【作者】 吴文清；

【导师】 刘嘉焜；

【作者基本信息】 天津大学 ， 应用数学， 2004， 硕士

【摘要】 网络流量特征分析和建模是网络技术研究的一个分支. 由于网络拥赛控制（包括丢包率、队列延时、网络吞吐率）和网络资源利用（如队列缓冲区容量、带宽利用率及 QOS 保证）等关键技术都依赖于特定的流量特性，因而网络流量的特性分析和建模对网络结构设计和性能优化具有重要的理论和实际意义. 在计算机网络模型的设计评价和优化中，网络业务模型起着非常重要的作用.其中，时间序列模型作为研究网络业务的工具，有着很好的应用前景. 然而，传统时间序列模型只能处理短相关，如泊松过程，Markov 过程，AR, MA，ARMA和 ARIMA 过程等. 随着网络测量技术的发展，网络研究人员发现高速网络的业务具有长相关性(long-range dependence), 亦称自相似性(self-similarity). 而这些传统模型是不能处理的. 于是，一些长相关模型，如 FGN（分数高斯噪声，Fractional Gaussian Noise）和 FARIMA(0,d,0)被用作网络业务模型. 新发展起来的FARIMA(p,d,q)模型克服上述长相关模型的缺点，可同时处理长相关过程和短相关过程. 但其预报是建立在概率基础上的，预报的步长也受到网络流量特性的制约，从而限制了它们运用于实际控制. 目前，用混沌动力学处理时间序列问题正在蓬勃发展，在许多领域得到或开始得到发展. 同时，网络流量的自相似性已被证实，自相似性和混沌具有紧密联系，这为我们研究网络业务模型开辟了新的途径. 我们注意到，用混沌方法研究网络流量的文献较少，仅有的研究仍停留在网络流量的混沌特性上. 这启发我们用混沌理论对网络业务进行系统研究，探讨网络流量混沌的成因. 神经网络具有并行处理及强大的非线性映射能力，对于未知的动力系统，可以通过它来学习混沌时间序列，然后进行预测和控制. 本文应用混沌理论对网络流量的系统动力学特征进行了认真分析，并结合网络流量的具体数据，计算了网络流量的 Hurst 参数，关联维数，Lyapunov 指数. 在此基础上，运用小波理论对网络流量数据进行了去长相关处理，进一步对网络流量混沌成因进行了探讨，指出网络流量的混沌与网络流量的长相关存在着联系.在此基础上，运用 BP 神经网络理论，建立了相关模型，对实际网络流量进行了预测. 研究表明，基于混沌理论建立的 BP 神经网络模型和 FARIMA(p,d,q)模型都能较准确地对网络流量进行预测，而 BP 神经网络模型能够经过学习，获得较长的预测步长.更多还原

【Abstract】 The analysis and modeling of the network traffic is one of the branches of thefield of traffic technology. Because network congestion detection and control (such asthe efficiency of bandwidth, providing QoS) is connected with the character ofnetwork traffic and because the utilization of network resource is dependent on it also,it has important theoretical and practical meanings on the study of network traffic todesign and optimize the network framework. In the design and optimization of network framework, the model of networktraffic is important. Especially, as a tool to analyze network traffic, time series has abright foreground of application. But traditional time series only deal withshort-dependence process, such as Poisson process, Markov process,AR(Auto-Regressive), MA(Moving Average) ， ARMA(Autoregressive MovingAverage) and ARIMA(Autoregressive Integrated Moving Average). With thedevelopment of network measuring, researchers have found that there islong-dependence in the high-speed network, which is also called self-similarity. Themodels above are not in point, so other long-dependence models such as FGN(Fractional Gaussian Noise) and FARIMA (0,d, 0) are used in network traffic. Therecent model of FARIMA (p, d,q) gets over the scarcity of above models, which candeal with long-dependence process and short-dependence process at the same time.While its forecast is based on the probability and its forecast length is confined to thecharacter of network traffic, this restricts it to put into practice. At the present, themethod about the question of time series by system dynamics is developing, which isused in many fields. On the other hand, the self-similarity of network traffic is reallycertificated and self-similarity is connected with chaos closely, so we can use thetheory about chaos to study the model of network traffic. We find that there are fewarticles using chaos theory to study network traffic, and these articles are attentive tothe character of network traffic. All of these suggest us to use chaos theory to studynetwork traffic and to argue about the cause about the chaos of network traffic. The paper analyses the kinetic characteristics of Internet traffic by using somecomplexity theories, such as fractal and chaos theories. We obtain the fractal<WP=5>characteristics, the extent of complexity, and the style of movement of the system.Based on the above, we discussed relation between the chaos of Internet traffic andlong-range dependence by filtering the long-range dependence using wavelet analysis.The conclusion as follow: there is fractal structure in the Internet traffic and the chaosof Internet traffic is connected with long-range dependence. With phase spacereconstruction, the paper demonstrates the Internet traffic chaos phenomena lies inInternet traffic, and computes some parameters such as correlative dimension,Lyapunov exponent. Based on this, the paper constructs the BP neutral network modelto forecast the Internet traffic. Comparing with FARIMA (p, d, q) model, the BPneutral network model has the same ability of forecast and has longer space offorecast.更多还原