Study Of Identification Methods For Adulterated Milk Based On Two-dimensional Correlation Near-infrared Spectroscopy

The safety of dairy products is a global problem, while the traditional test meansof dairy products can’t meet the needs of production quality control and safety control.With a lot of advantages such as efficient, fast, convenient, without pre-treatment, etc.Near-infrared(NIR) spectroscopy makes up the deficiencies of the traditionaldetection methods and was applied to the field of milk testing. Two-dimensional(2D)correlation spectroscopy has higher resolution than the conventional one-dimensionalspectroscopy, so NIR spectroscopy combined with2D correlation techniques and avariety of pattern recognition methods are applied to discriminant analysis foradulterated milk in this thesis.The range of wave number was fixed to4200~4800cm-1through analyzing theone-dimensional NIR spectrum of adulterated milk(urea, melamine).2D correlationsynchronous spectrum was chosen after compared with2D correlation asynchronousspectrum.2D correlation synchronous spectrum combined with pattern recognition wasused for discriminant analysis of adulterated milk. Mainly two methods,(1)Unfold-PLS,2D correlation spectrum matrix of each sample was unfolded alongthe direction of the wave number and combined with ordinary Partial LeastSquares(PLS)for discriminant analysis.(2)K-OPLS,2D correlation spectrummatrix(three-dimensional data) combined with K-OPLS were used for discriminantanalysis. The results showed that these two discriminant models could be effective formilk adulterated or not.In order to improve the efficiency of modeling, feature information of the hugeamount of the2D correlation spectrum data was needed. Mainly two methods,(1)Fivestatistical parameters(mean, standard deviation, center of gravity, skewness, andkurtosis) were extracted from2D correlation synchronous, then Partial Least SquaresDiscriminant Analysis(PLS-DA), Back Propagation(BP) neural network andProbabilistic Neural Networks(PNN) were used to establish discriminant modelsrespectively.(2)The first principal component’s feature vector was extracted from the2D correlation synchronous spectrum of the adulterated and the pure which used thesingular value decomposition, and then discriminant model was established by PNN.The results showed that these methods can extract effectively feature information of adulterant， reduce the input variables for modeling and made a good result fordiscriminant of the adulterated and the pure.In this study, the method that2D correlation synchronous spectrum combinedwith pattern recognition achieved the goal of milk testing and provided theoreticfoundation for the application of NIR spectroscopy in food detection.