基于多维关联的移动网络状态分析研究

发布时间：2018-05-19 18:09

本文选题：多维关联 + 机器学习　；参考：《中国科学技术大学》2017年博士论文

【摘要】：近几年来,移动用户的急剧增加,多样化的终端接入,数据业务的爆炸式增长,变革了移动网络的架构、技术等各方各面的发展,使得网络环境变得多样化、复杂化,运营商对于网络维护管理的复杂度,难度也随之增加。因此,为了降低移动网络运维管理的人工成本,提高网络的操作和维护性能,智能化网络运维成为未来移动网络运维的发展的必然趋势。移动网络状态分析作为智能化运维中的重要环节,包含发现问题、问题定界等方面的研究,具有极大的应用价值。因此,本文围绕移动网络状态分析展开,以多维关联分析作为理论基础,重点解决网络异常状态识别以及流量状态预测两大关键问题,具体研究内容包括以下几个方面:第一,本文针对移动网络KPI异常状态检测问题,引入基于密度聚类的异常检测方法,提出一种基于核密度的局部密度多维异常检测方法,从无监督学习的角度实现网络自动化异常检测功能。通过相邻点密度对比的方法判别异常,解决传统单维检测以及统计模型不稳定的问题;通过核函数赋予不同方向数据点不同的权重,相比于直接使用基于密度的算法,克服了移动网络KPI数据异常单向性、高性能点稀疏性、不同KPI变化范围差异性等数据特点,使得最终的异常检测精度有所提升。第二,本文针对移动网络KPI异常根因分析问题,设计了一种基于目标的分层聚类异常根因分析框架,实现训练数据未进行异常根因标定且KPI种类繁多的情况下,自动化进行KPI异常根因分析功能。该框架首先针对网络异常类别划分问题,设计了一种基于SOM和K-mediods混合的非监督学习的方法,通过参数分布特性,将异常划分为不同的类别,建立具有迭代特性的KPI异常类别划分系统,标定的异常类别用于下一步的异常根因相关维度选择;其次,针对异常根因分析问题,提出一种基于KS检验以及互信息熵的特征选择方法,结合KPI物理属性,利用KPI正常类别与异常类别之间参数分布的差异性,选择与异常类别相关的KPI进行聚类,得到不同类型的异常根因。异常类别及其对应的根因最终形成专家知识库,可以用于在线异常根因分析。实验证明,该框架结构能够有效解决异常模式样本稀疏且异常根因维度稀疏问题。第三,本文针对移动网络流量状态预测问题,引入多维度离散流量状态预测方法,以多源数据融合为基础,提出一种基于局部关联分析的双层数据驱动预测框架,相比于传统基于流量时间序列的预测方法,提升了时空尺度较小情况下,突发流量的预测的精度。在此过程中,首先针对如何确定流量相关维度问题,设计了一种基于多个独立二值分类器的分析方法,通过建立每个流量状态下的二值分类模型,得到不同流量状态下与之相关的维度;基于上述获取的与流量状态局部关联的维度,针对如何确定多维度与流量之间预测函数问题,基于迁移学习的概念,设计了一种集成式预测框架,分别对每个状态独立建模,预测时以自身状态模型为主,其它状态模型的结果辅助预测,最终将多个预测其的结果采用投票的方式集成,具体解决部分流量状态稀疏,无法建立准确模型的问题;不同时段内流量模型存在差异性,因此,针对如何确定建模时段问题,设计了一种基于层次聚类的相邻时间合并方法,将具有相似特性的相邻时段合并,对时段建模,具体解决用户行为"时间漂移"问题。实验证明,该预测框架在稀疏高流量状态上预测性能有明显提升,且整体性能较优。本文以现网采集的移动网络数据为基础,以分析移动网络数据特点为切入点,探讨以数据多维关联分析为理论的无监督学习、监督学习、迁移学习等方法在移动网络状态分析中的应用方式及其性能结果,从而为更为深入的智能化移动网络运维提供研究思路。
[Abstract]:In recent years, the rapid increase of mobile users, the diversified terminal access, the explosive growth of data services, the transformation of the architecture of the mobile network, and the development of all aspects of technology, make the network environment diversified and complicated, and the complexity of the network maintenance and management is increasing. In order to reduce the mobile network, to reduce the mobile network, The artificial cost of the network operation and maintenance management improves the operation and maintenance of the network. The intelligent network operation and maintenance has become the inevitable trend of the development of the mobile network operation and maintenance in the future. As an important link in the intelligent operation and maintenance, the state analysis of the mobile network includes the Research of the discovery problem and the problem delimiting. Therefore, this paper is of great application value. Around the state analysis of mobile network, taking multi-dimensional association analysis as the theoretical basis, it focuses on two key problems of network abnormal state recognition and traffic state prediction. The specific research contents include the following aspects: first, this paper introduces the anomaly detection based on density clustering based on density clustering to detect the abnormal state detection of KPI in mobile network. Method, a local density multidimensional anomaly detection method based on kernel density is proposed, and the network automation anomaly detection function is realized from the angle of unsupervised learning. The anomaly is discriminate by the density contrast method of adjacent points to solve the problem of the traditional single dimension detection and the unstable statistical model. The kernel function gives different direction data points to different points. The weight, compared to the direct use of density based algorithm, overcame the data characteristics of KPI data anomaly, high performance point sparsity, different KPI variation range and so on, which made the final anomaly detection precision improved. Second, this paper designs a kind of target based on the target of mobile network KPI anomaly root cause analysis. The hierarchical clustering anomaly root causes the analysis function of KPI abnormal root cause analysis in the case that the training data is not calibrated and KPI has a wide variety. The framework first designs a non supervised learning method based on the mixture of SOM and K-mediods, which is based on the network anomaly classification problem, through the parameter distribution. Characteristics, divide the exception into different categories, establish the KPI anomaly classification system with iterative characteristics, the calibrated exception category is used for the next step of the exception root to be selected because of the correlation dimension. Secondly, a feature selection method based on KS test and mutual trust entropy is proposed for the analysis of abnormal root cause, which combines the physical properties of KPI and uses K. The difference between the parameter distribution between the normal class and the exception category of PI, select the KPI which is related to the exception category, and get the different types of abnormal root cause. The exception category and its corresponding root cause the expert knowledge base, which can be used for the online abnormal root cause analysis. The experimental evidence shows that the frame structure can effectively solve the abnormal pattern sample. The sparse and abnormal root cause dimension sparsity problem. Third. In this paper, a multi dimension discrete flow state prediction method is introduced to the traffic state prediction problem of mobile network. Based on multi source data fusion, a two-layer data driven prediction framework based on local correlation analysis is proposed, compared with the traditional predictor based on the flow time series. In this process, a new analysis method based on multiple independent two value classifiers is designed, which is based on the establishment of a two value classification model in each flow state, which is related to different traffic states. Dimension; based on the dimensions associated with the local traffic state obtained above, based on the concept of how to determine the predictive function between multi dimension and traffic, based on the concept of migration learning, an integrated prediction framework is designed to model each state independently, the prediction is based on the self state model, and the results of other state models are assisted. In the end, a number of prediction results are integrated in the way of voting, which can solve the problem of sparse partial flow state and no accurate model. The flow model is different in different periods. Therefore, a method of merging adjacent time based on hierarchical clustering is designed to determine the time of modeling time. It is proved that the prediction performance of the prediction framework is obviously improved and the overall performance is better in the sparse and high flow state. This paper is based on the mobile network data collected in the present network, and discusses the characteristics of the mobile network data as the breakthrough point. The method and performance results of unsupervised learning, supervised learning, migration learning and other methods in the analysis of mobile network state are taken as the theory of data multidimensional association analysis, thus providing a research idea for more in-depth intelligent mobile network operation and maintenance.
【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TN929.5

【相似文献】