【作者】 王燚；

【导师】 郭伟；

【作者基本信息】 电子科技大学 ， 通信与信息系统， 2008， 博士

【摘要】 无线通信系统中利用了许多统计学相关技术完成对一些具体问题的最优化求解,如信道估计,信号检测,调制识别,但是随着系统本身的日益复杂,人们逐步需要从繁琐的数据中寻求简便智能的方法给出对这些问题的解决方案。统计学习理论试图帮助人们从繁多的数据和现象中揭示事物规律本质。与传统统计学相比,统计学习理论(Statistical Learning Theory,SLT)是一种专门研究小样本情况下机器学习规律的理论。该理论针对小样本统计问题建立了一套新的学习理论体系,在这种体系下的统计推理规则不仅考虑了对渐近性能的要求,而且追求在现有有限信息的条件下得到最优结果。论文主要围绕统计学习理论在通信系统中的一些应用问题进行了研究和探讨,其主要贡献在于:1、针对混沌时间序列的特性,根据混沌动力系统的相空间重构理论,利用SVM可以自动把输入向量映射到一个高维特征空间中实现数据的线性及非线性划分的功能,将具有统计学习功能的最小二乘支持向量机用于构建预测模型,并用该模型对混沌跳频序列进行预测,通过计算机仿真验证了该模型的正确。2、针对混沌时间序列的动态系统特性及普通神经网络在预测时存在的局限,论文提出一种将具有混沌动态特性的神经节用于构造混沌神经网络进行预测的方法,并以此对一般混沌动态系统建立预测模型,最后对Mackey-Glass混沌序列和Logistic-Kent映射的混沌跳频序列进行预测研究,并通过仿真验证了该模型的正确。3、不同于传统的经验风险最小化准则下的信道估计方法,论文研究了在结构风险最小化准则下,基于支持向量机的信道估计技术。由于MIMO及非线性通信信道的复杂性使得信道估计精度和速度相比SISO信道估计大大降低,在样本数量有限的情况下,这一问题就更加突出。现有信道估计方法往往局限于SISO信道的非线性估计,或者是非时变线性MIMO信道估计,而对于非线性时变信道估计则通常是将时变非平稳信道估计看作是在一段时间内的平稳信道估计进行处理,但在收敛精度和速度上不能达到满意的结果。针对以上不足,论文利用最小二乘支持向量技术将MIMO信道估计问题转变为求解多维信道函数的回归问题,利用支持向量技术的动态多维拟合方法对MIMO系统进行自适应非线性信道估计,显著提高了收敛精度和预测速度。4、在无先验样本可利用的情况下,仅依靠有限的信号先验统计特性,针对已有的盲信道估计和均衡算法在收敛速度上的不足,利用最小二乘支持向量技术改进传统的恒模盲均衡算法,并采用迭代权值方法进一步减小计算量,加快收敛速度,计算机仿真验证了该方法的高效性。5、根据跳频通信网中跳频序列族满足最大汉明相关距离最小化的正交特性,论文提出适用于非平稳条件下的多个子网混和跳频信号的盲分离模型,并将一种改进的短时相关算法和改进的最小互信息量算法应用于混和跳频信号的在线盲分离,验证了该算法的有效性。更多还原

【Abstract】 In mobile communication system plenty of technics of statistic theory was used to solve some practical problems, such as channel estimation, signal detection, modulation identification. The whole system is becoming more and more complicated, however. Simpler and smarter solutions are required from the numerous datus. Statistical Learning Theory (SLT) has become more and more popular to people since it was founded by Dr. Vapnik. This framework tries to help us explore the nature and rule among the various data and phenomenon. Compared to the trational statistics, SLT focuses on the research of machine learning based on the small number of samples. It constructs a new framework about learning. In this framework, the rule of statistical reasoning will consider not only the demand of convergence, but also the best optimal result under the circumvent that the limited information can be used. This dissertation focuses on research of the SLT’s application in some problems of communication system.The main contributions of this dissertation are achieved and listed as following:1、Having noticed that the chaotic characters lies in the FH-code, we build a LS-SVM model to predict the FH-code sequence. Because of the chaotic character of the FH-code, we utilized Taken’s embedding theory and SVM technology to construct a mapping between the one dimension input vector and a high dimension feather space, to realize the data linear or nonlinear classification. Then LS method is applied to train the SVM model and the function of determined by the FH-code’s attractor is regressed. Its performance is analyzed and some illustrative simulations are presented.2、Considering the static BP neural network’s insufficiency when predicting the chaotic dynamics, we propose a new neuro network including the chaotic neurons and apply it to regress the chaotic dynamic system. We applied the neuro network to predict the Mackey-Glass chaotic series and the L-K FH-code sequence and evaluate the performance of algorithm through the simulations.3、Once little prior samples are available, we utilize the modern statistic learning theory to solve the channel estimation problem. Due to the complication of MIMO nonlinear channel, the performance of channel estimation decreases on accuracy and speed with the nonlinear and nonstationary condition. In the circumstance of little amount of samples to use, this problem is more obvious. The current nonlinear channel estimation method lies in SISO channel or stantionary MIMO channel estimation. Although some methods treat the nonlinear time-variant channel as stationary in a sufficient short time period, the performance in convergenc and accuracy is still unsatisfying. We utilize the LS-SVM to convert the MIMO channel estimation problem to a multi-dimension channel function regression problem, and apply the AM-SVR framework to MIMO channel adaptive estimation. The simulation results demonstrate this method is efficient.4、Without the prior training samples in the non-cooperation condition, an approach for solve the channel estimation and equalization has been presented. This algorithm is based on the finite prior statistical character of received signal, and overcomes the shortcoming in slow convergence which exists in blind channel estimation. A relative weighted method is applied to train this SVM and reduces the amount of calculation in progress. The performance of blind channel estimation is evaluated by MATLAB simulation.5、An improved and more complex Blind Signal Seperation (BSS) algorithm for separating linear convolved mixtures of nonstationary signals in FHSS system is presented. This algorithm relies on the nonstationary nature of the sources to achieve separation, which assumes statistically stationary sources as well as instantaneous mixtures of signals. In practice, the FH/CDMA sources received are nonstationary and linear convolute mixing. A more complex BSS algorithm is required to achieve better source separation. The algorithm is based on minimizing the average squared cross output channel correlation. The mixture coefficients are totally unknown, while some knowledge about temporal model is available. At the same time, an improved MMI algorithm was applied to the BSS of the multiple FHSS signals. The simulation results show the effectiveness of the method in the blind detection of the multiple FHSS signals.更多还原