基于深度学习的GEO/LEO毫米波卫星网络路由策略研究

GEO/LEO millimeter wave satellite network has a series of advantages, including large coverage, large capacity and high-speed transmission rate. Therefore, it has become an important component of Chinese national space network architecture. In the satellite network, routing is the core technology to keep the stable and efficient information exchange. Under the scenario of large scale satellites and multi-services coexistence, traditional routing strategy degrades the system efficiency and impacts the user perception, thus becoming the bottle-neck of GEO/LEO millimeter wave satellite network. This project investigates on the key routing procedures, including topological division, message aggregation and distribution, service routing and quality of service (QoS) management. The major innovation includes three aspects: 1. In order to address the problem of large-scale hierarchical dynamic network topology, the hierarchical network virtual topology technology is studied and a snapshot-based self-tuning quasi-static topology scheme is proposed. 2. In order to address the collision problem of LEO node state information reporting, we study the multi-node information aggregation and distribution technology based on software defined network (SDN), and we propose the LEO uplink message collision avoidance scheme based on the link delay difference. 3. On the basis of above-mentioned two steps, the multi-service routing and QoS management techniques based on deep learning are studied, and further a routing scheme based on deep belief network is proposed, in order to address the multi-service routing and QoS management problems. The research results and the proposed routing strategy of this project aims at reaching the world-class level. This project employs the deep learning technique to enhance the smart capability of each satellite. In all, the project has both scientific value and application prospect for the future GEO/LEO millimeter wave satellite network.

GEO/LEO毫米波卫星网络以其广覆盖、大容量、高速率的优势成为国家空间网络体系的重要组成部分。路由策略是保证全网信息高效稳定交互的核心技术。在卫星大数量级和多业务共存场景下，传统路由策略成为制约系统效率和用户感知的瓶颈。本课题深入研究路由策略中的拓扑划分、消息汇聚分发、业务选路及QoS管理，创新点包括：1.针对大规模分层动态网络拓扑化问题，研究网络虚拟拓扑技术，提出基于快照的自修正准静态拓扑划分方案；2.针对LEO节点状态信息上报碰撞问题，研究基于SDN的多节点信息汇聚分发技术，提出基于链路时延差的上行防碰撞方案；3.在前两者的基础上，针对多业务选路及QoS管理问题，研究基于深度学习的路由策略，提出基于张量的深度信念网络路由方案。本课题研究成果以赶超和引领国际先进水平为目标，突破传统路由策略研究思路，运用深度学习有效提升卫星节点智能性，对于未来毫米波卫星网络具有重要的科学意义和应用前景。

GEO/LEO毫米波卫星网络中路由策略所面临的问题是拓扑划分、消息汇聚分发、业务选路及QoS管理。项目组针对以上问题开展研究，成果如下：.（1）基于虚拟拓扑的大规模分层网络动态网络拓扑化技术研究.项目组提出了一种基于网络状态的低轨卫星网络动态路由协议，通过对邻居节点连接状态的定期检测，以及节点负载信息的动态交互，动态检测网络拓扑，对异常情况及时感知并相应调整。项目组提出的动态路由协议包括路由表的操作与维护、节点负载获取、邻居节点状态探测、多任务处理机制等模块，在NS3平台对该协议进行了模拟仿真，并在Linux系统中实现了该动态路由协议，验证了在实际网络中该路由算法的可行性。.（2）基于SDN的多节点信息汇聚分发技术研究.项目组提出了一种适用于GEO/LEO网络的LEO上行异步协同多址接入方法，在GEO接收节点处通过串行干扰消除方法解决LEO上行消息在异步随机接入时的碰撞问题。该方法不需要系统同步，减少了用于系统同步所引入的功率和信令开销，降低了终端的功耗。.（3）基于深度学习的多业务选路及QoS管理技术研究.项目组提出了一种基于深度强化学习的低轨卫星网络路由策略，采用虚拟节点策略对低轨卫星网络拓扑结构进行建模，提出两跳链路状态感知策略，降低链路状态更新的开销并使链路状态更具实时性和准确性。该策略将深度强化学习与卫星路由算法相结合，采用DDQN模型感知卫星网络拓扑结构，并通过NS3仿真验证了策略的可行性。.在项目执行期间，项目组共发表学术论文8篇（SCI论文6篇，EI论文2篇），申请国家发明专利4项。研究结果表明，项目组所提出的路由策略可以有效降低端到端时延和丢包率，保证系统的整体吞吐量，为下一代卫星网络的研究和建设提供了理论依据和数据参考。