摘要
针对换流变压器振动信号复杂度高、数据量大、信息利用率低导致基于振动信号的换流变特征提取模型搭建困难、准确度不高等问题,该文研究了一种基于类格拉姆矩阵和卷积神经网络的换流变压器振动信号多层次特征提取模型。首先通过极坐标变换和自定义点积运算将一维振动时序序列和对应细化频率序列转化为类格拉姆矩阵得到时域、频域特征图谱,通过连续小波变换将原始序列转换为时频能量特征图谱,得到振动信号的时域、频域和时频能量图谱。然后利用卷积层和池化层并行对输入图谱进行多层次特征提取融合,解决了传统方法信息利用率低的问题。利用卷积神经网络对融合矩阵进行二次特征提取。分析结果表明,该文模型振动测点分布平均识别准确率为95.4%,工况平均识别准确率为97.82%,优于长短时记忆网络(long short term memory,LSTM)、一维卷积神经网络(one-dimensional convolutional neural network,1D-CNN)、残差网络(residual network,ResNet)、全卷积网络(fully convolutional network,FCN)等经典时间序列应用网络,可为基于换流变压器振动信号的故障检测、识别提供方法基础。
Aiming at the problems of the high complexity of the vibration signal of the converter transformer,the large amount of data,and the low utilization of information,the construction of the commutation variable feature extraction model based on the vibration signal are difficult and the accuracy are not good.This paper studied a kind of Gram-like matrix and volume.Multi-level feature extraction model of converter transformer vibration signal based on product neural network.First,transform the one-dimensional vibration time series and the corresponding refined frequency sequence into a Gram-like matrix through polar coordinate transformation and custom dot product operation to obtain time-domain and frequency-domain feature maps,and convert the original sequence into time-frequency energy through continuous wavelet transformation Characteristic map,obtain the time domain,frequency domain and time-frequency energy spectrum of the vibration signal.Then,the convolutional layer and the pooling layer were used to perform multi-level feature extraction and fusion on the input map in parallel,which solved the problem of low information utilization in traditional methods.The convolutional neural network was used to extract secondary features of the fusion matrix.The analysis results show that the average recognition accuracy of the model in this paper is 97.82%,which is better than long and short-term memory network(LSTM),one-dimensional convolutional neural network(1 D-CNN),residual network(ResNet).Fully convolutional network(FCN) Classical time series application networks such as those provide a method basis for fault detection and identification based on the vibration signal of the converter transformer.
作者
张占龙
肖睿
武雍烨
蒋培榆
邓军
潘志城
ZHANG Zhanlong;XIAO Rui;WU Yongye;JIANG Peiyu;DENG Jun;PAN Zhicheng(State Key Laboratory of Transmission and Distribution Equipment and System Safety and New Technology(Chongqing University),Shapingba District,Chongqing 400000,China;Maintenance&Test Center of EHV power Transmission Company,China Southern Power Grid,Guangzhou 510000,Guangdong Province,China)
出处
《中国电机工程学报》
EI
CSCD
北大核心
2021年第20期7093-7103,共11页
Proceedings of the CSEE
基金
国家自然科学基金项目(52077012)
中国南方电网有限责任公司科技项目(CGYKJXM20190190)。
关键词
特征提取
换流变压器
振动信号
卷积神经网络
信息融合
feature extraction
converter transformer
vibration signal
convolutional neural network
information fusion
