基于互信息方法的燃油品质光谱分析模型构建与简化  

Construction and Simplification of the Calibration Model for Spectral Analysis of Fuel Oil Properties Based on Mutual Information Method

作　　者：郝勇[1] 孙旭东[1] 蔡丽君[1] 刘燕德[1] 

机构地区：[1]华东交通大学机电工程学院,江西南昌330013

出　　处：《光谱学与光谱分析》2012年第1期175-178,共4页Spectroscopy and Spectral Analysis

基　　金：国家科技支撑计划(2008BAD96B04);江西省主要学科学术和技术带头人培养对象计划(2009DD00700);华东交通大学博士启动基金项目(01309021)资助

摘　　要：近红外漫反射光谱和紫外吸收光谱分别用于燃油的辛烷值和单芳香族化合物含量的测定,偏最小二乘回归(partial least squares regression,PLSR)用于光谱多元校正模型的构建。基于互信息(mutual infor-mation,MI)理论的变量筛选方法用于模型优化以提高模型的预测精度,降低模型的复杂度。结果表明,MI-PLSR可以有效的提高燃油品质模型的预测精度,简化分析模型。辛烷值的预测均方根误差(root meansquare error of prediction,RMSEP)由0.288减小为0.111,预测相关系数R从0.985提高到0.998,建模变量由401减小为112;单芳香族化合物含量的RMSEP从0.753减小为0.478,R由0.996提高为0.998,建模变量由572缩减为37。说明振动光谱结合MI-PLSR方法可用于燃油品质检测,具有高效率低成本的特点。Near infrared diffuse reflectance（NIRS） and ult raviolet（UV） spectral analysis were adopted for quantitative determination of o ctane number and monoaromatics in fuel oil.Partial least squares regression（PL SR） was used for construction of vibrational spectral calibration models.Variab les selection strategy based on mutual information（MI） theory was introduced to optimize the models for improving the precision and reducing the complexity.Th e results indicate that MI-PLSR method can effectively improve the predictive a b ility of the models and simplify them.For octane number models,the root mean s quare error of prediction（RMSEP） and the number of calibration variables were r educed from 0.288 and 401 to 0.111 and 112,respectively,and correlation coef fi cient（R） was improved from 0.985 to 0.998.For monoaromatics models,RMSE P and the number of calibration variables were reduced from 0.753 and 572 to 0.478 a nd 37,respectively,and R was improved from 0.996 to 0.998.Vibrational spe ctra l analysis combined with MI-PLSR method can be used for quantitative analysis o f fuel oil properties,and improve the cost-effectiveness.

关 键 词：近红外光谱 紫外光谱 偏最小二乘回归 互信息 燃油品质 

分 类 号：O657.33[理学—分析化学]