摘要
在解决卷烟厂设备效能分析中,因数据隐私限制而难以汇集多方数据,以及传统联邦学习在数据异构场景下个性化性能不足的问题.为此,提出了一种新型联邦学习方案,该方法融合了两项核心技术:首先,采用多中心全局聚类策略,将具有相同预测任务的客户端分组并进行组内聚合,以应对数据的非独立同分布特性;其次,在客户端引入注意力凝练网络,动态筛选与适配服务器下发的全局模型参数,从而生成更符合本地数据分布的个性化模型.在真实卷烟厂工业生产数据上的实验结果表明,本方案取得了卓越的性能,其均方误差与平均绝对误差分别达到0.198 5和0.192 4,为同类方法中的最优水平.结论表明,该方案有效地平衡了多方协作的全局效率与客户端的个性化需求,为工业场景下数据隐私与模型精准性兼顾的效能分析提供了可靠途径.
This research aims to address the challenges in analyzing equipment efficiency in cigarette factories,where data privacy concerns prevent data centralization and traditional federated learning falls short in handling data heterogeneity and personalized requirements.To this end,the study proposes a novel federated learning framework that integrates two core techniques:firstly,a multi-center global clustering strategy groups clients with similar prediction tasks and performs aggregation within each group to cope with the non-IID nature of the data;secondly,an attention refinement network is introduced on the client side to dynamically screen and adapt the global model parameters delivered from the server,thereby generating a personalized model that better fits the local data distribution.Experimental results on real industrial production data from cigarette factories demonstrate the superior performance of the proposed model,achieving Mean Squared Error and Mean Absolute Error of 0.1985 and 0.1924,respectively,which are the best among comparable methods.In conclusion,this solution effectively strikes a balance between the global efficiency of multi-party collaboration and the personalized needs of clients,providing a reliable approach for efficiency analysis in industrial scenarios that balances both data privacy and model accuracy.
作者
林熙
唐鑫
戴银波
李云平
赖华
张洲铭
LIN Xi;TANG Xin;DAI Yinbo;LI Yunping;LAI Hua;ZHANG Zhouming(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China;Kunming Cigarette Factory of Hongyun-honghe Tobacco(Group)Co.,Ltd,Kunming 650202,China)
出处
《昆明理工大学学报(自然科学版)》
北大核心
2025年第6期67-75,共9页
Journal of Kunming University of Science and Technology(Natural Science)
基金
云南省重大科技专项计划项目(202502AD080012)
红云红河烟草(集团)有限责任公司科研计划项目(HYHH2023ZN01)。
关键词
联邦学习
设备效能预测
时序预测
个性化联邦学习
深度学习
federated learning
equipment performance prediction
time series prediction
personalized federated learning
deep learning
