A comparison of variate pre-selection methods for use in partial least squares regression: A case study on NIR spectroscopy applied to monitoring beer fermentation

• Published in: Journal of food engineering 2009, Vol.90 (2), p.300-307
• Main Authors: McLeod, Georgina, Clelland, Kirsty, Tapp, Henri, Kemsley, E. Katherine, Wilson, Reginald H., Poulter, Graham, Coombs, David, Hewitt, Christopher J.
• Format: Article
• Language: English
• Subjects: absorbance; alcoholic fermentation; Beers; Biological and medical sciences; Brewing; dry cell weight; ethanol; Fermented food industries; Food engineering; Food industries; Fundamental and applied biological sciences. Psychology; General aspects; Genetic algorithm; least squares; linear models; model food systems; monitoring; near-infrared spectroscopy; NIR spectroscopy; PLS regression; pre-selection methods; prediction; specific gravity; Variate selection
• ISSN: 0260-8774; 1873-5770
• DOI: 10.1016/j.jfoodeng.2008.06.037

This work investigates four methods of selecting variates from near-infrared (NIR) spectra for use in partial least squares (PLS) regression models to predict biomass and chemical changes during beer fermentation. The fermentation parameters studied were ethanol concentration, specific gravity (SG), optical density (OD) and dry cell weight (DCW). The four selection methods investigated were: Simple, where a fingerprint region is chosen manually; CovProc, a covariance procedure where variates are introduced based on the magnitude of the first PLS vector coefficients; CovProc–SavGo, a modification to CovProc where the window size of a Savitzky-Golay filter applied to the spectra is also optimised; and genetic algorithm (GA), where variates are selected based on the frequency of appearance in 8-variate multiple linear regression models found from repeated execution of the GA routine. The analysis found that all four methods produced good predictive models. The GA approach produced the lowest standard error in prediction (SEP) based on leave-one-out cross-validation (LOO-CV), although this advantage was not reflected in the standard error in validation values, SEV, where all four models performed comparably. From this work, we would recommend using the Simple approach if a suitable fingerprint region can be identified, and using CovProc otherwise.