基于多特征声纹图谱的变压器绕组松动在线故障诊断方法  

On-line fault diagnosis method of transformer winding looseness based on multi-characteristic voiceprint maps

作　　者：马宏忠[1] 李楠 杨启帆 段大卫 朱昊 何萍[2] MA Hongzhong;LI Nan;YANG Qifan;DUAN Dawei;ZHU Hao;HE Ping(College of Energy and Electrical Engineering,Hohai University,Nanjing 211100,China;State Grid Nanjing Power Supply Company,Nanjing 210008,China)

机构地区：[1]河海大学能源与电气学院,江苏南京211100 [2]国网南京供电公司,江苏南京210008

出　　处：《电机与控制学报》2023年第5期76-87,共12页Electric Machines and Control

基　　金：国家自然科学基金(51577050);国网江苏省电力公司科技项目(J2021053)。

摘　　要：绕组松动故障是变压器安全稳定运行中的巨大隐患,目前尚缺乏有效的在线诊断方法。变压器运行产生的声音信号蕴含着大量反映设备状态的有效信息,依据声音信号的特征图谱对松动故障实现在线诊断。首先,构建4种特征图谱,包括通过格拉米角场构建时域特征图谱、通过傅里叶变化和马尔可夫变迁场构建频域特征图谱、通过小波变换构建时频域特征图谱、通过递归分析构建混沌特征图谱;然后,建立轻量化卷积神经网络模型,以4种特征图谱作为输入,通过卷积、池化等一系列操作提取有效故障特征;最后,利用分类器直接输出绕组松动的故障程度。实验结果表明,所提方法对25%、50%、75%及100%的松动程度均能实现可靠诊断,平均准确率为99.6%,对最为轻微的25%松动程度,准确率仍达98%。与仅采用单一特征的诊断相比,所提方法的准确率提升了9.9%;与采用AlexNet、MobileNetV2、GoogleNet、ShuffleNet、ResNet等经典神经网络的诊断相比,所提方法的准确率提升了18.1%,同时训练速度提高37%,占用内存减少20%。Loose winding fault is a huge hidden danger for the safe and stable operation of transformers,and there is still a lack of effective online diagnosis methods.The sound signal generated by transformer contains a large amount of effective information that reflects equipment status,thus the loose fault can be diagnosed online based on the characteristic map of sound signal.First,four kinds of characteristic maps were constructed,including the time-domain characteristic map extracted by Grammy angle field,the frequency-domain characteristic map extracted by Fourier and Markov transition fields,the time-frequency domain characteristic map extracted by wavelet transform,and the chaotic characteristics map extracted by recursive analysis.Then,a lightweight convolutional neural network model was constructed,where four kinds of characteristic maps were selected as input and the effective fault features were extracted through convolution and pool operation.Finally,the looseness degree of winding fault was directly deter-mined by classifier.Experimental results show that the proposed method can reliably diagnose the winding faults with looseness of 25%,50%,75%and 100%,and the average accuracy rate is 99.6%.Even for the moderate 25%looseness,the accuracy rate can also reach 98%.Compared with the accuracy of method using only one map,that of the proposed method is improved by 9.9%.Compared with the meth-od using classical neural networks such as AlexNet,MobileNetV2,GoogleNet,ShuffleNet,and ResNet,the proposed method has 18.1%improvement in accuracy,37%improvement in training speed,and 20%reduction in occupied memory.

关 键 词：电力变压器 声纹 卷积神经网络 多特征融合 绕组松动 

分 类 号：TM41[电气工程—电器]