基于拉曼光谱的变压器混合故障特征气体的改进卷积神经网络定量方法  

Improved Convolutional Neural Network Quantification of Mixed Fault Characterization Gases in Transformers Based on Raman Spectroscopy

作　　者：陈新岗 张文轩 马志鹏 张知先 万福[3] 敖怡 曾慧敏 CHEN Xin-gang;ZHANG Wen-xuan;MA Zhi-peng;ZHANG Zhi-xian;WAN Fu;AO Yi;ZENG Hui-min(School of Electrical and Electronic Engineering,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 University),Chongqing 400044,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400044

出　　处：《光谱学与光谱分析》2025年第4期932-940,共9页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(62275035);重庆市教育委员会科学技术研究项目(KJZD-K202101103);重庆理工大学研究生教育高质量发展项目(gzlcx20233150)资助。

摘　　要：激光拉曼光谱技术在变压器故障特征气体检测方面具有明显优势,随变压器状态监测智能化的发展,研究混合故障特征气体的快速、准确定量分析方法具有重要意义。传统拉曼光谱分析需要预处理过程,极大程度依赖人为经验,光谱特征提取虽可降低信号维度,但也会造成其特征部分缺失或改变。针对上述问题,提出基于改进一维卷积神经网络与最小二乘支持向量回归相融合的拉曼光谱定量分析方法,即引入全局均值池化与最小二乘支持向量回归改进传统卷积神经网络,并运用Dropout方法提高模型泛化性能,防止过拟合。设计并搭建变压器故障特征气体拉曼光谱检测平台,采集7种故障特征气体及N_(2)、O_(2)混合气体的拉曼信号,在谱图2900 cm^(-1)频移附近,CH_(4)、C_(2)H_(6)气体呈现谱峰重叠,且变压器过热或局部放电故障发生时,会产生主要故障特征气体CH_(4),选择不同含量比例下的CH_(4)、C_(2)H_(6)混合气体作为研究对象具有代表性,按不同比例配制146组不同含量的CH_(4)、C_(2)H_(6)混合气体样本,检测时选用氮气作为标气,采集不同含量比例下混合气体样本的拉曼光谱数据,利用光谱数据增强方法,构建适用于深度神经网络的气体样本数据集。通过不断实验,优化网络结构参数与网络权重,完成模型训练并测试其预测效果,与多种定量模型进行对比分析,并研究光谱预处理对不同定量模型的影响,进而评估模型性能。结果表明,使用原始数据集建模时,改进卷积神经网络模型的预测精确度与回归拟合优度最佳,决定系数可达0.9998,均方根误差仅为0.0005 MPa;使用预处理后数据集建模时,改进卷积神经网络模型均方根误差为0.0023 MPa,相比使用原始数据集建模误差上升了0.0018,而传统方法误差均有所下降。该研究结果表明,所提方法与传统拉曼光谱定量方法相比,集成光谱预处理、特征提�Laser Raman spectroscopy has obvious advantages in detecting transformer fault characteristic gases.With the development of intelligent transformer condition monitoring,it is of great significance to study the fast and accurate quantitative analysis method of mixed fault characteristic gases.Conventional Raman spectral analysis requires a preprocessing process that greatly relies on human experience and spectral feature extraction.Although it can reduce the signal dimensions,it can also result in partially missing or altered spectral features.Aiming at the above problems,a method for quantitative analysis of Raman spectra based on the fusion of improved 1DCNN and LSSVR is proposed;that is,the introduction of global mean pooling and least squares support vector regression improves traditional CNN,and the use of the Dropout method to improve model generalization performance and prevent over-fitting.Design and build the transformer fault characteristic gas Raman spectroscopy detection platform,collect the Raman signal of 7 kinds of fault characteristic gases and N_(2),O_(2)mixed gases,in the spectrogram near 2900 cm^(-1)frequency shift,CH_(4),C_(2)H_(6)gases show the overlap of the spectral peaks,and when transformer overheating or partial discharge fault occurs,it will produce the main fault characteristic gas CH_(4),choose different content ratio of CH_(4),C_(2)H_(6)mixed gas as a representative research object,146 groups of mixed gas samples with different contents of CH_(4)and C_(2)H_(6)are prepared according to different ratios.Nitrogen is chosen as the standard gas for detection,the Raman spectral data of the mixed gas samples with different content ratios are collected,and the spectral data enhancement method is utilized to construct the gas sample dataset suitable for deep neural networks.Through continuous experiments,we optimize the network structure parameters and network weights,complete the model training and test its prediction effect,compare and analyze with multiple quantitative models,study the effe

关 键 词：变压器 特征气体 拉曼光谱 改进一维卷积神经网络 定量分析 

分 类 号：O433.4[机械工程—光学工程]