潜变量聚类分析法在近红外光谱波长范围选择中的应用研究

介绍了潜变量聚类分析方法的基本原理,并将该方法应用于近红外光谱定量模型的谱区选择。以烟草样品为例,对107个样品的光谱进行处理,将光谱分为5簇,从化学角度分别解释了这5簇各自反映的信息。在此基础上,选择相应的波长范围用PLS方法建立了总糖、还原糖和尼古丁的定量分析模型。与全谱模型相比,3个模型的交互验证相关系数(Rtraining)分别由0.977 1,0.917 2,0.987 4提高到0.995 5,0.975 1,0.994 4;验证样品相关系数(Rtest)由0.977 8,0.941 2,0.993 2提高到0.992 7,0.967 9,0.994 0;交互验证均方差(RMSECV)由1.09,1.43,0.14降为1.05,1.05,0.13;预测残差均方差(RM-SEP)由0.92,1.17,0.16降为0.39,0.63,0.11;预测样品间平均标准误差(D)由1.274%,1.972%,0.829%降为0.711%,0.843%,0.768%,表明用该方法建立模型的预测准确度和精密度均有所提高,对实际应用有一定的指导作用。

Applied Study on Clustering of Variables around Latent Components Method in Wavelength Region Selection with Near-Infrared Spectroscopy

The present paper introduced the principle of clustering of variables around latent components method,and used this method in selecting spectrum range of the NIR quantitative analysis models.Taking tobacco samples as experiment materials,we dealed with 107 sample spectra,divided the spectra into 5 clusters,and explained the information reflected by each of these 5 clusters in terms of chemistry.On this basis,we chose the corresponding wavelength range to set up the quantitative models of the total sugar,reducing sugar and nicotine by PLS method.Compared with the model based on the full NIR spectral range,Rtraining of the models based on the chosen spectral range rose from 0.977 1,0.917 2 and 0.987 4 to 0.995 5,0.975 1 and 0.994 4;Rtest rose from 0.977 8,0.941 2 and 0.993 2 to 0.992 7,0.967 9 and 0.994 0;RMSECV dropped from 1.09,1.43,0.14 to 1.05,1.05 and 0.13,RMSEP dropped from 0.92,1.17 and 0.16 to 0.39, 0.63 and 0.11 and the D value dropped from 1.274%,1.972% and 0.829% to 0.711%,0.843% and 0.768% for the total sugar,reducing sugar and nicotine,respectively.These data indicated that this method can improve the forecasting precision and stability of the model,so offers certain guidance on practical application.