【作者】 刘鑫；

【导师】 赵鑫；

【作者基本信息】 内蒙古大学 ， 通信工程（含宽带网络、移动通信等）（专业学位）， 2025， 硕士

【摘要】 第六代移动通信(Sixth Generation,6G)和物联网(Internet of Things,Io T)的迅猛发展为智能交通系统(Intelligent Transportation Systems,ITS)和车联网(Internet of Vehicles,Io V)技术带来了前所未有的机遇,但也引发了诸多严峻挑战。传统车载边缘计算(Vehicular Edge Computing,VEC)网络由于车辆高速移动、网络拓扑频繁变化、边缘资源受限以及单一地面网络覆盖不足等原因,往往难以满足自动驾驶、实时导航等延迟敏感型应用对低时延和高可靠性服务的要求。同时,在大规模车辆接入和多任务并发卸载场景中,如何在有限资源下实现高效的计算任务分配、负载均衡以及能耗优化也成为亟待解决的关键问题;此外,多跳通信引发的下行链路延迟也会对整体用户服务质量(Quality of Service,Qo S)造成不良影响。针对车辆的高速移动和网络拓扑变化频繁导致的通信连接不稳定和高时延问题,本文提出了一种基于数字孪生(Digital Twin,DT)与深度强化学习(Deep Reinforcement Learning,DRL)算法的智能任务卸载优化方案。利用DT技术构建物理层与虚拟孪生层之间的双向映射机制,实现对车辆状态、路侧单元(Roadside Unit,RSU)负载及无线信道条件的实时感知与预测,从而提前识别因车辆移动可能引发的链路中断风险,为后续卸载决策提供准确的环境信息,进而降低整体任务延迟和通信风险。基于此,设计了一种轻量化的DRL算法,采用异步优势策略-价值(Asynchronous Advantage Actor-Critic,A3C)框架实现各智能体(例如车辆、RSU等)的实时在线决策,该方法在动态环境下能够并行采样与更新策略,并综合考虑任务延迟、能耗、负载均衡及用户满意度等多目标,通过折现奖励累计实现长期系统效用最大化,有效解决了传统静态优化方法计算量大、响应缓慢的问题。为了解决地面网络覆盖不足导致的车辆多任务并发卸载能耗高,时延大,服务质量低的问题,本文进一步将DT技术集成到空天地一体化网络(Space-Air-Ground Integrated Networks,SAGINs)中,以增强VEC的框架。通过整合卫星、无人机(UAV)和地面基站构建出天基广覆盖、空基灵活中继、地基高效处理的三维协同网络。扩展了服务覆盖范围,而且通过跨层资源协同机制实现了多域资源的动态分配和调度,能够在偏远或信号薄弱区域保障通信和任务卸载服务的稳定性。为进一步提升大规模系统中全局资源调度的性能,本文提出了一种基于多智能体深度强化学习(Multi-Agent Deep Reinforcement Learning,MADRL)算法与基于引导点的协方差矩阵自适应进化策略(Guiding Point-based Covariance Matrix Adaptation Evolution Strategy,GP-CMA-ES)的联合优化机制。在这一方案中,不同子智能体负责局部任务卸载决策,而GP-CMA-ES利用全局历史数据及引导点策略对各子智能体局部最优解进行全局优化,并通过奖励转发与集中协调反馈机制实现全局任务卸载、计算资源分配和能耗控制的协同优化,从而显著降低系统总延迟和能耗,同时提高任务成功率。仿真实验结果表明,本文所提出的整体方案在动态车联网环境中能够大幅降低任务卸载延迟,优化能耗和负载分配,并在超可靠低延迟通信(Ultra-Reliable Low-Latency Communication,URLLC)场景下实现更高的用户满意度和系统效益。该方案为应对高动态、高复杂度车联网场景中的任务卸载和资源调度问题,提供了一种兼具实时性、鲁棒性与扩展性的解决方案。更多还原

【Abstract】 The rapid evolution of Sixth Generation(6G)wireless communications and the Internet of Things(Io T)has brought unprecedented opportunities to Intelligent Transportation Systems(ITS)and the Internet of Vehicles(Io V),while simultaneously posing a range of critical challenges.Traditional Vehicular Edge Computing(VEC)networks often fail to meet the stringent low-latency and high-reliability requirements of delay-sensitive applications such as autonomous driving and real-time navigation.This inadequacy stems from several key factors:the high mobility of vehicles,dynamic changes in network topology,limited edge computing resources,and insufficient coverage provided by solely terrestrial networks.Moreover,in large-scale vehicular access scenarios with concurrent task offloading demands,achieving efficient task scheduling,load balancing,and energy optimization under constrained resources becomes a pressing issue.Additionally,the multi-hop communication process exacerbates downlink latency,thereby degrading the overall Quality of Service(Qo S)for end users.To address the instability of communication links and increased latency caused by vehicular mobility and frequent topological variations,this paper proposes an intelligent task offloading optimization framework based on Digital Twin(DT)technology and Deep Reinforcement Learning(DRL).A bidirectional mapping mechanism is established between the physical layer and the virtual twin layer,enabling real-time perception and prediction of vehicle states,roadside unit(RSU)workloads,and wireless channel conditions.This allows for early detection of potential link disruptions due to vehicle mobility and provides accurate environmental context for making informed offloading decisions,effectively mitigating communication risks and reducing overall task delay.Based on this foundation,we design a lightweight DRL algorithm that adopts the Asynchronous Advantage Actor-Critic(A3C)architecture to enable real-time,distributed decision-making among intelligent agents such as vehicles and RSUs.The proposed approach supports parallel sampling and policy updating in dynamic environments and simultaneously optimizes multiple objectives including task delay,energy consumption,load distribution,and user satisfaction.By accumulating discounted rewards,it maximizes long-term system utility while overcoming the computational overhead and sluggish response of traditional static optimization methods.To further address the issues of high energy consumption,large latency,and suboptimal Qo S resulting from limited terrestrial network coverage during multi-task vehicular offloading,we integrate DT technology into a Space-Air-Ground Integrated Network(SAGIN)architecture to enhance the VEC framework.By jointly leveraging satellites,unmanned aerial vehicles(UAVs),and terrestrial base stations,we construct a three-tier cooperative network characterized by broad satellite coverage,flexible UAV relaying,and efficient ground processing.This architecture not only extends service coverage but also supports dynamic cross-layer resource orchestration and allocation,thereby ensuring stable connectivity and task offloading even in remote or weak-signal areas.To enhance global resource coordination in large-scale systems,we further propose a hybrid optimization scheme combining Multi-Agent Deep Reinforcement Learning(MADRL)with a Guiding Point-based Covariance Matrix Adaptation Evolution Strategy(GP-CMA-ES).In this framework,individual sub-agents make decentralized offloading decisions,while GP-CMA-ES performs global optimization of local solutions by utilizing historical data and guiding points.A reward forwarding and centralized coordination mechanism ensures effective collaboration across agents for task offloading,computing resource allocation,and energy management.Simulation results demonstrate that the proposed solution significantly reduces task offloading delay,improves energy efficiency and load balancing,and achieves superior user satisfaction and system utility in Ultra-Reliable Low-Latency Communication(URLLC)scenarios.This work provides a robust,scalable,and real-time solution for addressing the complexities of task offloading and resource orchestration in dynamic vehicular environments.更多还原