多局放源检测分类用特征提取时频算子

Time-frequency Feature Extractors Used in Multi-PD Source Detection and Grouping

根据目前基于100 MS/s超宽带检测实现多局放源识别和干扰抑制依据的局部放电波形特征提取及分类,指出实现此系统的关键技术为脉冲群波形的快速特征提取。其一般通过对脉冲群时、频域波形进行非线性映射,形成特征参数的2D平面或3D空间,再进行无监督聚类分析即可实现局放源间和噪声源的分离。针对目前应用较多的等效时频法,提出了时频自相法和时频熵法,基于指数模型和蒙特卡洛法产生仿真脉冲群,对上述3种时频算子进行了特征参数2D平面和3D空间的可分类性评定。可分类性测度值表明,定义的2D和3D时频算子均具有较好的脉冲源可分类性,同一种时频算子3D空间的可分类性测度均优于2D平面。交、直流下局放试验数据均验证了3种特征提取时频算子的实用性。这为研制多局放检测及识别系统选择脉冲群波形特征提取方法提供了理论和试验依据。

In view of the presence of feature extraction and grouping method used in multi -partial discharge source （multi-PDs） identification and noise suppression system realized by ultra-wideband detection with sampling rate of 100 MS/s, this paper points out that one of the key technologies is the fast feature extractors used for the pulse sequence waveforms. A general way to make feature extraction resorts to nonlinear mapping of the waveforms in time and frequency（TF）domains, respectively, which are characterized by points in 2D Cartesian parameters plane or 3D Cartesian parameters space （3DCPS）, then muhi-PDs and noise separation can be realized using the unsupervised clustering analysis. According to the popular equivalent TF method（ETFM）, a TF autocorrelation method（TFACM）and a TF entropy method（TFEM）are proposed, and a separability criterion for those feature extractors is defined and ana- lyzed with pulse sequences generated with combination of exponential models and Monte Carlo simulation. The sepa rability evaluation results show those TF operators with good performance, especially in 3DCPS. The practicability of those TF operators is validated with the data from AC and DC PD tests. This study may provide experimental and theoretical basis for selecting feature extraction method in development of multi-PDs detection and identification system.