奶粉中蛋白质的含量是决定奶粉品质的重要指标。 奶粉可以通过掺杂三聚氰胺以虚假提高奶粉中蛋白质的检测值, 严重危害了消费者的身体健康。 采用傅里叶变换中红外光谱(FTIR)联合化学计量学的模式识别方法(模型)对奶粉中的三聚氰胺进行快速鉴别, 借助模式识别技术实现了对中红外光谱客观量化的解析, 克服了谱图比对鉴别的局限性、 复杂性及主观性。 分别配置纯奶粉样品和具有不同质量浓度(0.01‰~0.2%)三聚氰胺的掺伪奶粉样品。 扫描得到样品的中红外透射光谱数据后, 首先对原始光谱数据进行归一化预处理, 然后采用包括无监督(聚类)和有监督(判别)的多种模式识别(分类)方法进行综合分析比较。 其中, 传统的主成分分析(PCA)、 距离判别法(欧式距离和皮尔逊相关系数)和非负矩阵分解(NMF)无监督模式识别方法均无法准确鉴别出纯奶粉和掺三聚氰胺的奶粉样品; 采用有监督的偏最小二乘判别分析(PLS-DA)模型识别掺伪奶粉样品时, 识别灵敏度与特异度也较低; 最后采用线性判别分析方法(LDA)和非相关线性判别分析方法(ULDA), 成功地实现了含三聚氰胺掺伪奶粉的鉴别, 识别灵敏度和特异度均达到100%。 尤其ULDA方法最大化了两类样本之间的距离, 筛选出包含最佳分类信息的特征变量, 仅用一个判别矢量便可对样本进行区分。 利用ULDA方法进行了红外光谱重要变量(特征波长)的筛选, 考察了保留变量与识别准确率的关系, 在保留较少变量数目下实现了纯奶粉与掺三聚氰胺奶粉的鉴别, 对于奶粉中三聚氰胺的定性识别浓度可低至0.01‰。 因此, 提出基于中红外光谱快捷、 准确地识别奶粉中掺入低含量三聚氰胺的模式识别模型, 相比传统的化学分析方法有优势, 为奶粉的掺伪识别与质量控制提供了有效的途径, 并可拓展应用到其他食品的真伪优劣的鉴别中。

The content of protein in milk powder is an important factor to determine the quality of milk powder. The value of protein in milk powder can be falsely increased by adulterating melamine, which endangers the health of consumers seriously. In this study, melamine in the milk powder was identified quickly by Fourier transform mid-infrared (FTIR) spectroscopy combined with pattern recognition method (model). The objective analysis of the infrared spectrum is achieved with the help of pattern recognition technology, which overcomes the limitation, complexity and subjectivity of identification through spectra comparison. Pure milk powder samples and adulterated milk powder samples with different mass concentrations (0.01‰~0.2%) of melamine were prepared respectively. After collecting the mid-infrared transmission spectrum data of the samples, the original data were firstly normalized and then comprehensively analyzed using multiple pattern recognition (classification) models including unsupervised (clustering) and supervised (discriminating) methods. Among them, traditional principal component analysis (PCA), distance discrimination (Euclidean distance and Pearson correlation coefficient) and non-negative matrix factorization (NMF) unsupervised pattern recognition methods can not accurately identify pure milk powder and melamine-containing milk powder samples. The recognition sensitivity and specificity of the supervised model of partial least squares discriminant analysis (PLS-DA) method were also low. Finally, linear discriminant analysis (LDA) and non-correlated linear discriminant analysis (ULDA) were used, the identification of melamine-containing milk powder was successfully achieved, and the sensitivity and specificity of recognition were both 100%. In particularly, the ULDA method maximizes the distance between the two types of samples, selects the feature variables containing the best classification information, and distinguishes the samples with only one discriminant vector. Furthermore, the ULDA method was used to select the important variables (characteristic wavelengths) of the infrared spectra. The relationship between the retained variables and recognition accuracy was investigated. The identification of pure milk powder and melamine-containing milk powder was realized with fewer variables. The recognition concentration of melamine in milk powder can be as low as 0.01‰. Therefore, it is proposed to identify a small amount of melamine in the milk powder with simple and rapid way based on MIR spectroscopy, which has advantages over traditional chemical analysis methods and provides an effective way for the adulteration recognition and quality control of milk powder. The method is expected to extend to apply in the identification of other foods.