基于层次聚类分析的变压器油中溶解气体在线监测数据异常检测

Anomaly Detection of Online Monitoring Data of Dissolved Gases in Transformer Oil Based on Hierarchical Cluster Analysis

油中溶解气体分析是电力变压器常用的状态检测手段,在变压器运行与维护中发挥了显著作用,但因现有油中溶解气体在线监测系统可靠性问题及现场干扰,监测数据中充斥着大量的伪数据及错数据,易引起误判。针对此问题,监测系统常采用基于数据分布统计的阈值法判定数据真伪,因数据分布规律预先难以掌握,造成异常数据检出率普遍低下问题。文中依据油中溶解气体在线监测数据时间序列特点,提出了一种基于凝聚层次聚类的异常数据检测方法。首先,利用滑动时间窗对多种油中气体监测数据进行预处理,获得监测数据时间序列集,接着通过综合应用均值、阈值、标准差、小波变换周期性分量等指标对其进行分类建立异常类型的典型时序图谱;在此基础上,利用凝聚层次聚类模型,对不同特征的数据点与典型异常图谱的距离进行相似性聚类,以确定监测数据的异常类型。通过实际监测数据应用验证表明,该方法可实时检测在线监测数据流中数据异常并确定其类型,简单且易操作,具有较好的理论及应用价值。

Analysis of dissolved gases in oil is a common condition detection means for power transformer and it has played a significant role in operation and maintenance of power transformer.However,the due to the reliability of on⁃line monitoring system of the dissolved gases in the oil as well as site interference,there are plenty of false and wrong data in the monitoring data,which always leads to misjudgment.In response to this problem,the monitoring system often uses a threshold method based on data distribution statistics to determine the authenticity of the data.Since it is difficult to grasp the rules of the data distribution in advance,resulting in the problem of generally low detection rate of abnormal data.In this paper,according to the time series characteristics of on⁃line monitoring data of dissolved gas in the oil,a kind of an abnormal data detection method based on condensed hierarchical clustering is proposed.First,the sliding time window is used to preprocess a variety of gas monitoring data in the oil to obtain a time series set of monitoring data.Then,such index as average,threshold,standard deviation,wavelet transform periodic compo⁃nents are comprehensively applied to classify them and establish abnormal types of typical time series map.On this basis,the agglomerated hierarchical clustering model is used to perform similarity clustering on the distance be⁃tween different characteristic data points and the typical abnormal map so to determine the abnormal type of the moni⁃toring data.The verification of the application of actual monitoring data shows that this method can detect abnormal data in the online monitoring data stream and determine its type in real time.It is simple and easy to operate,and has good theoretical and application value.