摘要
针对工厂端数据量不均衡、新增数据样本多和联邦学习中通信成本较高的问题,提出一种基于云、边、端架构的分层联邦增量学习算法.首先将行业联合模型下发到设备端,训练得到本地模型,将本地模型上传到边缘端进行融合,经多次迭代后,上传到云端加权融合行业联合模型;同时,增量学习监控模块持续监控新增数据样本作为下一次模型训练的数据样本.此方案通过计算设备端的数据权重值和新增数据样本的模型影响度,修改模型加权聚合策略,持续优化行业联合模型,并防止模型发生倾斜.实验结果表明:分层联邦学习架构可明显提高通信效率,且相比于FedAvg算法的训练集和测试集准确率分别提高了12.34%和5.36%;增量学习算法可达到持续优化行业联合模型的效果,每次训练新增数据样本都可稳定提高行业联合模型的准确度.
Aiming at the problems of unbalanced factory-side data volume,large number of new data samples,and high communication costs in federated learning,a layered federated incremental learning algorithm based on cloud,edge,and end architecture was proposed.First,the industry joint model was sent to the device side,the local model was trained,and the local model was uploaded to the edge end for fusion.After multiple iterations,it was uploaded to the cloud for weighted fusion of the industry joint model.Meanwhile the incremental learning monitoring module continuously monitored the new incremental data sample,which was used as the data sample for the next model training.This solution modified the model weighted aggregation strategy by calculating the data weight value on the device side and the model influence degree of newly added data samples,continuously optimized the industry joint model,and prevented the model from being skewed.Experimental results show that the layered federated learning architecture can significantly improve the communication efficiency,and the accuracy of the training set and test set are increased by 12.34%and 5.36%,respectively,compared with the FedAvg algorithm;the incremental learning algorithm can achieve the effect of continuous optimization of the industry joint model,and new data sample for each training can steadily improve the accuracy of the industry joint model.
作者
路松峰
屠向阳
周军龙
王穆
LU Songfeng;TU Xiangyang;ZHOU Junlong;WANG Mu(School of Cyber Science and Engineering,Huazhong University of Science and Technology,Wuhan 430074,China;Shenzhen Huazhong University of Science and Technology Research Institute,Shenzhen 518063,Guangdong China;School of Computer Science and Technology,Huazhong University of Science and Technology,Wuhan 430074,China;Clinical Research Institute,Affiliated South China Hospital,University of South China,Hengyang 421001,Hunan China)
出处
《华中科技大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2023年第10期12-18,共7页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家重点研发计划资助项目(2021YFB2012202)
湖北省重点研发计划资助项目(2021BAA038,2021BAA171)
深圳市科技计划基础研究资助项目(JCYJ20210324120002006)
深圳市科技计划技术攻关项目(JSGG20210802153009028)
湖南省教育厅科研优秀青年项目(20B491)
南华大学衡阳医学院资助项目(衡医发(2021)1-2-7号)
湖南省科技创新重点工程项目(2020SK1012)
湖北省科技重大专项资助项目(2020AEA011)。
关键词
工业设备
联邦学习
增量学习
边缘计算
行业联合模型
industrial equipment
federated learning
incremental learning
edge computing
industry joint mode
