【作者】 赵悦；

【导师】 位寅生；

【作者基本信息】 哈尔滨工业大学 ， 信息与通信工程， 2010， 硕士

【摘要】 对空间谱估计技术进行研究的主要目的是提高空间信源方位角的估计精度以及分辨力,在众多的超分辨算法中,子空间类算法由于其得天独厚的优势:明确的物理概念和良好的估计性能而得到了更多的关注和应用。其中最具代表性也是最常用的两种方法为:多重信号分类算法(MUSIC)和旋转不变子空间算法(ESPRIT)。将子空间类算法在应用于实际中会发现存在着计算量大的缺陷,而在雷达信号处理中,对算法的实时性又有着较高的要求,虽然有着较好的估计性能,但难以进行实时处理制约了子空间类算法应用和发展。本文从并行处理的角度出发,对子空间类超分辨算法进行高效化进行。然而并行化的实现是并不是简单的将算法分割,而是要依赖于对算法本身的理解和实现步骤的划分。本文从经典的MUSIC算法入手,分析其性能和计算复杂度,对运算量主要集中的环节——特征分解提出改进算法。该方法针对MUSIC算法的计算复数域上运行,运算量大、不易于特征分解的缺陷,选择一种实值化预处理方法,该方法不仅可以在不影响算法性能的前提下将协方差矩阵化为实对称矩阵,有利于并行化算法的选取,而且处理本身即会减少算法的计算量。之后通过对比分析选取了两种适合并行化处理的特征分解方法,并对其原理进行了分析。在得到理论依据后,对MUSIC算法进行并行化实现:在构造协方差矩阵、特征分解和谱峰搜索三个阶段分别对其进行并行化改进,从而大大提高算法的实时性。其中在特征分解阶段,分别采用之前所分析的Jacobi方法和QR方法并对其进行并行化改进。通过仿真分析验证了并行算法的有效性,同时通过与经典的MUSIC算法的性能比较,对比分析这两种方法在估计性能及运算量方面的性能。在得到并行的MUSIC算法之后,本文将此方法推广到ESPRIT算法中。针对ESPRIT算法与MUSIC算法主要的不同之处,即在特征分解阶段ESPRIT算法涉及到对非对称矩阵进行广义特征分解,本文在QR方法的基础上进行改进,采用了Lanczos方法和带原点位移的QR方法,提出了一种适合ESPRIT算法的并行运算方法,通过仿真实验验证了该方法在应用中的可行性,并与ESPRIT算法进行对比分析其性能及运算量。更多还原

【Abstract】 The main goal of spatial spectral estimation is to study all the algorithms about increasing the accuracy, the angle resolution, and the operation speed in estimating the angles of the spatial signals within bandwidth.During many direction-of-arrival estimation algorithms, subspace-based method is widely used because of the clearness of the physical concept and well estimation performance. Two kinds of the most representative methods are: Multiple Signal Classification (MUSIC) method and Estimation of Signal Parameters via Rotational Invariance Technique (ESPRIT) method.Nevertheless, it needs huge quantity of computation, which baffles its further application in Modern Radar signal processing which has a high requirement on the speed of processing. Therefore, it is very necessary to study its high-speed implement methods. This article focuses on highly effective of high resolution approach from the aspect of parallel method. The parallel algorithms for sensor array processing algorithm of high-speed realization are complicated, not only depend on the performance of the chips, but also rely on the steps of the algorithm itself.Starting from MUSIC algorithm, this article made improvement in the stage of covariance matrix decomposition. After analyzing the structure of the array, found a simple and effective method of real value preprocessing, which changed this method from the plural field to the real field, thus reduced the computation greatly with little performance loss. Then selected two eigenvalue decompose methods which are suitable for prarllel proposing.According to the theoretical basis, this article developed parallel architecture for MUSIC algorithm, during the three stage of MUSIC algorithm: data covariance matrix stage, covariance matrix decomposition stage, power method stage, makes the parallelization improvement separately, thus improve the original method in real time. The effectiveness of the parallel algorithm has been verified through the simulations. At the same time, the advantages and disadvantages of the parallel algorithm has discussed by compared with the MUSIC method.After obtaining the parallel MUSIC algorithm, this article extends this approach to the ESPRIT algorithm. The parallel ESPRIT algorithm and the parallel MUSIC algorithm are similar except the stage of decomposition, where the ESPRIT algorithm involves decomposition of a generalized asymmetric matrix. This article has used the Lanczos method and the belt zero point displacement’s QR method based on the QR method’s foundation. The simulation results indicate the feasibility of the application of this method.更多还原