5G无线通信系统中基于机器学习的超密集网络技术研究

In future 5G communication system, demands of high capacity and high spectrum efficiency (SE) are effectively met by ultra-dense networks (UDN), where a lot of small cell base stations (SBSs) are densely deployed and share the spectrum resources with cellular networks. UDN has been proposed as one of the key technologies in 5G system, however, it still faces the bottlenecks such as limited spectrum resources, complicated interference, etc. Using machine learning (ML) of artificial intelligence (AI) area becomes one effective technical solution. Therefore, our works aim to establish the theoretical research framework of 5G UDN based on the ML tools of deep neural network, support vector machine and reinforcement learning. First, the user mobility pattern prediction and quality of experience are analyzed, where a reasonable recurrent neural network structure is designed for parameter training and prediction, thus, the SBS deployment strategy can be optimized in UDN. Second, by mining user historical information to form a big database, a method of user communication mode selection based on ML can be proposed to realize a low latency and low overhead mode selection in UDN. Last, consider different non-ideal conditions, the resource allocation can be optimized by ML to further enhance the UDN performance. Our research works are expected to provide solid technical reserves for future AI application in wireless communication field.

面对未来5G系统高容量高谱效的技术挑战，超密集网络（UDN）技术通过在网络中超密集部署大量小基站来降低传输损耗，大幅提高网络容量和谱效，成为5G的关键技术之一。然而，UDN仍面临资源受限、干扰复杂等瓶颈问题。采用人工智能领域中的机器学习理论，将成为解决这一瓶颈的有效技术方案。为此，本项目以机器学习中的深度神经网络、支持向量机以及增强学习作为基本理论工具，构建UDN技术完善的理论研究框架。首先，设计合理的循环神经网络进行样本训练和用户移动图案预测，推导UDN的最优用户体验质量，从而对UDN小基站部署策略进行优化；其次，挖掘用户历史数据形成大数据集，提出基于机器学习的用户通信模式分类，实现低时延的用户通信模式选择机制；最后，综合考虑各种非理想条件，利用机器学习对频谱资源分配进行优化，进一步提高UDN的性能。本项目可望为我国引领未来人工智能在无线通信领域的应用提供重要的理论基础和关键技术支撑。

本项目利用机器学习对5G超密集网络（Ultra-dense Networks，UDN）中的部署、通信方式以及频谱资源分配等问题进行深入研究。首先对5G UDN中的终端部署问题进行了分析，利用机器学习中的优化算法获得了合理的终端部署密度，为实际网络部署提供了理论指导；接着，对5G UDN中的通信方式进行了研究，利用机器学习SVM分类器对发射机码字进行训练，高效地完成了通信码字选择，大幅降低了通信系统开销；最后，对5G UDN中的频谱资源环境进行分析，通过强化学习中的双深度Q网络设计了频谱接入方式，使得用户合理进行频谱接入，同时避开可能造成的干扰。相关成果适用于海量用户存在的UDN网络，并能够有效提升网络的系统承载。共发表论文12篇，获授权专利2件。