Electrical insulation faults produce partial discharges (PD), which can be analysed to identify specific types of defects. PD clustering is a widely used method to identify PD sources, although its success depends largely on the feature maps used. In this paper, three widely used feature maps, or separation maps, are compared: chromatic, energy wavelet with principal component analysis (EW-PCA), and time–frequency (TF). To compare and evaluate, five scenarios with multi-PD environments with noise were developed. The clustering ability of the maps was evaluated using two performance indicators: intercluster distance and intracluster distance. The results indicate that the EW-PCA map performed the best in all scenarios, correctly identifying the largest number of data points and producing the clearest and most distinct clusters. The TF map created distinct clusters in several scenarios, but not all. The chromatic map created distinct clusters in all scenarios but was not as well defined as the other two separation maps. Given the results, it is important in fieldwork to use a wide range of PD clustering, accompanied by performance metrics that support a less biased decision tailored to the test object.

中文翻译：

---

用于多次局部放电分类的分离图：侧重于时间和频率特性的比较研究

电气绝缘故障会产生局部放电 （PD），可以对其进行分析以识别特定类型的缺陷。PD 聚类是一种广泛使用的识别 PD 源的方法，尽管它的成功在很大程度上取决于所使用的特征图。在本文中，比较了三种广泛使用的特征图或分离图：色度、带主成分分析的小波能量波 （EW-PCA） 和时频 （TF）。为了进行比较和评估，开发了 5 种具有噪声的多 PD 环境的情景。使用两个性能指标评估地图的聚类能力：集群间距离和集群内距离。结果表明，EW-PCA 地图在所有场景中都表现最佳，正确识别了最多的数据点并产生了最清晰和最明显的集群。TF 地图在多个场景中创建了不同的集群，但不是全部。色图在所有情况下都创建了不同的聚类，但不像其他两个分离图那样定义得好。鉴于结果，在现场工作中使用广泛的 PD 聚类以及支持为测试对象量身定制的较小偏差决策的性能指标非常重要。