Study On Rapidly Detecting The Quality Of Milk By Near -inrared Spectroscopy

Sugar content is an important parameter for assessing the internal quality of apples. Traditional methods for the main quality indexes (i.e. fat, protein and lactose) of milk determination don't apply to on-line analysis because they are time-consuming and apple-destructive. Near-infrared (NIR) spectroscopy analysis could detect the quality rapidly and nondestructively but without contamination. Recently, more and more attention has been paid to the use of NIR diffuse reflectance spectrum for milk quality indexes determination. Considering the existing shortcomings of NIR diffuse reflectance spectrum being used, this research focuses on several key techniques of determination model for milk main quality indexes in order to probe new method which may be suitable for detecting internal qualities of milk and other liquid state food. The main contents are as follow.The diffuse transmission was used to collecte the NIR spectra, according to the quality, composition of milk and the light diffuse very complicated in milk.In order to selecte the best preprocession method for milk, the precision of partial least square (PLS) models which built on the preprocession method was evaluated. The best PLS model for fat of milk that was based on multiplicative scatter correction (MSC) has 5 principal factors, and the correlation coefficient and the root mean square error of prediction (RMSEP) are 0．997 and 0．180, respectively. Meanwhile, the best preprocession methods for protein and lactose were first and second derivate, and the coefficient of the models for protein and lactose are 0．920 and 0．880 respectively, and the RMSEP of these models are 0．179 and 0．121．Although the resulting PLS model could predict the main quality indexes of milk well, it was built on the basis of 1501 wavenumber points from 3800 to 11000 cm brings about redundant information and even weakens predictive ability. Additionally, this simplified model needs enormous calculations and affects efficiency of on-line detection. A new wavelength selection methods, such as interval partial least square (iPLS), forward/backword interval partial least (FiPLS/BiPLS) were proposed and used to select wavenumber points for building PLS milk fat models from whole wavenumber range. Consequently, 2 subsets which contain 187 wavenumber points were selected, principal factors were 4 and RMSECV 0．09408 and RMSEP 0．04531． The PLS model built on selected wavenumber points shows better prediction than that of model built before wavelength selection. The other wavelength selection method, genetic algorithm combined with partial least square (GA-PLS) was also used to select wavenumber points. The PLS model after GA-PLS selection needs only 35 wavenumber points out of 1868 and 4 PLS factors and is thus more concise and better for prediction. Therefore, these methods are consequently used to simplify the protein and lactose models, and these models were more concise and better for prediction.