基于不平衡样本的中压电缆分布式故障诊断方法  

Distributed Fault Diagnosis Method for Medium Voltage Cables Based on Unbalanced Samples

作　　者：任广振 曹俊平 周铭权 陈荣鑫 陈森杰 REN Guangzhen;CAO Junping;ZHOU Mingquan;CHEN Rongxin;CHEN Senjie(State Grid Zhejiang Electric Power Co.,Ltd.,Hangzhou 310000,China;Electric Power Research Institute of State Grid Zhejiang Electric Power Co.,Ltd.,Hangzhou 310014,China;Hangzhou Juqi Information Technology Co.,Ltd.,Hangzhou 311400,China)

机构地区：[1]国网浙江省电力有限公司,浙江杭州310000 [2]国网浙江省电力有限公司电力科学研究院,浙江杭州310014 [3]杭州巨骐信息科技股份有限公司,浙江杭州311400

出　　处：《微型电脑应用》2025年第2期42-46,共5页Microcomputer Applications

基　　金：国网浙江省电力有限公司科技项目(5211DS20007P)。

摘　　要：由于高、中压电缆存在数据不平衡问题,难以满足深度网络训练要求,因此,提出基于不平衡样本的中压电缆分布式故障诊断方法。该方法采用双向长短期记忆网络为本地模型训练方式,在中央服务器聚合中利用联邦学习对模型聚合并更新。由多个用户分布学习,缓解模型训练压力,提高训练效率。同时引入数据共享策略,将云端数据下发给本地模型,降低数据的不平衡性。结果证明,所提方法实现了用户在独立且不平衡数据下的协同训练。Due to the problem of data imbalance in high and medium voltage cables,it is difficult to meet the requirements of deep network training.Therefore,a distributed fault diagnosis method for medium voltage cables based on imbalanced samples is proposed.This method adopts a bidirectional long short-term memory network as the local model training method,and uses federal learning to aggregate and update the model in the central server aggregation.Distributed learning by multiple participants to alleviate model training pressure and improve training efficiency.Simultaneously,introducing a data sharing strategy to distribute cloud data to local models,reducing data imbalance.The results show that this method achieves collaborative training of participants under independent and unbalanced data,and the accuracy of cable fault diagnosis can reach 96%.

关 键 词：分布式故障诊断 联邦学习 双向长短期记忆网络 数据孤岛 

分 类 号：TM75[电气工程—电力系统及自动化]