摘要
基于海量样本数据的深度神经网络可以显著提升电力变压器的故障诊断效果,但实际环境中各变压器本地数据被“数据孤岛”隔离,难以进行集中式深度训练。联邦学习可以实现多源数据的共同建模,但各变压器的非独立同分布不平衡数据会使算法性能严重下降。针对上述问题,提出一种基于联邦学习的含不平衡样本数据电力变压器故障诊断机制。与传统深度学习集中训练方式不同,文中采用多个参与者分布训练方式。每个参与者采用LeNet-5深度神经网络进行本地训练,中央服务器将本地模型进行聚合,同时引入改进的数据共享策略对云端共享数据进行选择性下发,以降低非独立同分布数据的不平衡性。实验结果证明,该机制实现了各变压器独立且不平衡数据的协同训练,对变压器故障类型的诊断精度可达到97%。
The deep neural network based on massive sample data can significantly improve the fault diagnosis effect of power transformers,but in the actual environment,the local data of each transformer are isolated by“data islands”,which makes it difficult to conduct the centralized in-depth training.Federated learning can realize the common modeling of multi-source data.However,the non-independent and identically distributed unbalanced data of each transformer will seriously degrade the performance of the algorithm.Aiming at the above problems,a power transformer fault diagnosis mechanism with unbalanced sample data based on federated learning is proposed.Different from the traditional intensive training method of deep learning,this paper adopts the distributed training method of multiple participants.Each participant uses the LeNet-5 deep neural network for local training,the central server aggregates local models,and introduces an improved data sharing strategy to selectively distribute the cloud shared data to reduce the imbalance of non-independently and identically distributed data.The experimental results show that the mechanism realizes the cooperative training of each transformer with independent and unbalanced data,and the diagnosis accuracy of transformer fault types can reach 97%.
作者
郭方洪
刘师硕
吴祥
陈博
张文安
葛其运
GUO Fanghong;LIU Shishuo;WU Xiang;CHEN Bo;ZHANG Wenan;GE Qiyun(College of Information Engineering,Zhejiang University of Technology,Hangzhou 310014,China;Kerun Intelligent Control Co.,Ltd.,Quzhou 324100,China)
出处
《电力系统自动化》
EI
CSCD
北大核心
2023年第10期145-152,共8页
Automation of Electric Power Systems
基金
国家自然科学基金资助项目(61903333)。
关键词
电力变压器
故障诊断
LeNet-5网络
联邦学习
非独立同分布数据
power transformer
fault diagnosis
LeNet-5 network
federated learning
non-independently and identically distributed data
