Rapidly Inspecting Of Liquid Dairy Products Quality By Near-infrared Spectroscopy Analysis

In the process of milk collection and production control, it is needed to determine the content of the main constituents in milk accurately and rapidly. As the content of fat and protein in milk is the key quality index, most of work on the milk quality determination placed emphasis on the measurement of the two indexes. Routine analytical methods used for milk composition measurement are destructive, expensive, time and labor consuming, and of-line by nature. In combination with the chemometrical methods, the near-infrared (NIR) spectroscopy analysis technology was used for milk and yoghurt quantitative analysis, in which the content of protein and fat are inspected. In order to provide new reference for the nondestructive examination and identification for the dairy products quality detection, several essential issues on near-infrared spectroscopy technology in building the NIR forecast modeling are investigated in terms of the spectral preprocessing methods choice, the modeling method optimize, the optimal wavenumbers selection. The main contents and conclusions are as follows:60 milk and yoghurt samples of different brands and production batchs are collected for this study, chemical value and the near-infrared spectrum of two nutrition ingredient consisting of protein and fat in milk and yoghurt samples are synchronous detected by means of near-infrared instrument of WQF-400N fourier near-infrared spectroscopy. PLS forecast models are established with the same wave number scope, same set division, and same preprocessing method, to the two targets, after using different processing methods, such as whole spectroscopy region, optimal wavenumbers selection and outlier detection, establish forecast models respectively. The correlation coefficient and root mean square prediction error of three protein forecast models are respectively 0.8551 and 0.1317, 0.9428 and 0.0714, 0.9128 and 0.1143; the correlation coefficient and root mean square prediction error of three fat models are respectively 0.826 3 and 0.273 5, 0.911 8 and 0.187 8, 0.874 3 and 0.229 6. From analyzing the results, we can know that after using some processing methods to the near-infrared spectrum, the calibration model's forecast ability is effectively improved.This article conducts the research into establishing forecast models with non-linear method--the partial least squares method union the pattern recognition method artificial nerve network (PLS+BP). And the correlation coefficient and root mean square prediction error of new protein forecast model are respectively 0.967 7 and 0.068 7; the correlation coefficient and root mean square prediction error of new fat forecast model are respectively 0.944 0 and 0.131 4.The results show us that using linear method--partial least square and non-linear method--the partial least squares method union the pattern recognition method artificial nerve network in rapidly inspecting of liquid dairy products quality by near-infrared spectroscopy analysis can get similar forecast accuracy and results, but the former is better relatively.The aim of this study was fundamental research on the measurement of liquid dairy products constituents by the near-infrared spectroscopy technique. And it laid the foundation for the further investigation.