概率聚类技术应用于变压器DGA数据故障诊断

传统的最优聚类、分类技术,需要对聚类原型做球形假设,若将其应用于溶解气体分析(DGA)数据表诊断故障分类问题将存在不符合聚类本质的问题。为此将密度聚类方法引入DGA数据的故障诊断,取消了对聚类原型做形状假设,实现了DGA样本聚类的无监督型分析。该方法实现如下:①利用非参数密度估计方法估计样本空间概率密度函数,并以概率密度函数作为聚类依据,密度函数值较大的区域将有可能作为类簇原型区;②利用非参数估计方法直接估计出概率密度函数的梯度场;③依据概率密度函数的梯度分布确定聚类原型,进而利用峡谷搜索法思想建立聚类划分;④最后利用类簇划分的边界确定变压器故障的区分边界。试验结果表明,该方法实现了基于密度的自然值域划分,能够做到比现有的人工划分方式更加细致地划分,为研究DGA样本表特性提供了一种新的可行途径。

Application of Density-based Clustering Technology in Diagnosis of DGA Data of Transformer

Advantages of a novel method named density-based clustering is used to realize the non supervised analysis of clustering of samples of DGA.The main steps are as follows:Firstly,by taking advantages of non-parametric density estimation,possibility density function(PDF) is determined,which is taken as the index of clustering,where areas with larger value of PDF would be taken as clustering prototypes;Secondly,the gradient field of PDF is estimated by means of non-parametric estimation;Thirdly,the prototypes of clusters are determined by virtue of gradient characteristic of PDF so that the partition of different clusters is completed based on valley seeking.Finally,the boundary of the fault in diagnosis table could be determined according to the boundary of clusters.The test results show that the proposed method realizes the partition based on natural essence of density of samples and yields a novel alternative for researches of characteristics of DGA samples.