基于深度学习的智能频谱认知技术研究

Spectrum sensing, which is the fundamental task of cognitive radio (CR), is a technology used by cognitive users (CUs) to detect the occupation of licensed spectrum by primary users (PUs). The existing methods only make detection by sole artificial features, such as energy or eigenvalue, which have limited detection performance. On the other hand, the existing methods also fail to explore the deep information of spectrum data, such as PU’s arriving rate and system traffic, which only have the single detection function. Motivated by this, this proposal tries to use the deep learning technology to explore the deep information of spectrum data, and proposes a novel idea of deep learning based smart spectrum cognition. Firstly, we design a deep auto encoder network model which is suitable for spectrum data, thus it can extract the feature information through feature learning; Secondly, we propose a PU’s state based user-level spectrum deep cognition method, in which we combine the spectrum state detection and state duration prediction together to design the optimization mechanism, achieving the task of the spectrum hole cognition; Finally, we further propose a PU traffic based system-level spectrum deep cognition method, in which we design the system traffic and spectrum hole cognition oriented deep learning network model to realize the goal of spectrum resource cognition. The proposed scheme can not only improve the detection performance of spectrum state, but also obtain the deep information of spectrum data. This research will develop a novel spectrum cognition method and contribute to the theoretical support for the next generation smart cognitive radio system.

频谱感知是认知用户对授权主用户频谱占用情况进行检测的技术方法，是认知无线电的基础和重要环节。现存算法仅通过单一的能量、特征值等人工特征进行检测，检测性能有限；而且算法没有发掘主用户到达率、系统流量等深层信息，感知功能单一。针对这两大局限性，本项目利用深度学习技术发掘频谱数据的深层信息，提出基于深度学习的智能频谱认知新思路。首先，提出适用于描述频谱数据特征的深度自编码网络模型，通过特征学习提取特征信息；其次，提出基于用户状态特征的用户级频谱深度认知方法，通过建立频谱状态检测和持续时间预测联合优化的机制，实现频谱空穴认知；最后，提出基于多用户传输特征的系统级频谱深度认知方法，通过设计面向系统流量认知和频谱空穴认知的深度学习网络模型，实现频谱资源认知。研究方案提高了频谱状态检测性能，并且能够认知频谱数据的深层信息。研究成果将形成新的频谱认知方法，为新一代智能认知无线电系统提供理论支撑。

频谱感知技术是认知用户对授权主用户频谱占用情况进行检测的技术方法，该技术在未来无线通信、救援通信、战场通信等场景有着广泛的应用前景。针对传统频谱感知检测性能有限、感知功能单一的局限性，本课题利用深度学习技术发掘频谱数据深层信息，提出了智能频谱认知新思路。课题组从面向频谱数据的深度学习模型、面向用户的智能频谱深度认知方案、以及面向系统的智能频谱深度认知方案展开研究，实现了基于深度学习的智能频谱认知。在自然科学基金的资助下，课题组发表学术论文14篇，其中，SCI期刊论文11篇。所提出的具有协方差意识的深度卷积神经网络频谱感知算法发表在期刊IEEE J. Sel. Areas Commun. (IF 9.144 一区，Top期刊)，提出的基于深度迁移学习的频谱共享方法发表在期刊IEEE Trans. Wireless Commun. (IF 7.016 一区)，提出的辅助无线电环境感知的深度残差网络信道估计算法发表在期刊IEEE Trans. Wireless Commun. (IF 7.016 一区)，提出的可频谱共享的共生无线电系统波束成形算法发表在期刊IEEE Trans. Commun. (IF 5.083 一区)，提出的基于拟合优度的频谱数据特征设计、基于无监督深度学习的频谱感知算法均发表在IEEE Trans. Vehicle Technology (IF 5.978 二区)。课题组还研究了基于主用户行为的深度卷积长短期记忆神经网络频谱感知算法和不同系统模型下无线电环境感知方法，这些研究成果发表在期刊IEEE Wireless Commun. Lett. (IF 4.348 二区)和 IEEE Commun. Lett. (IF 3.436 二区)。项目成果为实现高智能化的频谱认知提供了理论和方法保障，所提出的方法对智能无线电领域的研究具有普适意义和推动作用。