基于高维拉曼光谱数据的变压器油纸绝缘老化评估方法研究  

Study on the Evaluation Method of Oil-Paper Insulation Aging in Transformer Based on High Dimensional Raman Spectral Data

作　　者：陈新岗 陈姝婷 杨定坤 罗浩 杨平 崔炜康 CHEN Xin-gang;CHEN Shu-ting;YANG Ding-kun;LUO Hao;YANG Ping;CUI Wei-kang(Chongqing University of Technology,Chongqing 400054,China;Chongqing Energy Internet Engineering Technology Research Center,Chongqing 400054,China;State Key Laboratory of Power Transmission Equipment&System Security and New Technology,Chongqing 400054,China)

机构地区：[1]重庆理工大学电气与电子工程学院,重庆400054 [2]重庆市能源互联网工程技术研究中心,重庆400054 [3]重庆大学输配电装备及系统安全与新技术国家重点实验室,重庆400054

出　　处：《光谱学与光谱分析》2021年第5期1463-1469,共7页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(51977017);重庆市教委科学技术研究项目(KJ1400917);重庆理工大学研究生创新项目(YCX20192057)资助。

摘　　要：采用激光拉曼光谱技术对变压器油纸绝缘老化状态检测是一种有效的方法。随着样本量的扩充,亟待处理的数据集维度逐渐增大,研究适用于高维拉曼光谱数据的变压器油纸绝缘老化评估方法具有重要的意义。设计与现场变压器内部绝缘结构相似的油纸绝缘环境,进行加速热老化实验并定期采样,获取到10类老化程度依次递增的油样本,采用激光拉曼光谱技术对样本进行检测。选用复合稀疏导数建模法对样本原始拉曼光谱数据预处理,可以一步完成去噪与基线校正;引入差异特征选取方法筛选不同老化程度下光谱中变化显著的特征,计算同一拉曼频移下不同老化程度的特征点数据集方差,选择差异较大的数据序列所对应的拉曼特征变量,设定方差阈值为0.5进行特征选择,每个样本都从1023个光谱特征点抽取出304个特征点进行后续分析;针对变压器油纸绝缘老化拉曼光谱高维样本数据集,引入多种不同类型的算法对其处理。分别运用K-means聚类算法、Fisher算法与随机森林算法对获取到的样本预处理后的数据建立模型,引入评估准确度、提升度以及Kapp a系数对各算法建立的模型判别效果进行评估。结果表明:有监督学习的Fisher算法与随机森林算法效果较好,相对于无监督学习的K-means聚类算法,模型判别能力分别提升了1.1666和1.95,论证了有监督学习模型在变压器油纸绝缘老化的评估中具有判别优势;从模型判别准确度和Kappa系数来看,强分类器随机森林算法建立的判别模型均高于F isher判别模型,其准确度提升了10%,且Kappa系数上升了0.1115,论证了随机森林算法作为由多个单一分类器组成的强分类器,相对单一分类器来说,在变压器油纸绝缘老化的评估中模型的泛化能力较好,且模型较为稳定可靠。通过对三种不同类型的算法对比,确定了在变压器油纸绝缘老化评估中,有监�Laser Raman spectroscopy is an effective method for detecting the aging state of transformer oil-paper insulation.With the expansion of sample quantity and the gradual increase of data set dimension,it is of great significance to study the evaluation method of oil-paper insulation aging in transformer suitable for high-dimensional Raman spectral data.An oil-paper insulation environment similar to the internal insulation structure of the field transformer was designed,and the accelerated thermal aging experiment was carried out and regularly sampled to obtain ten types of oil samples with increasing aging degrees,then these samples were detected using laser Raman spectroscopy.The compound sparse derivative modeling method was used to preprocess the original Raman spectral data,which can complete the noise elimination and baseline correction in one step.The differential feature selection method was introduced to screen the spectral features with significant changes under different aging degrees,and the variance of the feature point data set with different aging degrees was calculated under the same Raman shift.Furthermore,the Raman feature variable corresponding to the data sequence with a large difference was selected,and the variance threshold was set to 0.5 for feature selection,each sample selected 304 from 1023 spectral feature points for subsequent analysis.In this paper,many different types of algorithms were introduced to process the high-dimensional sample data set of transformer oil-paper insulation aging Raman spectra.For instance,the K-means clustering algorithm,the Fisher algorithm and Random Forest algorithm were used to establish a model with the preprocessed data of the obtained samples.The evaluation accuracy,lifting degree and Kappa coefficient were introduced to evaluate the discriminant effect of each mathematical model.The results show that supervised learning Fisher algorithm and Random Forest algorithm have a better effect and discriminatory advantage compared with the unsupervised learning k

关 键 词：变压器 油纸绝缘 拉曼光谱 高维数据集 老化评估 

分 类 号：O657.37[理学—分析化学]