Speaker
摘要
摘要：为实现电力设备绝缘状态的非接触式可靠监测，本文提出了一种基于光学信号局部放电（PD）统计特性的故障诊断方法 。研究在充有 0.2 MPa SF6 气体的针型 GIL 实验腔体中，模拟了针尖缺陷、悬浮缺陷和沿面缺陷三种典型局部放电模型。采用高增益、抗磁场干扰的硅光电倍增管（SiPM）采集 PD 产生的光辐射信号 ，并通过最大值池化（Max-Pooling）算法进行降采样预处理以保留脉冲瞬态特征。本文提取了基于 PRPD 谱图的21 维统计特征向量，包括由最大放电量分布构成的特征集Mx和由平均放电幅值（放电频次）分布构成的特征集Mn。通过引入偏斜度Sk、陡峭度Ku及修正互相关系数MCC等算子定量描述图谱形状。最后，构建了基于BP神经网络、SVM及随机森林的决策层融合（DLF）诊断框架。实验结果表明，该决策层融合模型在测试集上的准确率达到 90.56%，显著优于单一分类模型，为气体绝缘设备的运行维护提供了有效参考 。
Abstract
Abstract. To achieve non-contact and reliable monitoring of insulation status in power equipment, this paper proposes a fault diagnosis method based on the statistical characteristics of Partial Discharge (PD) optical signals. Three typical PD models—needle, floating, and surface defects—were simulated in a needle-type GIL experimental chamber filled with SF6 gas. A high-gain, magnetic-interference-resistant Silicon Photomultiplier (SiPM) was employed to capture PD optical radiation signals. A Max-Pooling algorithm was used for downsampling to retain pulse transient features while reducing data redundancy. A 21-dimensional statistical feature vector was extracted based on Phase Resolved Partial Discharge (PRPD) patterns, including feature set Mx derived from the maximum discharge distribution and Mn from the pulse repetition rate distribution. Statistical operators such as skewness (Sk), kurtosis (Ku, and modified cross-correlation coefficient (MCC) were introduced to quantify pattern shapes. Finally, a Decision-Level Fusion (DLF) framework based on BP Neural Network, SVM, and Random Forest was constructed. Experimental results show that the DLF model achieves an accuracy of 90.56% on the test set, significantly outperforming single-classification models.
| 关键词 | 局部放电，光信号，决策级融合，统计特性 |
|---|---|
| Keywords | Partial Discharge,Optical Signal,Decision-Level Fusion,Statistical Characteristics |
