面向能量受限工业物联网设备的联邦学习资源管理  

Federated learning resource management for energy-constrained industrial IoT devices

作　　者：范绍帅[1] 吴剑波 田辉[1] FAN Shaoshuai;WU Jianbo;TIAN Hui(State Key Laboratory of Networking and Switching Technology,Beijing University of Posts and Telecommunications,Beijing 100876,China)

机构地区：[1]北京邮电大学网络与交换技术国家重点实验室,北京100876

出　　处：《通信学报》2022年第8期65-77,共13页Journal on Communications

基　　金：国家重点研发计划基金资助项目(No.2020YFB1807800)。

摘　　要：针对工业物联网联邦学习网络中由设备电池能量有限导致的设备失效、训练中断等问题,并考虑到无线资源受限的影响,提出一种动态的多维资源联合管理算法。首先,以最大化固定训练时间学习精度为目标,将优化问题解耦为相互依赖的电池能量分配子问题、设备资源分配子问题和通信资源分配子问题。其次,基于粒子群优化算法求解能耗预算下设备传输和计算资源分配策略。再次,提出资源块迭代匹配算法求解出最佳通信资源分配策略。最后,提出在线能量分配算法动态调整设备能量分配策略。仿真结果表明,与基准算法相比,所提算法能够提高模型学习精度,在能源不足场景下性能优势更明显。Given the impact of limited wireless resources,a dynamic multi-dimensional resource joint management algo-rithm was proposed,which intended to tackle the problem of device failure and training interruption caused by the limited battery energy in federated learning network in industrial Internet of things(IIoT).Firstly,the optimization problem was decoupled into battery energy allocation,equipment resource allocation and communication resource allocation sub-problems which were interdependent with the goal of maximizing the fixed-time learning accuracy.Then,the equip-ment transmission and computing resource allocation problem were solved based on particle swarm optimization algo-rithm under the given energy budget.Thereafter,the resource block iterative matching algorithm was proposed to optim-ize the optimal communication resource allocation strategy.Finally,the online energy allocation algorithm was proposed to adjust the energy budget allocation.Simulation results validate the proposed algorithm can improve the model learning accuracy compared with other benchmarks,and can perform better in energy shortage scenarios.

关 键 词：联邦学习 电池供电 资源分配 学习效率 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]