【作者】 王成；

【导师】 杨清海；

【作者基本信息】 西安电子科技大学 ， 通信与信息系统， 2023， 硕士

【摘要】 随着第六代移动通信技术(6G)的不断发展,使得低时延和高可靠通信成为可能,基于智能边缘设备与边缘计算(EC)相结合的边缘智能(EI)受到人们的广泛关注。边缘智能旨在通过传感器、智能边缘设备以及网络设施组成的多接入边缘计算(MEC)为用户提供人工智能服务。然而,边缘设备往往具有存储和功率有限的特点,不能独自承担人工智能(AI)任务,在资源受限的边缘设备上部署复杂的神经网络模型通常是不可行的。同时,某些边缘设备的数据具有一定的隐私性,这些隐私数据不能随意通过网络进行传输。为了进一步分析如何通过智能边缘设备进行数据处理,以及云边协同在边缘智能中的应用,本文研究了以云边协同的方式训练神经网络模型以及对神经网络模型进行在线优化。我们搭建了协同学习验证平台,执行了神经网络的训练、优化以及推理过程,验证了算法的高效性。针对传统联邦学习中存在的通信瓶颈问题,提出了多层级云边协同训练算法,在联邦学习架构的基础上增加了边缘云,为联合学习的边缘侧算法收敛提供了保障。仿真结果表明,相对于联邦平均算法,多层级云边协同训练算法在中心云服务器上具有更低的通信开销。此外,该算法与联邦平均算法具有相同的收敛速度。为了解决边缘设备算力不足的问题,提出了分布式云边协同优化算法,通过边缘设备的本地数据,以较低的计算开销实现了神经网络模型的优化,压缩神经网络模型的体积,与经典的泰勒评估准则相比,分布式云边协同优化算法只需要占用很低的上行通信链路,并且在相同的压缩率下,通过分布式云边协同优化算法压缩的神经网络拥有更高的性能。构建了协同学习验证平台,实现以多层级云边协同训练算法和边缘设备为基础的神经网络模型训练工作,以及通过分布式云边协同优化算法为神经网络提供不依赖数据传输的在线优化方法,最后将神经网络应用到智慧农业场景中,根据场景反馈调节策略函数,触发再训练,形成训练与推理的完整工作。更多还原

【Abstract】 With the continuous development of the Sixth Generation Mobile Communication Technology(6G),low latency and high reliability communication become possible.Edge Intelligence(EI)based on the combination of intelligent edge devices and Edge Computing(EC)has attracted extensive attention.Edge intelligence aims to provide users with artificial intelligence services through Multi-access Edge Computing(MEC)composed of sensors,intelligent edge devices and network facilities.However,edge devices often have limited storage and power,and cannot undertake Artificial Intelligence(AI)tasks alone.Deploying complex neural network models on resource constrained edge devices is generally not feasible.At the same time,the data of some edge devices has a certain degree of privacy,and these private data cannot be arbitrarily transmitted through the network.In order to further understand how to conduct data analysis through intelligent edge devices and the application of cloud edge collaboration in edge intelligence,this paper studies the training and online optimization of neural network models using cloud edge collaboration.We have built a collaborative learning platform,executed the training,optimization,and referencing processes of neural networks,and verified the efficiency of the algorithm.Aiming at the communication bottleneck problem in traditional federated learning,a multilevel cloud edge collaborative training algorithm is proposed,which adds edge clouds to the federated learning architecture,providing a guarantee for the convergence of the edge side algorithm of federated learning.Simulation results show that the multi-level cloud edge collaborative training algorithm has lower communication overhead on the central cloud server compared to the federated average algorithm.In addition,this algorithm has the same convergence speed as the federal average algorithm.In order to solve the problem of insufficient computing power for edge devices,a distributed cloud edge collaborative optimization algorithm is proposed.Using local data from edge devices,the optimization of neural network models is achieved with low computational overhead,compressing the volume of neural network models.Compared to the classic Taylor evaluation criterion,the distributed cloud edge collaborative optimization algorithm only requires a very low uplink consumption,and at the same compression rate,Neural networks compressed through distributed cloud edge collaborative optimization algorithms have higher performance.A collaborative learning verification platform was constructed to implement neural network model training based on multi-level cloud edge collaborative training algorithms and edge devices,as well as provide online optimization methods for neural networks that do not rely on data transmission through distributed cloud edge collaborative optimization algorithms.Finally,the neural network was applied to intelligent agricultural scenarios,and the strategy function was adjusted based on scene feedback to trigger retraining,forming a complete work of training and referencing.更多还原