【作者】 孙娜；

【导师】 杨义先；

【作者基本信息】 北京邮电大学 ， 信号与信息处理， 2010， 博士

【摘要】 通信电台个体识别是通信对抗领域的研究热点之一。通过对电台信号细微特征的研究,可以获得信息特征、信号格式和目标个体特征等诸多信息,是进行数据融合、破译解码、模式/目标识别与跟踪、态势分析的重要技术手段。长期以来,对通信信号指纹特征的识别不论在军事或民用领域都具有重要的意义。但与早一步发展起来的雷达信号识别不同,由于电台个体差异极小,特征提取困难,通信电台个体特征研究目前还处于初步研究阶段,识别效果不佳,与实际应用存在很大差距。本文摒弃电台个体识别研究中流行的暂态信号分析方法,重点对通信电台稳态信号进行分析,对电台信号的常规特征与高阶统计特征在分类中的重要度进行了系统地评估,提出一种有别于以往研究思路的基于高阶谱核函数的特征提取与分类算法,采用邻域粗糙集算法进行特征集的约简,以支持向量机作为分类器,构造了二类分类器和组合分类器。提出的所有方法都用外场实测数据做了仿真。由于扩频电台信号的分析与处理方式与一般信号有所区别,且难度更大,文中还特别对在军事通信中最为通用的跳频信号的细微特征研究进行了初探。本文的贡献主要包括以下几个方面：1)研究特征重要度评估方法。提出基于邻域粗糙集的特征评价法,对信号的常规特征(载频偏差、信号的调制参数、电台的杂散特征)和高阶统计特征分别作重要度分析,以所提取特征的实验数据为基础,得到各特征在分类中的重要度参数。这种方法证明了各特征的重要度有明显差别,在分类中应予以区别对待。与大多数研究不同,对各特征区别对待以及赋予权值的思想在文中的很多方面都有所体现。2)针对高阶谱用于分类时会引发“维数灾难”进而导致的性能和效率问题,提出一种基于高阶谱核函数的特征提取与分类算法,此方法与目前流行的双谱(主要是积分双谱和选择双谱)分析方法思路迥异,利用高阶谱核函数,将低维的不可分样本映射到高维空间,从而有效降低支持向量的个数,提高样本的可分性,达到提高识别率的目的,同时还兼顾了分类正确率和分类效率的平衡。此外,对高阶谱核函数中阶数的选择进行研究,通过实测数据的分析结果表明,分类器的正确率与高阶谱的阶数并不成正比,因而一味地提高高阶谱阶数是不明智的。通过特征评估结果与高阶谱阶数分析结果的一致性表明,特征评估可以作为高阶谱核函数阶数选择的一种参考方式,并基于这个思路构造了优化的加权高阶谱核函数。3)研究电台信号特征选择及分类器设计方法。引入基于邻域粗糙集的数据集约简算法,对文中所定义的细微特征集进行降维处理以获得优化特征子集。在SVM分类器设计中,将各特征重要度作为权值构造加权特征以提高识别率,并利用各特征的重要度设计加权投票组合分类器,实验结果表明,基于加权投票组合分类器的效果明显优于单个分类器,能够取得较为满意的分类结果。4)研究跳频信号细微特征。用小波包分解与重构的方式抑制WVD方法产生的交叉项,进而提取跳频信号的参数特征,并基于核函数的原理,提出小波变换核函数算法。通过实测信号的实验对相关参数的优化给出了说明。实验结果表明,小波抑制WVD交叉项法取得的跳频信号参数特征有一定的聚类性,且在参数选择合适时小波核函数用于跳频信号分类可以得到较高的正确率。更多还原

【Abstract】 Identification of individual communication transmitters is one of the hotspots in the field of communication countermeasure. Researches on the subtle features of communication transmitters have gained some progress with the increasing of new techniques used in signal processing. Information such as signal features, types and individual features can be derived from the research on the subtle features maintained from the transmitters, which is an important measure of decoding, data fusion, pattern recognition, object tracking and situation analyzing. Analyzing of the subtle features from the individual transmitters has played an important role in either the military or civil field for a long time. However, differences of the technical parameters between individual transmitters are fine, thus analyzing of subtle features is rather difficult, which is a little different from that of radar.In this paper, the characteristics of stable signals, such as carrier frequency offset, modulation parameters and stray features of the transmitters, and their significances in the course of the identification are both studied systematically. Thereby, a novel feature extraction and classification algorithm based on polyspectral kernel function is presented, in which polyspectral theory and kernel technique are combined, and support vector machines (SVMs) are chosen as the classifiers. Feature sets are reduced by the neighborhood rough set algorithm. Furthermore, two-class classifiers and combined classifiers are designed based on the discussion above. The proposed algorithms are verified to be efficient using the measured radio data. Besides that, researches on the subtle features of the frequency-hopping spread spectrum signals are carried out.The main contributions of this paper are concluded as follows.1. Evaluation method of the significance of the features based on the neighborhood rough set is studied. General signal characteristics including carrier frequency offset, modulation parameters, and stray features, and high order statistical characteristics of the transmitters are studied in this paper. The results show that general signal characteristics play an important role in the signal identification. We can get the significance of every feature in classification on the basis of the experimental data. It is proved that the significance of every feature is obviously different, thus they should be treated distinctively. The distinctive treatments of every feature and thinking of the weight are well reflected in the paper.2. To address the curse of dimensionality caused by the higher order spectrum used in classification, a new feature extraction algorithm based on poly spectral kernel function is presented. The method can reduce the numbers of the support vectors effectively, and shorten the training and classifying time in comparison to the bispectral analysis (including integral bispectral and selected bispectral). Moreover, it can improve the identification rate.Then how to choose the order of the polyspectral kernels is studied. The neighborhood rough set theory is proposed for evaluating the significances of the order of the polyspectral kernels, which raises a novel weighted polyspectral kernel. The experiment shows that better and stable classification rate can be achieved.4. The selection of the signal feature and the design of the classifiers are discussed. Data set reduction algorithm based on the neighborhood rough set theory is proposed in the selection of the signal features. An optimized feature subset can be obtained through reducing the dimension of the subtle feature set defined in the paper. In the design of the SVM classifiers, the significance of every attribute is used as the weight to construct weighted feature in order to improve the identification rate. And the weighted voting combination of multi-classifiers are designed on the basis of the significance of every attribute. The experimental results show that the efficiency of the weighted voting multi-classifiers is better than the single classifier.5. Subtle features of the frequency hopping signals are studied. Wavelet packet decomposition and reconstitution are used to constrain the cross-term brought about by WVD method. Then the extraction of the feature from the frequency hopping signals is carried out. Wavelet kernels function algorithm is put forward based on the principle of kernel function. Relevant parameter optimizations are discussed through the experiments on the measured radio signals. The experimental results show that features of FH signals extracted by the way of Wavelet constrained WVD cross-term have a certain clustering performance. A pretty good identification can be achieved if the appropriate parameter is chosen.更多还原