基于电场测量特征的开关柜绝缘状态可分性评价方法

Classification and Detection Method of Switchgear Faults Based on Electric Field Test Features

摘要:

针对扩展参数空间下开关柜绝缘状态细粒度区分困难的问题，考察电场测量特征对状态区分的支持程度。构建五类等效绝缘状态的10000个仿真样本，采用26项可追溯电场特征、固定训练/验证/测试划分和随机森林模型，结合相关性、混淆矩阵、置换重要性及分层抽样重复实验进行分析。结果表明，26特征模型测试集和验证集Macro-F1分别为0.3379和0.3479；从训练集中分层抽取2000个样本进行10次重复训练，测试集和验证集Macro-F1分别为0.2820±0.0056和0.2915±0.0110。主要混淆集中于表面污秽与内部气隙、悬浮颗粒与正常状态等类别。多维电场统计特征可提供一定状态区分信息，但扩展参数空间下存在明显特征重叠，不宜解释为现场高水平诊断能力。

Abstract:

To address the difficulty of fine-grained switchgear insulation condition discrimination in an expanded parameter space, the support provided by electric-field measurement features for distinguishing insulation conditions was examined. A total of 10000 simulation samples covering five equivalent insulation conditions were constructed. Twenty-six traceable electric-field features, a fixed training/validation/test split and a random forest model were used, together with correlation analysis, confusion matrices, permutation importance and repeated stratified-sampling experiments. The 26-feature model achieved Macro-F1 values of 0.3379 on the test set and 0.3479 on the validation set. Ten repeated training experiments using 2000 samples drawn by stratified sampling from the training set yielded Macro-F1 values of 0.2820±0.0056 and 0.2915±0.0110 on the test and validation sets, respectively. The main confusions involved surface contamination versus internal void and floating particles versus the normal condition. Multidimensional electric-field statistics provide some information for distinguishing insulation conditions, but substantial feature overlap remains in the expanded parameter space; these results should not be interpreted as high-level field diagnostic capability.