Feasibility Study On Discriminating Adulterated Milks By Near Infrared Spectroscopy

In this paper, hundreds of raw cow milk samples were collected from the pastures of Shanghai and neighbour areas in 2011 and 2010. A series of adulterated milk samples were prepared through adding 5%～15% adulteration solution of dextrin or starch that contained melamine, urea or ammonium nitrate to the raw milks. And the 'protein' detected by Kjeldahl method in the adulteration solutions was 3%. In order to investigate the feasibility of applying near infrared (NIR) technology to discriminating adulterated milk added different pseudo-proteins and solid contents, and furthermore to judge if the NIR technique can discriminate type of pseudo-proteins and added solid contents, various NIR pattern recognition models were established based on the different sample sets.First, the modes of testing milk samples and the methods of pretreating samples were studied and compared. It was indicated that high quality milk NIR spectra can be ensured by ultrasounding milk samples 20 min at 40℃before collecting their NIR spectrum with diffuse reflectance mode.Secondly, the pattern recognition methods of partial least squares discrimination (PLS-DA), K-nearest neighbour (KNN), improved and simplified KNN (IS-KNN) and support vector machine (SVM) were used to discriminate true and adulterated milk in this paper. The two-class classification results of true and false milks based on the four methods showed that the IS-KNN and SVM discriminating methods always gave better and more robust results.For two-class classification of true and false milks, the NIR pattern recognition models established by IS-KNN and SVM based on different sample sets can predict more than 90% samples correctly in most cases. The prediction accuracy of multi-class classification models built by the two methods for discriminating species of adulterated solids was usually higher than 80%, however, that for discriminating type of pseudo-proteins added to adulteration milks was only 40%～78.5%. The further analysis showed that the prediction accuracy of the NIR pattern recognition models was positively correlated with the concentration of the adulterated solutions and was independent of the species of pseudo-proteins and adulterated solid contents. The main reason that the established NIR modles can well discriminate adulteration milks maybe that water added to milks is well over the detection limits of NIR, so its effect on NIR spectra of adulteration milks can be reflected in the NIR models.In summary, the NIR classification models established by IS-KNN and SVM can well identify adulterated milk samples, which are added more than 5% adulteration soultions. The accuracy of the NIR models increases with the concentration of the adulterated solutions. The type of pseudo-proteins in adulterated milks could be determined by kits, chromatography etc aiming to a specific prohibited pseudo protein such as melamine if necessary.