基于机器学习与红外光谱技术的变压器油老化行为研究  

Study on the Aging Behavior of Transformer Oil Based on Machine Learning and Infrared Spectroscopy Technology

作　　者：肖忠良[1] 袁荣耀 付壮 刘成[1] 尹碧露 肖敏之 赵亭亭 匡尹杰[1] 宋刘斌[1] XIAO Zhong-liang;YUAN Rong-yao;FU Zhuang;LIU Cheng;YIN Bi-lu;XIAO Min-zhi;ZHAO Ting-ting;KUANG Yin-jie;SONG Liu-bin(College of Chemistry and Chemical Engineering,Changsha University of Science and Technology,Changsha 410114,China)

机构地区：[1]长沙理工大学化学化工学院,湖南长沙410114

出　　处：《光谱学与光谱分析》2025年第2期434-442,共9页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(21501015,31527803,21545010);湖南省自然科学基金项目(2022JJ30604)资助。

摘　　要：为解决现阶段油品老化分析工作复杂、误差大等问题,提出一种红外光谱与机器学习(ML)相融合的技术。借助傅里叶变换中红外(FT-MIR)光谱仪采集三种变压器油在不同老化时间的样本光谱,运用多种预处理方法对样本光谱进行预处理,以自动寻峰并求得特征峰面积之和。采用偏最小二乘回归(PLSR)和粒子群优化-支持向量机回归(PSO-SVR)算法建立了变压器油老化程度定量分析模型,研究并分析了多种光谱数据预处理方法对红外光谱降噪、基线校正等处理效果以及对两种模型定量分析效果的影响。结果表明,油品光谱预处理效果最好的是平滑法,其中SG+SVR和SG+PLSR模型拟合优度(R^(2))分别为0.9814、0.9913,平均绝对误差(MAE)为0.3124、0.2880,均方根误差(RMSE)仅有0.0977、0.3790。在合适的预处理条件下,两种机器学习算法鲁棒性和可靠性均较强,模型预测值与实际值间差异极小。To solve the problems of complexity and large errors in oil aging analysis at the present stage,a technique integrating infrared spectroscopy and machine learning is proposed.With the help of a Fourier-Transform Mid-Infrared(FT-MIR)spectrometer,the sample spectra of three kinds of transformer oils were collected at different aging times.Various preprocessing methods were used to preprocess the sample spectra,and then the peaks were automatically sought and the sum of the characteristic peak areas was obtained.PLSR and PSO-SVR were used to establish a quantitative analysis model of transformer oil aging degree,and the effects of multiple spectral data preprocessing methods on the processing effects of infrared spectral noise reduction and baseline correction,as well as on the quantitative analysis effects of two models were investigated and analyzed.The results show that the best oil spectral preprocessing is the smoothing method,in which the SG+SVR and SG+PLSR model fitting Goodness-of-Fit(R^(2))are 98.14% and 99.13%,respectively,and the mean absolute error(MAE)is 0.3124 and 0.2880,and the root-mean-square error(RMSE)is only 0.0977 and 0.3790.Under the appropriate preprocessing conditions,both machine learning algorithms are robust and reliable,and the difference between the predicted and actual values of the models is extremely small.

关 键 词：机器学习 傅里叶变换中红外光谱 变压器油 老化程度 粒子群优化-支持向量机回归(PSO-SVR) 偏最小二乘回归(PLSR) 

分 类 号：O657.33[理学—分析化学]