摘要
目前变压器故障诊断最常用的方法为三比值法,但在大量实际应用过程中,单凭经验或统计学原理很难进一步提升故障诊断准确率。为此,提出通过构建深度卷积神经网络(deeply convolutional neural network,DCNN)模型以提升设备故障诊断准确率,DCNN模型能够识别设备监测数据的局部特征以及不同时刻监测数据间的相关信息;深度神经网络(deep neural network,DNN)模型可以无限逼近目标函数,能够以任务为导向,提高设备故障诊断的准确率。结合这2种网络模型,并使用残差网络(residual network,ResNet)结构、批量归一化来提高模型的收敛速度以及模型泛化能力。实验表明:DCNN模型在设备故障诊断时F-Score值、准确率和召回率均优于传统卷积神经网络(convolutional neural network,CNN)、循环神经网络(recurrent neural network,RNN)、XGBoost和三比值法。DCNN模型能够对设备监测数据特征进行自主学习,减少人工干预,降低误报率;此外,基于DCNN提取的设备指纹(表征设备特征信息)为后续设备故障诊断积累了数据基础。
The most common method for transformer fault diagnosis is three-ratio method at present,but in numbers of actual applications,it is hard to further promote fault diagnosis accuracy only depending on experiences or statistic principles.Therefore,this paper proposes to construct the deeply convolutional neural network(DCNN)model to improve fault diagnosis accuracy because the DCNN is capable of identifying partial features of equipment monitoring data and relevant information between monitoring data at different time.Meanwhile,the deep neural network(DNN)model is able to approach the objective function infinitely and can be task-oriented to improve fault diagnosis accuracy.Thus,the paper combines these two models and makes use of the residual network(ResNet)structure and Batch Normalization to improve convergence speed and generalization ability of the models.The experiment indicates that the DCNN is superior to the traditional CNN,the RNN,XGBoost and three-ration method in fault diagnosis.The DCNN model can perform self-learning on characteristics of equipment monitoring data,reduce manual intervention and false positive rate.In addition,equipment fingerprints based on DCNN extraction that represents feature information of the equipment have accumulated data foundation for subsequent equipment fault diagnosis.
作者
王峰
毕建刚
万梓聪
闫丹凤
WANG Feng;BI Jiangang;WAN Zicong;YAN Danfeng(China Electric Power Research Institute Co.,Ltd.,Beijing 100192,China;State Key Laboratory of Networking and Switching,Beijing University of Posts and Telecommunications,Beijing 100876,China)
出处
《广东电力》
2019年第9期177-183,共7页
Guangdong Electric Power
基金
国家电网有限公司科技项目(GY71-17-012)
关键词
变压器
故障诊断
深度卷积神经网络
循环神经网络
XGBoost
设备指纹
transformer
fault diagnosis
deeply convolutional neural network(DCNN)
recurrent neural network(RNN)
XGBoost
equipment fingerprint
