利用近红外光谱进行婴幼儿乳粉中营养物质含量的快速检测

Fast monitoring nutritive material content in infant milk powder based on near infrared spectroscopy

婴幼儿乳粉的质量和安全日益受到人们的关注,研究了利用近红外漫反射光谱进行乳粉中各类营养物质含量的快速无损检测的适应性。以156个乳粉样本作为样本集,分别采用偏最小二乘回归(PLSR)和KNN保形映射(KNN-KSR)算法建立含量范围分别为14.5%~23.1%(蛋白质、脂肪)、(5.6~9.2)mg/g(钙)、(1.29~10.2)mg/100 g(锌、V_(B2))、(0.34~1.47)mg/g(Vc)的营养物质近红外定量分析模型。PLSR与KNN-KSR预测各类营养物质含量的平均相对误差分别小于5%与3%,2种方法对Vc的预测误差最大。不同光谱预处理方法所得结果显示一阶导预处理结果更为理想,可使每种营养物质的预测平均相对误差降低0.1%以上。将样品按照Vc浓度的量级划分为2个区间分别建立模型,平均相对误差较整个浓度区间建模结果减小1%以上。利用PLSR与KNN-KSR方法基于近红外光谱信息可快速预测乳粉中不同浓度量级的营养成分,KNN-KSR的预测效果更佳。为获得准确的预测结果,建议采用同量级浓度样品建立乳粉营养成分的近红外定量模型并分浓度区间进行应用。

The quality and safety of infant milk powder has attracted more and more attention from the public. A quick and nondestructive method using near-infrared （NIR） diffuse reflectance spectroscopy to determine the content of nutritive materials in milk powder was presented. 156 milk powder samples were analyzed. Partial least square （PLS） regression and an innovative method of keeping same relation between X and Y space based on K nearest neighbors （KNN-KSR for short） were utilized to predict the protein, fat, calcium, zinc, Vc and VB2 content in milk powder based on NIR spectra, respectively. The content distributions of these nutrient substances are greatly different： 14.5%-23.1% （protein, fat）, （5.6-9.2）mg/g （calcium）, （1.29-10.2）mg/100 g （zinc, VB2）, （0.34-1.47）mg/g （Vc）. Average prediction relative error given by PLS and KNN-KSR were less than 5% and 3%, respectively. And the results of both methods based on the first derivative spectra were best. Prediction error of Vc was the largest for both of methods. When the samples were divided into two groups to build NIR quantitative analysis models, the prediction error of concentration of Vc, and the average relative error was reduced more than 1%. Nutritive materials with the different of concentration magnitude could be detected quickly based on near-infrared spectroscopy and the prediction results using KNN-KSR were better than PLS. In order to obtain more accurate prediction results, it might be better to build and apply NIR models based on samples with the same concentration order of magnitude.