未来5G中基于机器学习的动态频谱高效感知协同控制机理研究

Spectrum sharing is one of the international research fronts in the field of wireless communications, which has gained increasing interest of academia, industry, standard associations, and governmental spectrum regulators, and has been recognized as one of the most direct and efficient methods to resolve the issue of ever-increasing serious spectrum deficit risk in future wideband wireless mobile communication systems. However, due to the potential new features of future 5G networks, including hyper-densification, high heterogeneity, high dynamics, and intelligence, there are new problems and new challenges faced by the discovery process of spectrum sharing, such as the huge gathering cost of spectrum big data, low valued data density, inaccurate detection outputs, insufficient sharing opportunity discovery, etc. This project proposes to perform the research on basic theories and key technologies from an interdisciplinary perspective of machine learning, data analysis processing and cognitive radio techniques. Then, we begin with building spectrum sensing model, aiming at improving the level of spectrum-efficiency, Energy-efficiency and timeliness in 5G systems, and framing it in terms of spectrum situation pedigree. Furthermore, we will focuse mainly on developing these three aspects of joint sparse sampling, deep cooperative sensing, and multidimensional collaborative forecasting theories and methodologies. As a result, we expect to form an integrated collaborative control mechanism of multi-quality features, and its methods of modeling and solving, which can be employed to significantly improve the spectrum utilization in future 5G systems.

频谱共享是当前无线通信领域国际研究前沿，引起学术界、工业界、标准化组织、政府频谱管理部门等多方力量的密切关注，被认为是解决未来宽带无线移动通信（未来5G）系统中“频谱赤字危机”这一问题直接有效的手段之一。然而，由于未来5G网络超密集、高异构、高动态、智能化的新特征，频谱共享发现面临着海量数据获取成本高、价值密度低、检测结果不准确、机会发现不充分等问题与挑战。本项目拟利用机器学习、数据分析处理和认知无线电技术等多学科交叉融合的方法，以频谱感知过程建模为主线，以综合提高系统频谱效率、能量效率和时效性三个关键性能为目标，以频谱态势的“谱系”研究为切入点，重点开展频谱大数据信息的多维联合稀疏采样（获取）、深度协同感知（融合）与多维协同预测机理（认知）三方面内容的研究，以期形成较为完整的频谱感知多质量特性“一体化”协同控制机制和建模求解方法，为实现未来5G频谱高效利用提供理论和使能技术支撑。

针对5G网络频谱资源紧缺和利用率低的难题，考虑实际无线通信环境频谱高动态性和噪声不确定性的多重因素影响，基于机器学习方法，研究了频谱效率、能量效率和时效性综合提高的协同控制机制。首先，提出了基于单元平均恒虚警率（CA-CFAR，Cell Average-Constant False Alarm Rate）的频谱感知算法以自适应估计能量检测的判决阈值，推导了自适应差分双门限能量检测概率的闭合数学表达式，构建了基于软件无线电外设（USRP, Universal Software Radio Peripheral）的认知网络原型实验平台分析验证噪声功率的自适应估计，以及判决系数与噪声波动系数对检测性能的影响。接着，针对海量频谱大数据感知成本大，信息处理复杂度高的问题，提出了基于降维和聚类的深度协同感知方法。该方法将主成分分析（PCA, Principal Component Analysis）算法分别与改进型K均值聚类（K-Means++）和高斯混合模型（GMM, Gaussian Mixture Model）相融合，先利用PCA算法将所构建的能量特征矩阵转换成低维特征矩阵以降低频谱数据量和训练复杂度，再利用低维特征矩阵分别训练无监督学习K-Means++和GMM聚类器进行频谱状态判决分类，仿真分析结果表明该方法不仅大幅度降低了计算复杂度和学习训练时间，而且提高了复杂动态网络频谱感知的准确度和可靠性。最后，提出了频谱感知多质量特性一体化控制的多维协同优化方法。基于协同频谱感知理论和动态时分多址接入机制，构建了具有能量收集功能的5G认知网络模型，设计了自适应动态协同感知时间与能量收集工作时间帧结构，推导出能量有效性函数表达式，提出了关于感知时间与传输功率分配的多目标联合优化函数，并利用帕累托最优化方法将非凸函数转化为凸函数求取全局最优解。理论分析和仿真结果表明所提方案能够有效提高频谱利用率和能量效率，同时保证了感知时间敏捷性和感知结果的准确性，实现了多质量特性的一体化协同控制。本项目研究成果为5G频谱高效利用提供理论和使能技术支撑。