Quantitative Comparison of Correction Methods in Composition Concentration Measurement of Milk Using Near-Infrared Spectroscopy

To establish the analytical calibration models with strong prediction ability and robust performance, it is very important to eliminate the interference information in diary product measurement using near-infrared (NIR) spectroscopy techniques. In this paper, five correction methods including extended multiplicative scatter correction, standard normal variate, orthogonal signal correction, similar background correction and wavelet transform were used to remove the interferences of irrelevant information during the process of composition concentration measurement in milk. Then, coupled with partial least squares (PLS) regression, the performances of these methods were studied, respectively. In order to further eliminate the effects of interference information, the combination of these preprocessing methods were applied and the improvements on prediction precision of fat and protein were also investigated in detail. Experimental results show that the prediction precision for fat and protein with the combinatory correction method can be improved by at least 31.6% and 30.0% compared with that using only one method, which indicate that the combinatory correction methods are more efficient and can achieve better performance. However, the combination order of these correction methods needs further study and optimization.