基于卷积神经网络的变压器振动信号分析  

Analysis on Transformer Vibration Signal Based on Convolutional Neural Networks

作　　者：苏世玮 郭盛[1] 高伟[1] 杨涛[1] 赵家毅 SU Shiwei;GUO Sheng;GAO Wei;YANG Tao;ZHAO Jiayi(School of Energy and Power Engineering,Huazhong University of Science and Technology,Wuhan,Hubei 430074,China)

机构地区：[1]华中科技大学能源与动力工程学院,湖北武汉430074

出　　处：《广东电力》2018年第6期127-132,共6页Guangdong Electric Power

基　　金：国网湖北省电力公司科技项目(52191614004V)

摘　　要：为研究变压器振动与运行状态之间的关系,在小波分析方法的基础上,结合卷积神经网络的方法进行变压器振动信号分析。首先对1台油浸式变压器模拟绕组松动和铁心松动两种故障状态并分别测量其振动信号,然后对测试所得的振动信号进行小波变换,生成小波灰度图,并进行卷积神经网络训练分析。根据卷积神经训练的结果,该方法准确率在84.03%,说明卷积神经网络结合小波灰度图的分析方法可以有效识别振动信号中故障信息。比较2类故障验证样本中错误结果的分布情况可以发现,错误结果受变压器振动测点位置影响较大,在改善测点和增加训练数据的前提下,准确率还能有所提升。In order to study relationship between vibration and operating state of the transformer, the paper makes analysis on transformer vibration signal on the basis of wavelet analysis method combining with the convolutional neural networks. It firstly makes simulation on two failure states including winding looseness and iron core looseness of one oil-immersed trans- former and respectively measures vibration signals. Then it makes wavelet transform of measured vibration signals and cre- ates wavelet gray scale maps, and conducts the convolutional neural networks training analysis. The training results indicate accuracy of this method is 84.03%, which explains the analysis method combining the convolutional neural networks and wavelet gray scale map can effectively recognize fault information in vibration signals. After comparing distribution of error results of verification samples, it discovers error results are greatly affected by locations of measuring points. It points out in the premise of improving measuring points and increasing training data, accuracy rate can be promoted.

关 键 词：卷积神经网络 小波分析 灰度图 变压器 振动信号 

分 类 号：TM41[电气工程—电器]