Method Research For Milk Powder Doping Discrimination And Multi-brand Classification With Vis-NIR Spectroscopy

Food safety is a major strategic issue related to national economy and human's livelihood.Since the incidents such as“melamine doping”,the safety of milk powder has received much attention.And some related detection methods have been developed.However,in order to make huge profits,there are still incidents of unscrupulous merchants mixing or counterfeiting high-end infant milk powder with cheap milk powder.Due to the similarity of sample components,the identification?doping/counterfeiting?of branded milk powder is very tedious and difficult.The discriminant analysis with Vis-NIR spectroscopy is based on the spectral similarity of samples in the same type and the spectral differences of different types of samples for spectral pattern recognition.Vis-NIR spectroscopy has the advantages of direct,rapid and efficient dection,and has the application potential in the identification of milk powder.In this paper,Vis-NIR spectroscopy combined with partial least squares-discriminant analysis?PLS-DA?was used to carry out,there were two aspects of research:1)Infant milk powder?I?was mixed with different proportions of milk powder from other brands to carry out research on the“binary classification”spectral discriminant analysis method;2)Four infant milk powder brands?I,II,III,IV?were used to carry out research on the“four classification”spectral discriminant analysis method.At the same time,the optimization method of the relevant wavelength model was studied to explore the possibility of Vis-NIR spectral discriminant analysis for milk powder identification?doping/counterfeiting?.The main content and results are as follows:1.Research on the discriminant analysis method of doped milk powder based on Vis-NIR spectroscopy:1)I brand milk powder was used as the identification sample?negative,96?.And different proportions of II and III brand milk powder were added to I brand milk powder for preparation doped samples?positive,91?.The independent sample system for calibration-prediction-validation was established to ensure the objectivity of validation.Nine recognition accuracy rates?RAR,%?and their standard deviations(RAR_SD)for each sample attribute were proposed as model evaluation indicators.The total recognition accuracy rate(RAR_Total)was used as the optimization index to determine the parameters,taking into account the balance of discrimination effect for each type.2)Based on the PLS-DA algorithm,the discrimination efficiency curve of single wavelength was determined.The wavelength selection method of wavelength efficiency priority combination?WEPC?was proposed further,which was applied to the original spectra and the standard normal variable transformation?SNV?pretreatment spectra.The number of wavelengths?N?for optimal models were 25 and 134,and the PLS latent variables?LV?both were 7.3)According to the spectral population,the separation spectrum of I,II,and the relative separation spectrum of I,II were proposed respectively.The four wavelength selection method of separation degree priority combination?SDPC?were proposed,which were applied to the original spectra and SNV spectra.The best validation results were obtained from the relative separation of II priority in SNV model?N=19,LV=10?,and three recognition accuracy rates for negative,positive,and total were all 100%.2.Research on the discriminant analysis method of four brand milk powders based on Vis-NIR spectroscopy:1)The model evaluation system of calibration-prediction-validation for discriminant analysis of 60 samples from each of the four brands of milk powder?I,II,III,IV?was established.The“one-to-other method”was used to establish four“binary classification”discriminant analysis models,which including I-non I,II-non II,III-non III,and IV-non IV;The validation samples was judged and scored four times,and the“four classification”discrimination result of the sample was determined according to the score vector.2)Based on the four SDPC methods,the PLS-DA models of I-non I,II-non II,III-non III,IV-non IV discrimination were established and compared,and the number of wavelengths?N?for optimal models were 1,3,1,3.The 72 validation samples that did not participate in the modeling were judged by“four classification”,and the total recognition accuracy rate reached 100%.The research shows that:Vis-NIR spectroscopy combined with PLS-DA method can be used for discriminant analysis of doped milk powder,but also for multi-brand milk powder classification.The four proposed methods of separation degree priority combination are novel,which reflect the classification characteristics of spectral populations.The established wavelength models have achieved a high-precision discriminant analysis effect,and can also provide a reference for specialized spectrometers design for milk powder identification.The technology is fast and simple,and has application potential in food safety of milk powder.