【作者】 张伟；

【导师】 袁业术；

【作者基本信息】 哈尔滨工业大学 ， 信息与通信工程， 2006， 硕士

【摘要】 通信信号调制方式的识别是通信侦察的关键技术之一,是目前电子战的重要研究课题。本文针对通信信号的调制识别问题,深入研究了信号的特征提取方法,在此基础上探讨了分类器的设计方法,其中特别提出了由分形、L-Z复杂度、熵和相像系数组成的特征集并设计了基于最小二乘支持向量机算法的分类器。本文首先介绍了调制信号瞬时幅度、瞬时频率和瞬时相位的计算方法,为提取有效的分类特征奠定基础,进而介绍了常用的调制技术,给出一些常用调制信号的数学模型,考察了基本模拟调制信号和数字调制信号的时域波形和频谱图。在特征提取部分研究了信号的零中心瞬时特征、小波分析特征以及由分形、L-Z复杂度、熵和相像系数组成的特征集。仿真实验表明3种特征提取方法均能提取出有效的分类特征。通过绘制特征分布图重点考察了分形等特征的特性,结果表明本文提出的由分形等特征组成的特征集类内聚集程度和类间分离程度高,能够在低信噪比时获得较高的信号识别率。分类器设计部分首先研究了神经网络分类器的设计方法,然后重点研究了支持向量机分类器,详细推理和分析了支持向量机算法,构建了多类别分类支持向量机。支持向量机算法是基于结构风险最小化原则的,其目标是得到现有信息下的最优解而不仅仅是样本数趋于无穷大时的最优值,因此在小样本情况下具有很大的优势;另一方面,该算法将特征向量映射到高维空间中加以分类,解决了样本在低维空间中的非线性不可分问题,避免了判决门限的确定,与传统的神经网络方法相比,具有更好的泛化推广能力。在仿真实验中,提取信号的小波分析特征,针对BP网络的不足采用改进的BP网络和径向基函数神经网络分类器进行分类,实现了多种信号在较大信噪比变化范围内的自动识别,并在大样本低信噪比条件下取得满意的效果。仿真研究了最小二乘支持向量机分类器在不同观测数据长度时、采用不同核函数和不同分类方法时的性能,得出最小二乘支持向量机分类器对不同的模型具有一定不敏感性的结论。最后对比了低信噪比情况下和小样本情况下神经网络和更多还原

【Abstract】 The modulation recognition of communication signals is one of the key techniques of communication reconnaissance, and it is also an important task in the field of Electronic warfare. This dissertation has a study on the modulation identification of communication signals. It introduces how to extract characteristics of signals and how to design classifiers. Especially, it gives a feature set consisting of fractal dimension, L-Z complexity degree, entropy and resemblance coefficient and designs a classifier based on an algorithm called least squares support vector machine.Firstly, the dissertation introduces mathematical models of common communication signals and simulates to get basic signal figures. It lays the groundwork for future feature extraction.It researches three ways of characteristics extraction of signals including zero-center instantaneous feature, wavelet analysis feature and a feature set consisting of fractal dimension, L-Z complexity degree, entropy and resemblance coefficient. The experiment shows that features are effective for classification. It discusses congregation degree among same kinds and separation degree among different kinds of signals through drawing feature distribution figures. It shows that the feature set consisting of fractal dimension, L-Z complexity degree, entropy and resemblance coefficient congregates among same kinds and separates among different kinds highly. It can gain a high recognition rate of signals.To design the classifier, firstly, it researches how to design neural network. Then it analyzes support vector machines classifiers in theory at length and gives some algorithms for classification question beyond two kinds. On the one hand, based on structure risk minimization least squares support vector machine algorithm can gain best results with existing information. The condition of infinite stylebooks is unnecessary. So the outcome is good based on this algorithm when the stylebooks are few. On the other hand, it maps key features into the high dimension space. The更多还原