Abstract: Protein content is one of the important quality evaluation indexes for milk and dairy products. It plays an important role in the adjustment of product processing technology and feed formula. In order to improve the accuracy and stability of protein content measurement using near-infrared (NIR) spectroscopy technology, the spectral absorption of milk was first analyzed in detail. Using the two-dimensional correlation spectroscopy technique, the wavelength points related to protein absorption were obtained and the wavelength region for protein content measurement was initially determined. Then, to avoid interaction between measurement factors, the effects of detection bands, preprocessing algorithms and modeling methods on the prediction accuracy of the model were studied by single factor experiments. It showed that all three factors have a great impact on the performance of the protein model, and the factors in descending order affecting the prediction precision of model were modeling method, detection band and preprocessing method. Among the preprocessing algorithms, the SNV algorithm and the MSC algorithm can effectively eliminate the interference of scattering. According to the modeling method, the performance of linear models was obviously superior to that of nonlinear models. At last, the optimized modeling conditions were determined as that detection band ranged from 1800nm to 2300nm, the preprocessing method was MSC, and the modeling method was PCR. Under the optimized modeling conditions, the squared correlation coefficient (R2) and the RMSEP value of the protein model are 0.993 and 0.106, respectively. Compared with the single factor optimization, this research can provides a feasible technical way for parameter optimization using NIR detection technology in practice.

Key words: milk, near-infrared spectroscopy, orthogonal design, protein