Determination Of Adulteration In Raw Milk Using Pattern Recognition Technology Base On Near Infrared Spectroscopy

Near Infrared Reflectance Spectroscopy (NIRS) have merits of rapid, clean and low cost, which plays an important part in rapid multi-component determination on line and quality controlling. This paper presented the advantages of applications of NIRS with chemometrics methods used in rapid determination and pattern recognition.The research emphasized on chemometrics and NIRS used in determination of contents of raw milk, differentiating pasteurized milk and reconstituted milk, raw milk and adulterated milk.The Partial least squares (PLS) prediction model for determinating Fat, Protein, Lactose, Total Solid and Solid Not-Fat in raw milk established by NIRS combined with chemometrics methods. Correlation Coefficients of the five nutrition components were 0.987, 0.982, 0.978, 0.991 and 0.992 respectively; While Root Mean Square Error of Cross Validation (RMSECV) was respectively 0.053, 0.143, 0.446, 0.111 and 0.125. By paired samples test, the results of prediction were compared with those of Reference data with no significant difference. Reliability of the model was verified by cross-validation and external-validation with good repeatability.PLS and Artificial Neural Networks (ANN) prediction model for Fat, Protein, Lactose and Total Solid in raw milk had been established with good veracity and recurrence. 14 peak value data from 3 principal components straight ahead compressed from original data by PLS, which were taken as inputs of Back-Propagation (BP)Network, while 4 predictive targets as outputs. According to Kolmogorov theorem and experiment, 29 nerve cells were taken as hidden nodes. Its training steps were 300. Predictive correlation coefficient (R~2) of the four components by the model was 0.961, 0.974, 0.951 and 0.997 respectively. The result showed an idea for multi-component quantitative analysis of other samples.This paper introduces an application for NIRS multi-component quantitative analysis by building a kind of Recurrent Network (Elman) model. Elman prediction model for fat, protein and lactose had been established with good veracity and recurrence. 14 peak value data from 3 principal components straight ahead compressed from original data by PLS, which were taken as inputs of Elman while 3 predictive targets as outputs. 53 nerve cells were taken as hidden nodes with the lowest error compared with 50 and 57. Predictive correlation coefficient (R2) of three components by the model were 0.985,0.950,0.967. The results showed that Elman using in NIRS is a rapid, effective means for measuring multi-component in other samples.This paper also studied on the construction of discrimination analysis model for differentiating pasteurized milk and reconstituted milk, raw milk and adulterated milk added in vegetable cream, whey powder, powdered oil using NIRS and Principle Component Analysis(PCA)- Mahalanobis Distance. With the spectra of samples, the original spectra data were processes by pretreated methods and compressed by PCA, the most optimum factors of each model were 8, 3, 6 and 5. The results were important in dairy quality controlling and evaluating in dairy industry.This paper showed a new method to identify pasteurized milk and reconstituted milk, raw milk and adulterated milk added in vegetable cream based on NIRS. 113 and 21 peak value data from principal components straight ahead compressed from original data by PLS, which were taken as inputs of Self-organizing Competitive network. Learning rates of each model were selected. Predictive samples provided a favorable recognition result of differentiating two classes using Self-organizing Competitive network.We Proposed a pattern recognition model for differentiating pasteurized milk and reconstituted milk, raw milk and adulterated milk added in dextrin, whey powder, powdered oil simultaneously with good veracity. 113 and 326 peak value data from principal components straight ahead compressed from original data by PLS, which were taken as inputs of Self-organizing Map (SOM) network. The competition layer structure were selected as [60 40] and [100 80]. After the network's training, the correct rate for recognition of two classes was 100%.