基于数据挖掘的IPTV故障诊断研究与实现

发布时间：2018-09-12 09:48

【摘要】：随着多媒体通信技术的迅猛发展,越来越多的用户能够在家中享受IPTV(Internet Protocol Television)服务。为了保证极佳的用户体验,IPTV运营商们尽力支撑高质量视频节目并确保视频传输通畅。与此同时,运维部门的压力也越来越大,通常每日都要解决不同程度、不同层面的困难与故障。因此,如何客观准确地评估IPTV服务质量并及时维护IPTV网络正常运行已成为当下的研究热点。本论文提出了全新的IPTV运维质量评估方案,该方案是一套能高效精准定位并诊断IPTV网络故障的综合系统。该IPTV业务服务质量评估系统在构建过程中,特别融合潜在用户报障与潜在设备预警这两方面。在采用故障设备预测模型来解决潜在设备预警过程中,不仅使用传统服务质量QoS(Quality of Service)指标,也充分考虑并采用能反映用户观看行为的客观用户体验质量QoE(Quality of Experience)指标。此外,本文针对IPTV故障定位诊断还进行以下工作:一方面,本文提出了一个综合故障设备预测模型,该模型首先挖掘机顶盒指标参数与网络状况之间的关联,使IPTV网络故障诊断更加高效准确,特别是针对卡顿花屏故障现象。该故障设备预测模型包括质差指标、记录、用户三方面。首先,质差指标模型作为替代传统决策树生成中的特征提取部分,负责从候选机顶盒指标中筛选出有效有限指标。其次,质差记录模型是基于改进的决策树算法,通过不断调整阈值以达到高准确率以及低误判率。最后的质差用户模型将反映用户观看体验质量的观看时长考虑在内。实验结果表明,质差用户模型的预测准确率能高达83.25%。另一方面,针对传统算法在非均衡IPTV数据集下用户报障预测效果不理想的问题,本文将影响网络服务质量的传统网络参数QoS和主观反映用户体验质量QoE的MOS(Mean Opinion Score)评分结合来预测用户是否报障。本文在已有的ODR-BSMOTE-SVM算法基础上,针对过采样算法产生噪声以及核参数没有进行优化的缺陷,提出了一种改进型算法。该改进算法首先采用欠采样和过采样算法及数据清洗算法对原始非均衡数据进行处理,然后通过自适应变核参数寻找近似最优值,最终实现提升分类效果。实验结果表明,较传统标准支持向量机(SVM)算法和ODR-BSMOTE-SVM算法,本文所提出的算法能获得更佳的预测效果。
[Abstract]:With the rapid development of multimedia communication technology, more and more users can enjoy IPTV (Internet Protocol Television) services at home. IPTV operators try to support high-quality video programs and ensure that video is transmitted smoothly in order to ensure an excellent user experience. At the same time, the pressure of operations and maintenance departments is increasing, usually to solve different levels of difficulties and failures on a daily basis. Therefore, how to evaluate IPTV QoS objectively and accurately and maintain the normal operation of IPTV network in time has become a hot research topic. In this paper, a new IPTV operation and maintenance quality evaluation scheme is proposed, which is a comprehensive system which can locate and diagnose IPTV network faults efficiently and accurately. In the process of constructing the IPTV service quality assessment system, the potential user reporting barrier and potential equipment warning are especially integrated. In the process of using fault equipment prediction model to solve the potential equipment warning process, not only the traditional quality of service (QoS (Quality of Service) index is used, but also the objective user experience quality (QoE (Quality of Experience) index, which can reflect the user's viewing behavior, is fully considered and adopted. In addition, this paper also does the following work for IPTV fault location diagnosis: on the one hand, this paper proposes a comprehensive fault equipment prediction model, which first mining the relationship between set-top box index parameters and network condition. Make the IPTV network fault diagnosis more efficient and accurate, especially for the phenomenon of Carton flower screen fault. The fault equipment prediction model includes three aspects: quality index, record and user. First, the quality difference index model, as the feature extraction part of the traditional decision tree generation, is responsible for screening out the effective limited index from the candidate set-top box index. Secondly, the quality difference record model is based on the improved decision tree algorithm, by constantly adjusting the threshold to achieve high accuracy and low error rate. The final quality difference user model takes into account the length of viewing time that reflects the quality of the user's viewing experience. The experimental results show that the prediction accuracy of the quality difference user model can reach 83.25%. On the other hand, aiming at the problem that the performance of the traditional algorithm is not satisfactory under the unbalanced IPTV data set, In this paper, the traditional network parameter QoS which affects the quality of service of the network is combined with the MOS (Mean Opinion Score) score, which reflects the quality of user experience. In this paper, based on the existing ODR-BSMOTE-SVM algorithm, an improved algorithm is proposed to overcome the defects of over-sampling algorithm which produces noise and the kernel parameters are not optimized. The improved algorithm firstly processes the original unbalanced data by using the under-sampling and over-sampling algorithm and the data cleaning algorithm, and then finds the approximate optimal value through adaptive variable kernel parameters, and finally realizes the improvement of classification effect. The experimental results show that the proposed algorithm can obtain better prediction results than the traditional standard support vector machine (SVM) algorithm and the ODR-BSMOTE-SVM algorithm.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN949.292

【参考文献】