VBR Video Traffic Modeling And Prediction Based On Neural Network

With the development of visual services such as VOD,video conferencing, multimedia especially video has gradually become the mainstream of the communication system. Differing from the traditional constant bit rate (CBR) data services, the compressed real-time encoded video flow is variable bit rate (VBR). In order to adapt to the changes in this type of service, people demand high-speed network especially the next generation wireless networks to be able to well support the multimedia service whose mainstay is video service. Among them, the mathematical description of the video data flow as the source of the data becomes a prerequisite to a variety of network computing and strategy design. By modeling and predicting of the VBR video traffic, we can design effective congestion control mechanisms,resource allocation and scheduling strategies to reduce or avoid congestion, improving efficiency of the use of network resources while ensuring available bandwidth efficient and fair.VBR video signal has uncertain, sudden and nonlinear characteristics and takes a lot of bandwidth resources in the network transmission. To achieve the accurate forecast of VBR video traffic is an important means to improve the information transmission speed and the utilization of network bandwidth resources .Linear prediction method is one of the ways. For its low computational complexity, it's fit for online learning, but with low forecast precise. In recent years, the improvement of VBR video traffic online forecast are mainly: recursive least squares method and time-delay neural network method has been applied to JPEG and MPEG video flow forecast, multi-multiplexing of MPEG video traffic expansion recursive least squares method, linear prediction method based on the minimum mean squares, as well as the adaptive wavelet prediction methods for fast convergence. The artificial neural network (ANN) with nonlinear capturing sequence, non-stationary and unexpected performance capabilities, has unparalleled advantages in handling the issue which requires high real-time and can not be accurately described.Compared with the traditional methods based on mathematical prediction method, it can adapt to a more complex video network transmission requirements, so intelligent video traffic prediction has become a hot point of the current study.As a new forecasting method, the neural networks, with the characteristics of good nonlinear mapping capabilities, flexible and effective way of learning, shows great strength and potential in the application in the forecasting area. Artificial neural network is a kind of model imitating brain information processing. Because of its characteristics that do not rely on the issue of the high degree of parallel processing, and do not rely on prior knowledge or rules as he prerequisite for adaptive learning ability, in solving intelligent information problem with high real-time requirement and no exact mathematical model to describe, the neural network method has incomparable advantages with other methods. BP neural network is by far the most widely used back-propagation network, which is widely used in function approximation, pattern recognition / classification, data compression, and so on. Many researchers are using BP neural network to predict the video stream flow or network traffic flow, and it is applied to dynamic bandwidth allocation scheme, which supports a good quality of service.Based on artificial intelligence methods, this paper, on the basis of the analysis of the MPEG-4 video traffic characteristics, use a method by combining the BP neural network and discrete wavelet transform ,to establish a new VBR video traffic forecasting model.The model uses the multiresolution analysis method in the wavelet transform which can decomposit the original signal to the corresponding space by scale. Use two filters constructed with Mallat algorithm to decomposit the samples to high-frequency and low-frequency component, low-frequency component here for the scale factor, high-frequency component for the wavelet coefficients. The model introduces a new variable called scaling factor, which is defined as the ratio of wavelet coefficients and scale coefficients. The neural network input vector is composed of scale factors. This paper put forward two methods to deal with the largest-scale low-frequency component, by processing to get a group of new datas ,which , with the new scale factor trained by the neural network, are used to reconstruct wavelet from the largest scale to the smallest scale, then get the data which is the final forecast value. The above is the overall thinking of establishing.flow forecasting model.This paper first outlines the mathematical theory which is applied in the process of establishing the model and a brief overview of the wavelet analysis theory, fractal theory, MPEG coding theory and neural network theory. Secondly, it introduces the main characteristics of MPEG-4 video stream, gives definition and analytical methods of variable characteristics, and analysis of the experimental data. Then, list some existing video traffic forecasting model, and outlines several typical modeling method. Finally, it detailedly introduces the modeling idea of two method of how to design, model ,analysis and get the analysis results.This paper uses the datas trained by BP neural network through a large number of simulation experiments to determine the network structure..By seting up the forecast model with Matlab simulating enviroment,we get two methods which achieve forecasting function in some degree ,but the forecasting effect of the first one is better than the second. At the same time, by selecting video of different image quality as the data flow analysis object, we get the conclusion that the model has s good performance for low quality VBR video data flow, and the model shows better ability to forecast video flow with lower sudden property. At the same time, by comparing the traffic performance indicators in different scales, we can conclude that the model has multi-step forecasting ability of low quality video.For high quality video, a single-step prediction is suitable.