可见光/近红外光谱技术快速测定橙汁柠檬酸含量

为了快速无损测定橙汁的柠檬酸含量，提出了一种用可见光/近红外光谱技术进行检测的新方法。选用高效液相色谱法作为光谱柠檬酸测定的标定方法。采用平滑点数为5的移动平滑法对原始光谱进行预处理消除噪声。由于采集的光谱数据量非常大，为了减少建模时间，建模之前采用小波变换对经过预处理的大量光谱数据进行降维压缩，并在Matlab7.01中通过自编程序实现此变换。利用光谱专用分析软件Unscrambler 9.5，对压缩后的新变量进行分析，建立偏最小二乘(PLS)校正模型。考虑到不同小波基及分解尺度对数据压缩的影响，采用预测平方和PRESS值最小的评价标准，选择最佳的小波基Db4及分解尺度5。用于本实验的样本总数为40，其中30个样本进行建模，10个用于预测。用校正集相关系数(r)和标准偏差(SEC)作为校正模型的评价指标,预测结果采用预测相关系数(r)和预测标准偏差(SEP)来评定。文章将基于小波变换的PLS模型与直接建立的PLS模型进行了比较，偏最小二乘法结合小波变换的模型预测相关系数为r=0.901, 预测标准偏差SEP=0.937；而由PLS建立的模型其预测相关系数r=0.849，预测标准偏差SEP=1.662。由此可见，由偏最小二乘法结合小波变换所得模型效果优于单独使用偏最小二乘法的结果。

Rapid Measurement of Citric Acids in Orange Juice Using Visible and Near Infrared Reflectance Spectroscopy

Visible and near infrared reflectance spectroscopy(Vis/NIRS) as a new method was proposed for the rapid and non-destructive measurement of citric acids in orange juice.High performance liquid chromatography(HPLC) was used as a reference method for the spectral analysis of citric acids.The original spectral data were preprocessed by the smoothing method with five smoothing points in order to eliminate the noise.Before modeling,large spectral data were compressed by wavelet transform(WT) in Matlab7.01 with the edited program to reduce the dimensions and modeling time,and then the new variables after being compressed were used to build PLS calibration in spectral software Unscrambler 9.5.Considering the effect of different wavelet functions and decomposed scales on the data compressed,the optimal wavelet function Db4 and decomposed scale 5 were determined by predictive residual error sum of squares(PRESS).A total of forty samples were used in our experiment,including thirty samples for the calibration model and ten unknown samples for the prediction.The quality of the calibration model was evaluated by the correlation coefficients(r) and standard error of calibration(SEC),and the prediction results were assessed by correlation coefficients(r) and standard error of prediction(SEP).Comparing WT-PLS model with PLS model,the result of WT-PLS model was r of 0.901 and SEP of 0.937,while the result of PLS model was r of 0.849 and SEP of 1.662,indicating that the prediction result from PLS model with wavelet transform was better than that from PLS model.