Research On Adaptive Routing Optimization For LEO Satellite Networks

With the rapid development of the global Internet,the era of satellite Internet communication is accelerating.Countries around the world are seeking technological breakthroughs and striving to seize the next opportunity,while our country has clearly included satellite Internet as a communication network infrastructure into the scope of new infrastructure,and strives to take the lead in the future communication network infrastructure.Due to its unique composition and operation mode,the satellite network has different characteristics from the ground network,especially in routing optimization,which has different challenges from the ground network.First of all,the satellite network topology is in periodic changes.Traditional satellite network routing schemes adopt topology control strategies to shield topology dynamics,but this method is simple in mechanism and lacks flexibility and fault tolerance;Secondly,the traditional satellite network routing lacks active network status awareness,only guarantees the reachability of the route,and the user experience is poor.The route calculation method based on the shortest path also needs to be considered.The development of emerging network technologies has brought new light to solve the problem of satellite network routing optimization.In-band network telemetry technology not only avoids the "observer" effect and "uncertainty" effect of traditional network measurement,but also minimizes the additional burden on the network as much as possible to achieve accurate and efficient lightweight measurement.The development of deep reinforcement learning,graph neural network,simplifies the complex modeling work between network state and routing decision.This topic focuses on the adaptive routing optimization problem of low-orbit satellite network,with the help of two technologies of in-band network telemetry and deep reinforcement learning,and has achieved the following results:(1)A lightweight implementation of in-band network telemetry is designed.With the help of the idea of software-defined network,we decouple the satellite network,use the data plane programming language P4 to realize lightweight in-band network measurement,and actively report network status information.Through the verification of the simulation platform,this architecture can realize active network status awareness and provide decision-making basis for intelligent routing algorithms.(2)A generalizable intelligent satellite network routing algorithm is proposed.In the face of complex network state information,deep reinforcement learning technology is used to learn the mapping relationship between state information and routing strategies.At the same time,in view of the weak generalization ability of traditional deep reinforcement learning technology,a graph neural network is introduced to learn graph features to enhance the generalization ability of routing algorithms.The experimental results show that the intelligent routing algorithm proposed in this topic reduces the maximum link utilization by an average of 18.5%and 15.5%respectively compared with the SP and ECMP algorithms;At the same time,the generalization of the project is tested.The experimental results show that the routing algorithm of this project has a certain generalization ability and can cope with general topology changes.