一种近红外光谱定量分析软件预测性能评价方法  

An Evaluation Method of Quantitative Analysis Software for Near-Infrared Spectroscopy

作　　者：李蓉 郝璐 袁洪福 何桂梅 邓天龙 杜彪 龚丽[4] 岳欣 LI Rong;HAO Lu;YUAN Hong-fu;HE Gui-mei;DENG Tian-long;DU Biao;GONG Li;YUE Xin(College of Marine and Environmental Sciences,Tianjin University of Science&Technology,Tianjin 300457,China;Research Institute for Environmental Innovation(Binhai,Tianjin),Tianjin 300457,China;Chinese Research Academy of Environmental Sciences,Beijing 100012,China;Beijing Yixingyuan Petrochemical Technology Co.,Ltd.,Beijing 101300,China;China University of Petroleum(Beijing),Beijing 102249,China)

机构地区：[1]天津科技大学海洋与环境学院,天津300457 [2]天津市滨海新区环境创新研究院,天津300457 [3]中国环境科学研究院,北京100012 [4]北京易兴元石化技有限公司,北京101300 [5]中国石油大学(北京),北京102249

出　　处：《光谱学与光谱分析》2025年第1期213-221,共9页Spectroscopy and Spectral Analysis

基　　金：国家生态环境标准制修订计划项目(2020-60)资助。

摘　　要：常用的多元分析模型评价指标尚缺乏评价近红外分析软件多项重要预测性能指标的能力,成为近红外光谱仪预测性能以及实际近红外应用中模型适用性评价的痛点。为此,旨在发展一种近红外定量分析软件预测性能的评价方法。以近红外测定汽油烯烃浓度为研究对象,收集了192个国Ⅵ汽油样品,包括92#、95#和98#;采集其近红外光谱;按照GB/T 30519-2014测定其烯烃浓度作为参考值,分别使用两种不同的多元分析软件(1种是偏最小二乘(PLS)建模软件,另1种是非PLS的软件),建立了两个校正模型。研究发现,与参考值相比,PLS模型对低浓度样品预测值呈正偏差,高浓度的呈负偏差,即“均值化”现象。常用的模型预测性能评价指标尚不能评价模型预测值的均值化程度,也不能评价:(1)预测值与参考值偏差大于参考方法再现性的样本占比,(2)模型泛化能力。本文针对上述问题,提出了4项新评价指标包括均值化指数(AE)、预测偏差超限值样本占比(Ratio)、异常样本预测偏差(DAS)和孤立样品预测偏差(DIS)。综合常用的评价指标和新评价指标(共12项),对仪器选型的近红外光谱定量分析软件预测性能的评价、实际近红外分析应用中模型适用性的评价均具有实际意义,对近红外分析学术研究也具参考意义。The commonly used evaluation indexes of multivariate models lack the ability to evaluate many important predictive performance indicators of near-infrared quantitative analysis software.This has become a pain point in evaluating the predictive performance of near-infrared instrument selection and the applicability of models in practical near-infrared analysis applications.Therefore,this study aims to develop an evaluation method for the predictive performance of near-infrared quantitative analysis software.192 national VI gasoline samples,including 92#,95#,and 98#,were collected for determination of olefin concentration of gasoline using near-infrared spectroscopy;their near-infrared spectra collected and olefin concentrations were measured as a reference value according to GB/T 30519-2014,and two different multivariate software(one is partial least squares(PLS)modeling software,and the other is non-PLS software)were used to study.It has been found that compared to the reference value,the PLS model has a positive bias in predicting low-concentration samples and a negative bias in predicting high-concentration samples,which is known as the phenomenon of“averaging”.The commonly used performance evaluation indicators for model prediction cannot yet evaluate the degree of the averaging,nor can evaluate(1)the proportion of samples with deviation from the reference value greater than the limit value for the reproducibility of the reference method and(2)the model's generalization ability.In this paper,four new evaluation indicators are proposed to address the above issues,including Averaging Index(AE),Ratio of samples with prediction bias exceeding the limit value(Ratio),Deviation of Abnormal Sample(DAS),and Deviation of Isolated Sample(DIS).The comprehensive use of commonly used evaluation indicators and new ones(12 items)has practical significance in evaluating the predictive performance of near-infrared quantitative analysis software for instrument selection and the applicability of models in practical near-infra

关 键 词：近红外光谱 多元分析模型评价 近红外分析软件评价 汽油烯烃浓度 

分 类 号：TE622.11[石油与天然气工程—油气加工工程]