基于边缘计算的深度学习模型推断研究毕业论文
2021-10-26 22:36:03
摘 要
近些年来,随着通信技术的不断发展,云计算趋于成熟。加之5G技术的开始普及,越来越多的移动设备和物联网终端接入到移动网络中。移动设备爆炸式增长的计算任务数量使得移动网络的负载越来越大,造成频谱资源紧缺,使得云服务器逐渐无法应对越来越庞大的计算任务。此外,深度学习技术的不断发展,使得移动设备具备了利用深度学习模型进行推断的能力。但是限于移动设备的大小和计算能力,复杂的深度学习模型依旧无法很好地应用于移动终端。移动边缘计算技术将大量服务器部署在距离移动终端更近的移动网络边缘,用户在进行任务计算时可以优先将任务卸载到边缘计算服务器中,由于距离用户更近,因此传输时延和传输能耗更低,且由于服务器位于移动网边缘,相对独立于移动网络的其他部分,因此可靠性更高。本文针对移动设备的深度学习模型推断,设计了一种任务层分类卸载策略,将深度学习模型按照层数划分为多个子任务,根据每层的数据量、运算量等要求进行分类卸载,将各层分别卸载到本地设备、移动边缘计算服务器和云服务器中计算,以此达到三地的深度学习模型的协同推断。仿真实验结果表明,与未采用此分层分类卸载算法的方案相比,本文提出的卸载策略在时延降低方面具有更好的效果,提高了卸载效率的同时,合理地利用了各部分的计算资源。
关键词:移动边缘计算、深度学习、卸载策略、协同推断
Abstract
In last few years, with the rapid development of communication technologies, cloud computing has ripe. In addition, 5G technology has become widespread, more and more smart terminals and Internet of Things devices are connected to mobile networks. The explosive increase in the number of computing tasks on mobile devices has made the load on mobile networks larger and larger, resulting in a shortage of spectrum resources, making cloud servers increasingly unable to cope with increasingly large computing tasks. In addition, with the continuous development of deep learning technology, mobile devices have the ability to make inferences using deep learning models. However, limited to the size and computing power of mobile devices, complex deep learning models still cannot be well applied to mobile terminals. Mobile edge computing technology deploys a large number of servers at the edge of the mobile network closer to the mobile terminal. Users can preferentially offload tasks to the edge computing server when performing task calculations. Because it is closer to the user, the transmission delay and transmission energy consumption are lower, and because the server is located at the edge of the mobile network and is relatively independent of the other parts of the mobile network, the reliability is higher. This paper designs a task level classification offloading strategy for mobile device deep learning model inference. The deep learning model is divided into multiple sub-tasks according to the number of layers, and uninstalled according to the data amount and calculation amount of each layer, and each layer is offloaded to the local device, mobile edge computing server and cloud server for calculation. In this way, the collaborative inference of the deep learning models of the three places is achieved. Through simulation experiments, we can conclude that the unloading strategy proposed in this paper has a better effect in terms of delay reduction compared to the scheme that does not use this hierarchical classification offloading algorithm. While improving the efficiency of unloading, it reasonably uses the computing resources of each part.
Key words: Mobile Edge Computing, Deep Learning, Computing Unloading Strategy, Co-inference
目 录
摘 要 I
Abstract II
第1章 绪论 1
1.1研究的背景及意义 1
1.2国内外研究现状的介绍 4
1.3研究的目的 4
第2章 移动边缘计算和深度学习相关理论 5
2.1 移动边缘计算的架构 5
2.2 移动边缘计算的卸载策略 6
2.2.1 任务卸载技术简介 6
2.2.2 MEC计算任务卸载策略 7
2.3 深度学习模型 8
2.3.1 深度学习简介 8
2.3.2 深度学习模型结构 9
第3章 基于移动边缘计算的深度学习模型卸载策略设计 10
3.1 问题描述 10
3.2 模型设计 10
3.3 算法设计 12
第4章 基于MATLAB的仿真结果及分析 16
4.1 环境及参数设置 16
4.2 仿真结果分析 16
第5章 总结及展望 19
5.1 总结 19
5.2 展望 19
参考文献 21
致 谢 23
第1章 绪论
1.1研究的背景及意义
近些年来,随着科技的不断快速发展,智能移动终端设备和物联网终端设备已经全面普及,逐渐成为了人们日常生活中必不可少的工具。手机、平板电脑、蓝牙音箱等设备为人们的生活带来了极大的便利,甚至改变了人们的生活习惯。科技带来的生产力正逐渐将人类从枯燥的机械劳动中解放出来。然而,智能移动终端和物联网终端设备的工作需要进行大量的计算任务,仅仅依靠终端的计算能力远不足以完成如此巨量的工作任务。因此,移动云计算(Mobile Cloud Computing,MCC)技术应运而生,用来承接终端设备的大量计算任务[1]。移动云计算将移动网络和云服务器相结合,利用云服务器的大容量和计算能力,通过移动网络与各类终端设备进行计算任务数据的传递,从而实现代替智能终端设备进行计算的目的。