基于学习排序的计量装置故障严重程度评估方法  

Fault severity assessment method of metering device based on learning ranking

作　　者：郭光 王海元 彭潇 王智 梁睿琪 赵景波 GUO Guang;WANG Haiyuan;PENG Xiao;WANG Zhi;LIANG Ruiqi;ZHAO Jingbo(State Grid Hunan Electric Power Corporation Limited,Changsha 410000,China;Hunan Province Key Laboratory of Intelligent Electrical Measurement and Application Technology,Changsha 410000,China;College of Information Science and Engineering,Southeast University,Nanjing 210096,China;College of Electrical and Information Engineering,Hunan University,Changsha 410082,China)

机构地区：[1]国网湖南省电力有限公司,湖南长沙410000 [2]智能电气量测与应用技术湖南省重点实验室,湖南长沙410000 [3]东南大学信息科学与工程学院,江苏南京210096 [4]湖南大学电气与信息工程学院,湖南长沙410082

出　　处：《中国测试》2024年第6期176-182,共7页China Measurement & Test

基　　金：国家电网公司科技项目资助(5216AG21000K)。

摘　　要：现有的电能计量装置状态检测研究通常仅涉及故障的检测,未考虑故障严重程度的区分或排序问题。为解决这一问题,该文提出一种基于学习排序的电能计量装置故障严重程度的评估方法。首先,根据电能计量装置的运行监测数据,设计包含14个分量的特征向量;其次,选择sigmoid函数对特征进行概率和评分的转换;再次,采用RankNet神经网络计量装置故障程度的分级。最后,该文采用国家电网公司某省2020-2021年部分地区电能计量装置运行监测数据进行测试。训练集包含2万组样本。所提方法在测试集上判断正确和错误的样本对数分别为9769和231,准确率达97.69%。此外,对于故障严重程度评分结果前50的故障样本,模型的平均排序偏移量为0.96,说明该文方法对于排序靠前的故障具有很好的排序效果。同时,模型仅需10次左右迭代即可收敛,能有效帮助工作人员提高电能计量装置检修效率。Existing researches on the state detection of power metering devices usually only involve the detection of faults,and do not consider the problem of distinguishing or sorting the severity of faults.To solve this problem,this paper proposes a method for power metering devices evaluation based on learning ranking.Firstly,we design a feature set containing 14 components according to the monitoring data of the electric energy metering device,designing a;secondly,selecting the sigmoid function to convert the probability and the score of the features;thirdly,using the RankNet to classify the failure degree of the metering device.Finally,this paper uses the operation monitoring data of electric energy metering devices in some areas of a province of State Grid Corporation of China from 2020 to 2021 for testing.The training set contains 20,000 sets of samples,the test set contains 10,000 sets of samples.The number of correct and incorrect samples of the proposed method on the test set is 9769 and 231,respectively,with an accuracy rate of 97.69%.In addition,the average sorting offset of the model is 0.96 for the top 50 fault samples in the fault severity score,indicating that the method in this paper has a good sorting effect on the top faults.At the same time,the model only needs about 10 iterations to converge,which can effectively help the staff to improve the maintenance efficiency of the power metering device.

关 键 词：RankNet 计量装置故障 排序 评分函数 

分 类 号：TB9[一般工业技术—计量学] TM933.4[机械工程—测试计量技术及仪器]