【作者】 杨镭；

【导师】 肖怀铁；

【作者基本信息】 国防科学技术大学 ， 信息与通信工程， 2010， 硕士

【摘要】 高分辨雷达一维距离像包含有目标丰富的结构特征,在雷达自动目标识别技术领域具有独特的优势。建立在结构风险最小化原则上的支持向量机,具有泛化能力强、小训练样本、非线性、无局部极小等许多优良性质,使其成为高分辨距离像自动目标识别应用中的有力工具。在此背景下,本文研究了支持向量机多目标参数优选和支持向量机多阶段选择性集成的算法设计及在高分辨雷达目标识别中的应用问题。本文主要内容安排如下:第一章简要介绍了课题的背景、意义以及相关技术的发展和研究现状。第二章全面分析了高分辨雷达一维距离像和支持向量机的基本原理。本章推导了一维距离像的获取,结合目标姿态敏感性和平移敏感性等问题对一维距离像进行了特性分析,然后详细阐述了支持向量机的理论基础、基本算法及特点,为后续章节算法的设计与应用提供了广泛依据。第三章研究了支持向量机多目标参数优选算法的设计问题。本章分析了参数对算法泛化性能的影响,指出了参数优选的必要性。传统的参数优选是基于训练集上单个泛化误差界的,本文通过实验分析指出这是不充分的,参数优选是个多目标优化问题。本文提出了基于非支配排序遗传算法的支持向量机多目标参数优选算法,实验结果表明与单目标参数优选算法相比多目标参数优选的参数值比较适中并获得了较高的识别率。第四章研究了支持向量机多阶段选择性集成算法的设计问题。集成学习可以提高学习机的泛化能力,选择性集成可以在不损失甚至提高集成泛化能力的基础上大幅削减集成成员的数量,这对高分辨雷达自动目标识别是有益的。本文提出了基于遗传算法的支持向量机多阶段选择性集成算法,实验结果表明该算法能够在使用较少集成成员的条件下获得比传统Bagging集成和单个SVM更高的识别率。第五章对论文进行了总结,指出了下一步要做的工作。更多还原

【Abstract】 The high resolution range profile (HRRP) obtained by the high resolution radar contains abundant structure signatures of the target and embraces unique advantages in the field of Radar Automatic Target Recognition (RATR).The Support Vector Machines (SVMs) founded on the Structural Risk Minimization (SRM) principles has many good properties, such as better generalization, small training samples, nonlinear, no local minima, etc, which make it become a powerful learning machines in the application of ATR based on HRRP.In this background, the multiobjective parameter selection algorithm and the multistage selective ensemble algorithm for SVMs along with their application in the high resolution radar target recogniton have been researched by the paper.The main contents of the paper are arranged as follows:In chapter one, the background and significance of the subject, the progress of the related technology and the current status of the research are briefly introduced. Chapter two gives an extensive materials about the basic principles of the high resolution range profile and the support vector machines. The obtaining of HRRP is deduced,and the target-aspec and time-shift sensitivity problems are conclued to analyze the features of the HRRP samples, then the theory foundation, the basic algorithms and the characters of SVMs are detailedly stated, which give a comprehensive bases for the design and application of the algorithm in the successive chapters.In chapter three, the design of the multiobjective parameter selection algorithm for SVMs is studied.Based on the simulation, the necessary of optimal parameter selection is demonstrated by their direct influence on the algorithm’s generalization ability. The traditional parameter selection methods are usually conducted through a single generalization error bound , and that seems insufficient according to the results of our experiments. From our view, the issue of parameter selection should be treated as a multiobjectice optimization problem. A multiobjective parameter selection algorithm based on the nondominated sorting genetic algorithm (NSGA-II) is proposed in the paper, and the experiments’results show that compared with single-objective parameter selection algorithm the multiobjective parameter selection algorithm can obtain a better moderate parameter values which ensure a higher correct recognition rate.In chapter four, the design of the multistage selective ensemble algorithm for SVMs is investigated. Ensemble Learning can ensure a better generalization ability for a learning machine, and the selective ensemble can hugely reduce the number of the members that constitute a ensemble without lossing or even improving the ensemble’s generalization ability, which is a very helpful property in the application to the ATR based on HRRP.The paper proposes a multistage selective ensemble algorithm for SVMs based on genetic algorithm (GA). The simulation results indicate that the correct recognition rate achieved by the algorithm is higher than the bagging ensemble and a single SVMs while just a less number of ensemble members have been used.Finally, the dissertation is concluded in chapter five. Several aspects for future work are also pointed out.更多还原