近红外光谱结合非线性模式识别方法进行牛奶中掺假物质的判别

Identification of Adulterants in Adulterated Milks by Near Infrared Spectroscopy Combined with Non-Linear Pattern Recognition Methods

以287例上海及上海周边地区牧场的生鲜奶作为真奶样本集组成3个真奶样品集合,配制了526例含有糊精(或淀粉)+三聚氰胺(或尿素、或硝酸铵)的掺假牛奶形成6个不同种类的假奶样品集合,其中糊精、淀粉在掺假奶中的含量为0.15%~0.45%;硝酸铵、尿素和三聚氰胺的含量分别为700~2 100,524~1 572与365.5~1 096.5mg·kg-1,以保证掺假奶中凯氏定氮法测得的蛋白含量不低于3%。所有样本的近红外光谱均经过标准正态变换(SNV)预处理。将3个真奶样品集合和6个假奶样品集合进行不同的组合并对其采用改进与简化的K最邻近结点算法(IS-KNN)和改进与简化的支持向量机法(ν-SVM)建立了判别糊精、淀粉、三聚氰胺、尿素、硝酸铵这5类掺假物质的近红外判别模型,探寻掺假物质的浓度与识别正确率之间的关系。结果表明IS-KNN和ν-SVM两种方法对含三聚氰胺、尿素、硝酸铵的掺假牛奶的平均判别正确率分别在49.55%~51.01%,61.78%~68.79%与68.25%~73.51%区间波动,说明在该研究的掺假物浓度范围内,很难用近红外模型良好区分不同类型伪蛋白的掺假奶;IS-KNN和ν-SVM两种方法对含淀粉的掺假牛奶的判别正确率分别为92.33%与93.66%、对含糊精的掺假牛奶的平均判别正确率分别为77.29%与85.08%。从整体结果上来看ν-SVM法进行建模判别的结果大部分优于IS-KNN法进行建模判别的结果。对判别正确率与样品中掺假物质的含量水平分析表明近红外光谱结合非线性模式识别方法能良好地区分掺假奶中含量较高(0.15%~0.45%)的糊精和淀粉,而对含量偏低的三聚氰胺等伪蛋白的判别效果不佳,说明近红外光谱技术不适于鉴别牛奶中含量低于0.1%的掺假物质。

In the present work ,two hundred and eighty seven raw milks collected from pastures in Shanghai and surrounding ar-eas of Shanghai were used as true milk samples and divided into three true milk sets .Five hundred and twenty six adulterated milk samples ,which contained dextrin (or starch) mixed with melamine (or urea ,or ammonium nitrate) ,were prepared as six different adulterated milk sets .The concentrations of these adulterants in the adulterated milks were designed to be 0.15% ~0.45% (starch or dextrin) ,700~2 100 mg · kg -1 (ammonium nitrate) ,524~1 572 mg · kg -1 (urea) ,and 365.5~1 096.5 mg · kg -1 (melamine) to guarantee the protein content of adulterated milks detected by Kjeldahl method not lower than 3% .All the near infrared spectra (NIR) of the samples should have a pretreatment of normal variable transformation (SNV) before they were used to build discriminating models .The three true milk sets and six adulterated milk sets were combined in different ways in order to build NIR models for discriminating different kinds of adulterants (i .e .,dextrin ,starch ,melamine ,urea and ammo-nium nitrate) based on simplified K-nearest neighbor classification algorithm (IS-KNN) and an improved and simplified of sup-port vector machine (ν-SVM ) method .The relationship between mass concentration of the adulterants and the rate of correct discrimination was also investigated .The results show that the average discrimination accuracy of IS-KNN andν-SVM for identi-fying melamine ,urea and ammonium nitrate were in the region of 49.55% to 51.01% ,61.78% to 68.79% and 68.25% to 73.51% ,respectively .Therefore within the concentration regions designed in this study ,it is difficult to distinguish different kinds of pseudo proteins by NIR spectroscopy .However ,the average accuracy of IS-KNN andν-SVM for identifying starch and dextrin are 92.33% and 93.66% ,77.29% and 85.08% ,respectively .Most discrimination results of ν-SVM are better than those of IS-KNN .The correlative analysis between the discrimination accuracy rate and the content levels of the adulterants indi-cated that near infrared spectroscopy combined with non-linear pattern recognition methods can distinguish dextrin and starch in milks with higher concentration levels (>0.15% ) ,but do not work well on identifying the adulterants with lower concentrations such as melamine (365.5 to 1 096.5 mg · kg -1 ) ,urea (524 to 1 572 mg · kg -1 ) ,ammonium nitrate (700 to 2 100 mg · kg -1 ) . Therefore near Infrared Spectroscopy is not suitable for identifying the adulterants with concentrations are below 0.1% .