Near-infrared modelling with temperature compensation based on multilevel principal component regression

To compensate the effect of temperature on predictive performance of calibration modelling, a multilevel principal component regression modelling technique is presented to improve the accuracy of regression models with temperature fluctuation. First, the multilevel simultaneous component analysis is used based on the decomposition of the spectral data into two parts, named as “between-part” and “within-part, in order to split the characteristics caused by temperature variation and concentration variation respectively. Then, the score and loading matrices of within-part are calculated to eliminate the temperature effect from spectra. Next, principal component regression model is established to represent the relationship between within-part of the spectra and concentration matrix. Finally, the theoretical results are utilized for the viscosity measurement of bisphenol-A which shows the effectiveness of the developed techniques.