【作者】 陈筱倩；

【导师】 王宏远；

【作者基本信息】 华中科技大学 ， 信息与通信工程， 2009， 博士

【摘要】 空间中各种无线电信号的模式、频率以及功能复杂化，形式多样化，客观上要求无线电终端能够适应多模式、多频段、多功能环境，因此期望能够在使用、规划和创建无线电过程中，事先认知并描述复杂空间的无线电状态(RKRL：Radio KnowledgeRepresentation Language)(包括波达方向、频率、功率、调制制式、信道编码、传输帧结构等)，进而对这个复杂的状态空间进行合理的解析，从而作出理性选择，形成了“认知无线电”的总和，作为一个不可缺少的内在功能要求，调制制式的自动分类也开始得到关注。截获信号通常被埋藏在噪声中(高斯白噪声、脉冲噪声等)，传统的决策理论和统计模式识别的方法，分类信息的利用较低，特征提取带有较大的盲目性，且受噪声影响大，因而识别性能不太理想。如何在特征参数的提取中减弱噪声和干扰的影响，设计具有较强适应能力和容错能力的分类器，是研究的主要方向。本文在前人工作的基础上，围绕“无线电认知研究”中调制分类子课题，在信噪比变化较大，存在干扰信号的情况下，对数字通信信号调制制式自动分类预处理过程的改进、新的有效分类特征构造，分类器建模的自适应优化算法理论基础和实现进行了深入的研究，形成了一套有效的数字通信信号调制制式分类解决方案。所做的工作主要包括：首先对调制分类前预处理的几个关键环节整合的基础上，分析了信号序列边界处瞬时频率估计以及精确的载频估计方法，设计了一种模糊加权滤波算法，可较好地保护信号的细节信息，为提高分类性能打下了好基础。接着讨论了有效分类特征的选择与提取，基于对广义相关函数形式的分析，引出高阶正交累积量和高阶正交循环累积量，并根据理论分析和测试选择性能最优的分类特征构造形式，新特征实际上定义了一种调制信号同相分量与正交分量的高阶相关统计量，蕴涵更多调制制式的本质结构特征，实验证明具有明显的抑制噪声和干扰的效果。利用模糊聚类方法确定了特征空间的聚类中心等信息，可作为后续分类器设计的参考。最后利用神经网络的学习机制实现调制制式自适应模糊分类器的非线性动态建模，先根据特征训练样本的大致分布状况建立蕴涵初始经验的并联模糊分类器，再采取分层决策的级联结构，提高了特征与分类器的契合度，最大程度上减少了隶属度函数和模糊规则的冗余。蕴涵初始经验的模糊推理系统，其知识推理结构明确可控，通过样本训练实现结构参数自适应调整和优化，完成其逼近求精，系统在信噪比等环境参数变化较大情况下具有更好的适应性和容错性，正确识别率和效率相对神经网络识别器和模糊识别器有明显提高。更多还原

【Abstract】 The various signal pattem,frequency and function of radio are more complicated inspace,and the form diversifies,objectively,require that the radio terminal is able to adapt tomany patterns,multifrequency,multi functional environment.Therefore,in the process ofusing,planning and creating,the radio consumer longs to be able to predict and describ thestates of complicated space radio signals using RKRL(Radio Knowledge RepresentationLanguage),include DOA(Direction Of Arrival),frequency,power,modulation format,signal channel encode,frame structure,etc..Then analyse on this complicated state spacerationally and make reasonal choices.All the above issues form the concept of“CognitiveRadio”.As a necessary function module,Automatic modulation classification has attractedmore attention.The intercepted signals are often concealed in noise including Gaussianwhite noise and impulse noise.Moreover,since the former methods of statistical patternrecognition and decision theory are mostly used,the information of classification isn’tadequately utilized.Sometimes,the features are extracted subjectively and easily affected bynoise,so the classification performances aren’t satisfying.How to reduce the influence ofsignal noise ratio and interference in features? How to design the classifier with greateradaptability and fault tolerance? All of these problems are the topic way of the paperresearch.Based on the previous works,this thesis has engaged in extensive research on Sub-topicof modulation foramat classification in“Radio Cognition Studies”.Considering the largevariation range of communication signal’s SNR and the interference from other signals,themodified preprocessing,the novel classification features with excellent quality and theoptimal adaptive classifiers are presented for automatic digital modulation classification.Finally an improved global solution is formed.The main works can be summarizedasfollows:Firstly,several key problems of the preprocessing of modulation classification areintegrated into a part.The instantaneous frequency estimation method at signal sequenceboundary and the accurate carrier frequency estimation method are analysed.A fuzzyweighted filter algorithm with preferable signal details protection capability is presented,which is contributed to improve the classification performance.Then,the selection and abstraction of classification features are discussed.This thesisintroduces the high order orthogonal cumulant (HOOC)and high order orthogonal cycliccumulant (HOOCC)based on the analysis of generalized correlation function.The optimalstructure form of the classification features are selected according to the theoretical analyseand simulation estimation.The novel features actually defines a high Order correlation statistics of the in-phase component and quadrature component of the modulated signal,which contains more essential information of modulation formats and has evidente effect tosuppress noise and interference.The clustering centers of the feature space,which can be thereference information for modulation classifier design,are determined by the fuzzyclustering algorithm.Finally,the non-linear dynamic modeling of an adaptive fuzzy modulation classifier,which based on the training mechanism within the neural network,was presented.Themodel adopted the parallel and hierarchical decision-based structure,which made thefeatures match the classifier and reduced the redundancy of the membership functions andfuzzy rules.The system with initial experience guarantees the controllability of theknowledge inference structure.By applying the training data,the algorithm adaptivelyadjusted and optimized the structure parameter and completed the approximation process.The simulation results verify the better adaptability and fault-tolerance of the system in thepresence of various environment parameters (SNR etc),as well as the improvement of theaverage probability of correct classification and the algorithm efficiency,compared with theneural network classifier or fuzzy classifier.更多还原