Comparison of Five Svd-Based Algorithms for Calibration of Spectrophotometric Analyzers

Abstract

Abstract Spectrophotometry is an analytical technique of increasing importance for the food industry, applied i.a. in the quantitative assessment of the composition of mixtures. Since the absorbance data acquired by means of a spectrophotometer are highly correlated, the problem of calibration of a spectrophotometric analyzer is, as a rule, numerically ill-conditioned, and advanced data-processing methods must be frequently applied to attain an acceptable level of measurement uncertainty. This paper contains a description of four algorithms for calibration of spectrophotometric analyzers, based on the singular value decomposition (SVD) of matrices, as well as the results of their comparison - in terms of measurement uncertainty and computational complexity - with a reference algorithm based on the estimator of ordinary least squares. The comparison is carried out using an extensive collection of semi-synthetic data representative of trinary mixtures of edible oils. The results of that comparison show the superiority of an algorithm of calibration based on the truncated SVD combined with a signal-to-noise ratio used as a criterion for the selection of regularisation parameters - with respect to other SVD-based algorithms of calibration.