基于分层联邦学习的无人机小基站RAN切片方法  

Hierarchical Federated Learning-Based RAN Slicing for Drone-Small-Cells

作　　者：殷珉 沈航 王天荆[1] 白光伟[1] YIN Min;SHEN Hang;WANG Tian-jing;BAI Guang-wei(College of Computer and Information Engineering,Nanjing Tech University,Nanjing,Jiangsu 210000,China)

机构地区：[1]南京工业大学计算机与信息工程学院,江苏南京210000

出　　处：《电子学报》2023年第7期1774-1780,共7页Acta Electronica Sinica

基　　金：国家自然科学基金(No.61502230,No.61501224);江苏省自然科学基金(No.BK20201357);江苏省六大人才高峰高层次人才资助项目(No.RJFW-020)。

摘　　要：针对多架无人机共同为地面用户提供差异化服务的场景,本文提出一种基于分层联邦学习的动态RAN(Radio Access Network)切片框架,目的是提升切片性能隔离效果、减少协同训练过程的通信代价.考虑到无人机动态部署和数据不足等特点,本文通过数据增广促进本地模型训练.为了使得距离地面基站较远的无人机有更多机会参与联邦学习并降低通信代价,本文根据位置和数据分布信息设计支持边缘模型聚合的无人机分簇策略.在此基础上,本文探索基于注意力机制的边缘和全局模型聚合方案,以增强全局模型的泛化能力.仿真结果表明,与联邦平均和分布式LSTM(Long Short-Term Memory)相比,所提方案在切片性能隔离的时长占比上分别有8.4%和16.5%的提升,并降低了无人机协同训练的通信代价.For the scenario where multiple drone-small-cell provide differentiated services for ground users,a dynam⁃ic radio access network(RAN)slicing framework based on hierarchical federated learning is proposed.The goal is to im⁃prove slice performance isolation and reduce the communication cost in collaborative model training.Data augmentation is introduced to promote local model training and improve model performance,considering the dynamic deployment and in⁃sufficient data of drones.Then,a clustering strategy for drone-small-cell depending on geographic location and data distri⁃bution is designed to support edge model aggregation.Hence,member drones at the edge have more opportunities to partic⁃ipate in federated learning with reduced communication costs.On this basis,an attention mechanism-based aggregation scheme for edge and global models is explored to improve the generalization ability of the global model.Simulation results show that compared with federated averaging and distributed long short-term memory(LSTM),the duration of slice perfor⁃mance isolation of the proposed scheme increases by 8.4%and 16.5%,respectively,with reduced communication cost of drone collaborative model training.

关 键 词：无人机小基站 RAN切片 分层联邦学习 边缘模型聚合 注意力机制 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]