面向业务的智能软件定义蜂窝网研究

As cellular networks are becoming the infrastructure provider in the Information Communications Technology (ICT) industry, it also faces more serious conflict between various service demand and simple provisioning mechanism as well as complex network entities. Therefore, it is imperative to revolutionize current cellular networks by introducing full functionalities (e.g., sensing, mining, prediction and reasoning) of artificial intelligence and designing intelligent, efficient and sustainable schemes. This project first leverages the flexibility and open interfaces of software defined networks (SDN) and fully consider the difference between access networks and core networks, so as to provide a practical software-defined cellular network architecture and lay the foundation to adopt the artificial intelligence. Afterwards, based on the practical records from operating cellular networks in China, this projects studies the intelligent means to properly model different types of services and classify the embedded distinctions, so as to timely find one new and unknown type of service and determine the required network provisioning sources. Furthermore, based on the aforementioned results, this project tries to answer how to conduct an effective prediction method in terms of both the service prediction accuracy and the induced computational resources, as well as how to design an efficient and scalable service-oriented network resource management schemes.

随着蜂窝网逐渐成为信息通信领域的基础设施提供者，业务需求多样化同网络架构复杂化、承载能力单一化矛盾日益加深，迫切需要引入机器学习算法，分析过载的业务数据，在业务数据和资源管理方面上实现面向业务进行感知、挖掘、预测、决策的智能化，实现智能、高效、可持续的业务承载能力。本项目首先挖掘软件定义网络灵活、开放的特点，考虑蜂窝网接入网和核心网的差异，研究软件定义网络与蜂窝网的融合方式，为智能蜂窝网打下基础；接下来，研究如何更好地依托蜂窝网实测数据，对业务进行统计建模，凝练业务间模型和需求的差异性，探索基于深度学习的分类方法，从而快速地对层出不穷的未知新业务完成分类、确定需要的网络资源；最后，基于业务的分析结果，研究平衡准确性和资源开销的业务预测方法，建立于业务需求相匹配的、扩展性强、基于学习的网络资源管理方法。

随着智能无线设备的不断增多，移动蜂窝网流量加速增长的同时，所承载的内容从单一的语音变得丰富多样。为了服务日益增多的移动用户和层出不穷的无线应用，无线网络架构正在变得日益复杂，这给运营维护带来前所未有的挑战，迫切需要面向业务的蜂窝网智能化。本项目首先挖掘软件定义网络灵活、开放的特点，研究和提出了智能软件定义蜂窝网的基本架构，并以蜂窝网的安全为例构建了群能软件定义网络与安全的仿真平台。接下来，依托蜂窝网实测数据，对业务和基站站址密度进行统计建模，基于此提出了网络资源覆盖概率与时延性能分析方法，这些结果有助于更准确地评估关键技术的网络服务性能。基于业务的分析结果，构建了数据模型多驱动的预测方法，可以更好平衡预测准确性和资源开销。最后，提出了业务驱动的网络切片资源调配方法，从而快速及时准确地根据业务需求分配所需的网络资源。本项研究成果可以提高蜂窝网业务承载能力，为实现智能、高效、可持续的 5G 蜂窝技术演进提供有益尝试。在本项目的资助下，共发表和录用各类期刊、国际会议论文32篇，授权发明专利3项；培养本科毕业设计3人，研究生11名。项目申请人也成功晋升副教授，并获得了面上项目的进一步支持。