工业物联网中的个性化联邦学习算法的研究
Research on Personalized Federated Learning Algorithm in Industrial Internet of Things
刘洋 1吴旭 2刘承坤1
作者信息
- 1. 广西大学计算机与电子信息学院,南宁 530004
- 2. 广西大学计算机与电子信息学院,南宁 530004;海南师范大学信息科学技术学院,海口 571158
- 折叠
摘要
为了在不直接共享原始数据的前提下构建联合模型,联邦学习应运而生.然而,在复杂的工业物联网环境中,联邦学习的应用面临两大挑战:1)工业物联网设备之间彼此异构,掉队和离线设备的存在极大拖慢了联邦学习的训练速;2)不同数据所有者拥有的数据彼此异构,客户端的本地模型差异较大,简单对本地模型进行平均无法获得适用于所有客户端的高质量模型.为了解决上述挑战,本文设计了一个融合数字孪生的联邦学习架构,实现对设备资源的高效调度.此外,本文提出了一个基于参数解耦和聚类的个性化联邦学习算法,既能满足用户的个性化需求,又能实现同构客户端的深度协作.实验结果验证了提出的个性化联邦学习算法的有效性.
Abstract
In order to build a joint model without directly sharing the original data,federated learning came into being.However,in the complex industrial IoT environment,the application of federated learning faces two major challenges:1)Industrial IoT devices are het-erogeneous with each other,and the existence of outdated and offline devices greatly slows down the training speed of federated learn-ing;2)The data owned by different data owners are heterogeneous with each other,and the local models of the clients are quite differ-ent.Simply averaging the local models cannot obtain a high-quality model suitable for all clients.In order to solve the above challen-ges,this paper designs a federated learning architecture that integrates digital twins to achieve efficient scheduling of device resources.In addition,this paper proposes a personalized federated learning algorithm based on parameter decoupling and clustering,which can not only meet the individual needs of users,but also realize the deep collaboration of isomorphic clients.Experimental results verify the effectiveness of the proposed personalized federated learning algorithm.
关键词
联邦学习/工业物联网/数字孪生/个性化Key words
federated learning/industrial internet of things/digital twin/personalization引用本文复制引用
出版年
2025
NETL
NSTL
万方数据