做了分析，确定了通用并行数字信号处理器（DSP）是完成雷达信号处理机的最佳选择；接着从实用角度出发，分析了分布式处理和共享总线处理两种基本网络结构在不同处理机规模时所能达到的加速比和并行效率，确定了不同处理机规模时的网络结构选择；接着讨论了在并行处理任务特征不同时的两种任务分配策略—静态分配和动态调度方法所能达到的性能。 第三章对雷达信号处理的各种主要运算类型的运算量和数据通信量作了统计和分析，用时频域变换方法在频域快速实现了长阶数横向滤波，表明现在的通用DSP在雷达信号处理的各种类型的运算中都可以取代专用DSP，接着我们基于两种高性能的通用DSP设计和制作了多种并行处理模块，模块具备可重构的特点，能用于多种类型雷达信号处理。 第四章针对雷达信号处理的瓶颈—矩阵分解算法研究了脉动阵/波前阵的实现问题。作者对Cholesky分解和QR分解算法用于自适应波束形成的可实现性做了比较，用通用DSP以较高的速度实现了脉动阵/波前阵的自适应波束形成，并且针对某些需要显式求取自适应权的应用，设计了一种能流水提取自适应权值的高效率波前阵结构，证明了通用DSP完成脉动阵、广义旁瓣相消、自适应权值提取的有效性。 第五章研究并提出了用全通用DSP模块设计空时二维自适应信号处理机的新方案，首先用快速付立叶变换和对长序列截断处理的重叠保留法以大大减少的运算量完成了处理机通道处理部分的正交插值、脉压、自适应均衡三个功能块的长阶数线性卷积运算，接着对自适应波束形成提出了多种实现方案，分析了多种方案下的运算量和设备要求，指出在不同通道数、不同数据吞吐率、不同自适应要求时这几种方案有不同的可行度和实现难度，并针对实际要求确定了一种运算田N、的实现方案，对自适应波束合成部分的大量贝玉以良元组咸肋冈络结拘和处凰闺和羽凶挝娠昨了优化设计以威少设备复杂性，构成了设备量大夕父伯小、们g团回高、辜于生回阴田闩引愧快的空时二维自适囱戏升引毗四机。 第六章进行了龄肌径地毗USAR）实时狮的研究和实砚，首先根据目标槽向成像原理，分析了各回浚间羹瞩0氧元的相伍瞎日以市，研究了侠速自司og歇回波抽取间隔的掀，并用第三章设计的并行处理板块完成了附飞行目标的ISAR实时成但；接着分析了机动飞行目标成倡田闪宇在的间回和解抉J引日，并确定了一种 量小、无需人工于预的自遁脑沮爿刽围离团恻固巧n阮尤法，林J劲印并 率出发，对这和一h法中的 卜确定性运算，用动态同度方法设计了一种分布 案，威少了处理系统的设备量。 第七章对工作颁了总结，并对雷达信号的实隙觎－和娜腑作了腿。 本文工作的主要目的是 憎号处娜点，蛐是针对B地我们捕的重要深图，瞅如何以高性能 域的高跟暇岖，提出了栅的样机研制方柔并作了初步实现。更多还原

【Abstract】 The massive processing volume of large scale scientific computation and high speed real-time signal processing should be realized by parallel processing technology instead of the conventional single-processing systems with limited processing capability. Among the applications needing massive processing, the advanced radar signal processing systems with powerful function and complicated construction and sophisticated processing manner need fast response, high data precision, wide dynamic range and very high data throughput besides the processing speed of billion to tens of billion floating-point operations per second. The paper is mainly on the question of design and construct real-time parallel processing system used in radar signal processing, and is responsible of proposing schemes of the realization of two kinds of application in this region. To design a parallel processor means much synthesis work in algorithm research, VLSI design, system configuration and network topology. In Chapter 2, the author begins with establishing a timing model reflecting the performance of parallel processor, with this model, the three main factor having important influence on the performance are displayed: the processing element, the configuration of network, the parallel algorithm and task allocation method. We can see that the general purpose DSP are the best choice of processing element for radar signal processing, the distributed system and shared-bus system will present different performance in differed processing dimension. And we prefer one kind of task allocation method than the other depending on the characteristic of task. In Chapter 3, a statistics of computation and communication burden on the main kinds of radar signal processing is given, a method of realizing linear convolution of time domain in the frequency domain is proposed. All these prove that the general purpose DSP is capable of processing all kinds of radar signal processing. Then the author designs several kinds of parallel processing module which can be used in a wide area of real-time processing. The bottleneck of radar signal processing is matrix decomposition which is more complicated. In Chapter 4, A detailed comparison between the realization of Cholesky decomposition and OR method is given, a systolic/wavefront array made of general purpose DSP is used to realize the adaptive beamforming at a moderate speed, and a wavefront array to extract the adaptive weights which is necessary in some applications is designed. In Chapter 5, the author proposed a new structure to realize the space time adaptive processing. First 1~q using the FFr and the truncating infinite sequence and sone optimizing method, a much simpler architecture can be acquired in realizing the three portion of channel processing: quadrature sampling, pulse compression and adaptive equalization. Then proposed several methods in adaptive beamforming and analyze their performance, complexity and speed under different data sampling speed, channel number and adaptive requirement. Aiming at the actual in-progress system designing, one method is applied using the unified type of parallel processing module. To decrease the hardware ccm~lexity, the author optimizes the network connection and data conununication scheme. So a STAP parallel processor with much less equipaent, higher performance and high degree generalization is acquired. In order to realize the real-time Inverse Sy更多还原