【作者】 郭双冰；

【导师】 肖先赐；

【作者基本信息】 电子科技大学 ， 通信与信息系统， 2003， 博士

【摘要】 大量的自然现象显示出复杂的、不可预测的、伪随机的行为，而混沌动力学对其中许多现象提供了相对简单十分合理的解释。即使简单确定的系统也能产生复杂、伪随机的行为。混沌理论的目的是揭示貌似随机的现象背后可能隐藏的简单规律，以求发现一大类复杂问题普遍遵循的共同规律。随着混沌理论的发展，人们对时间序列预测的复杂性有了更深刻的认识。对混沌行为即使近似的长期预测也是不可能的，但短期内却可能做到准确的预测，因而研究混沌时间序列的预测可以解决以前许多难以解决的问题。此外，混沌理论提出了不同的特征参数对混沌时间序列进行刻划和描述，而通信信号也是一种非线性时间序列，调制类型的差异表现在时间序列的差异上，因而可以应用混沌理论对通信信号调制类型来识别。本文围绕混沌时间序列预测和混沌应用在通信信号调制类型识别两个方面内容展开研究。主要内容为：（1）非线性自适应滤波器对混沌时间序列及跳频码的预测；（2）常用的神经网络（BP网络RBF网络）的建模和对混沌时间序列及跳频码的预测。（3）一种对混沌时间序列的快速预测方法；（4）基于混沌理论提取通信信号调制类型的特征参数；（5）利用模糊测度和模糊积分理论设计了组合分类器。本文的创新之处在：1. 研究了二阶volterra自适应滤波器的收敛性；由于volterra自适应滤波器的预测性能主要依赖于自适应算法，而自适应滤波器的收敛性是评判自适应算法好坏的一个重要标准，因此研究了：控制参数与预测误差的收敛关系，初始滤波器系数与预测误差收敛的关系，滤波器系数之间的收敛关系，预测误差与滤波器系数的收敛关系，滤波器阶数m与滤波器系数h(n)的收敛关系。2. 非线性自适应滤波器的定阶；由于基于非线性自适应滤波器技术提出的非线性自适应滤波器的阶数是人为确定，而滤波器的阶数不但影响时间序列的预测精度，而<WP=6>且关系到实现该滤波器的计算复杂度。从理论和实验得出非线性自适应滤波器（包括二阶和高阶自适应滤波器）的最优维数，并给出一种算法确定滤波器最优输入维数。3. 针对实际工程应用要求（实时性），提出了对混沌时间序列的一种快速预测方法；利用hopfield神经网络具有快速的收敛性，可利用软硬件实现的优点，提出了基于hopfield神经网络的快速预测方法，该方法把二阶volterra预测滤波器和hopfield神经网络结合起来构成了新的预测方法，该方法对工程上要求实时性的问题提供了一种解决方案。4. 跳频码的动力学特征分析；跳频码是由确定性的规则产生的伪随机序列。它与混沌时间序列都表现出外在的随机性。利用动力系统方法研究了几种跳频码(m序列,RS序列,非线性序列及混沌序列)的动力学特性(相关维数与最大李雅普诺夫指数),结果表明跳频码表现出混沌的特征。5. 基于混沌理论提出的混沌预测方法对实际跳频码进行了预测研究；跳频码是由确定的机制产生的伪随机序列，表现出混沌的特征，因而可用混沌预测方法对跳频码进行预测，研究表明神经网络和非线性自适应滤波预测都能对跳频码进行有效的预测，但非线性自适应滤波预测预测速度快，且准确率高。6. 混沌理论在通信信号调制识别方面的应用；利用混沌理论可对混沌时间序列的复杂度描述和刻划，提出了提取通信信号调制类型的特征参数（分形特征和Lempel-Ziv复杂度特征），这两种特征对于通信信号信噪比在较大范围的变化，体现出良好的类内聚集程度和类间分离程度，因而提高了分类识别率，同时也降低了分类器的复杂度，关于分形特征参数，从理论上说明了该特征参数对噪声不敏感，另外，利用模糊测度和模糊积分理论，设计了组合分类器，该组合分类器利用了单个分类器不同的重要测度，采用模糊积分作为组合规则，显著地提高了通信信号调制类型整体识别率，克<WP=7>服了单个分类器对某种调制类型的局限性。更多还原

【Abstract】 A wide variety of natural phenomena exhibit complicated, unpredictable and seemingly random behavior. Chaotic dynamics appears to provide a relatively simple and possibly more satisfactory explanation to a lot of complicated phenomena among them. Chaos theory shows that simple deterministic systems with only a few variables can generate seemingly random behavior.The purpose of chaotic theory is discovering possible simple rules, which hide in the seeming random phenomena and obtaining common rules which a big group of complicated problems follow. With chaotic dynamics developing, predicted complexity of time series have been profoundly comprehended. Even if it is impossible that long-term approximate prediction for chaotic behaviors, but it is possible that accurate short-term prediction for it. So studying prediction of chaotic time series probably solve a lot of complicated problems. In addition, chaotic theory provided different characteristic parameters describe chaotic time series and communication signal is also a sort of nonlinear time series, so difference of modulation types of communication signal is embodied by difference of time series. This dissertation studys prediction of chaotic time series and chaotic application in the aspect of recognition of communication signals. Main contents studied include: (1) nonlinear adaptive filter predicting chaotic time series and hopping frequency code;(2) common used neural networks (Backpropagation neural networks and radial basis function neural networks) modeling and predicting chaotic time series and hopping frequency code;(3) a fast method of predicting chaotic time series;(4)feature extraction of communication signals based chaotic theory;(5) classifier design based fuzzy integral.Several value and important results which bring forth new ideas are<WP=9>achieved and listed as following: 1. Studying convergence of two-rank volterra adaptive filtersAs predicting performance of volterra adaptive filters mainly depending on adaptive algorithm and convergence of adaptive filters is important criterion evaluating adaptive algorithm good or bad, we have studied convergent relationship between a control parameter and predicting error, between coefficient of initial filters and predicting error, among coefficient of filters, between predicting error and coefficient of filters and between ranks of filters and coefficient of filters.2. Determined ranks of nonlinear adaptive filters.As ranks of nonlinear adaptive filters based nonlinear adaptive technology is determined arbitrarily by people, ranks of filters not only influence predicting accurateness of time series, but also involve computing complexity of realizing this kind of filters. Via theory analyzing and experiments, optimal input dimensions of nonlinear adaptive filters is obtained, at the same time it is presented a simple algorithm which determine optimal input dimensions of nonlinear adaptive filters.3. In order to meet engineering demands of real time, a fast method predicting chaotic time series is proposed.As hopfield neural networks have many advantages such as fast convergence, software or hardware realizing hopfield neural networks, it is presented that fast predicting method based hopfield neural networks, which combines volterra predicting filters and hopfield neural networks. This method provides a solution to real time engineering problems.4. Analysis of dynamic characteristics of hopping frequency codes Hopping frequency codes are pseudo-random sequences generated<WP=10>by determinate rules. It is seemingly random the same as chaotic sequences. It is studied that dynamic characteristics(Correlation dimension and largest Lyapunov exponents )of some kind of FH codes (m sequences, RS sequences, Nonlinear sequences and chaotic sequences) by dynamic system method，the results show that FH codes are of chaotic characteristics5. Predicting performance has been studied for hopping frequency codes by a kind of method predicting chaotic time series via chaotic theory更多还原