Comparison of principal components regression, partial least squares regression, multi-block partial least squares regression, and serial partial least squares regression algorithms for the analysis of Fe in iron ore using LIBS

Abstract

The objective of the current research was to compare different data-driven multivariate statistical predictive algorithms for the quantitative analysis of Fe content in iron ore measured using Laser-Induced Breakdown Spectroscopy (LIBS). The algorithms investigated were Principal Components Regression (PCR), Partial Least Squares Regression (PLS), Multi-Block Partial Least Squares (MB-PLS), and Serial Partial Least Squares Regression (S-PLS). Particular emphasis was placed on the issues of the selection and combination of atomic spectral data available from two separate spectrometers covering 208–222 nm and 300–855 nm ranges, which include many of the spectral features of interest. Standard PLS and PCR models produced similar prediction accuracy, although in the case of PLS there were notably less latent variables in use by the model. It was further shown that MB-PLS and S-PLS algorithms which both treated available UV and VIS data blocks separately, demonstrated inferior performance in comparison with both PCR and PLS.