Abstract: In order to improve the accuracy and stability of protein content measurement using near-infrared (NIR) spectroscopy, the spectral data of milk samples with different protein contents were analyzed using the two-dimensional correlation spectroscopy to identify the characteristic wavelengths region of protein. Then, the effects of wavelength bands, preprocessing algorithms and modeling methods on the prediction accuracy of the model were studied by single factor experiments. On this basis, the modeling conditions were optimized by orthogonal array design to avoid interactions. The results showed that all three factors had a great impact on the performance of the prediction model in the descending order of modeling methods, wavelength bands and preprocessing methods. Among the preprocessing algorithms tested, standard normal variable (SNV) algorithm and multiplicative scatter correction (MSC) algorithm could effectively eliminate the interference of scattering. The linear models such as principal component regression (PCR) and partial least squares (PLS) were significantly better than non-linear support vector machine regression (SVR). Finally, the optimized modeling conditions were determined as follows: detection wavelength range from 1 800 to 2 300 nm, MSC preprocessing, and PCR modeling. Under these conditions, the correlation coefficient (R2) and root mean square error of prediction (RMSEP) of the prediction model were 0.993 and 0.106, respectively. This research provides a feasible technical way to develop a new device for the rapid detection of protein content in milk in the future.

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