Reduction of Model Complexity by Orthogonalization with Respect to Non-Relevant Spectral Changes

Abstract

A method is presented to remove changes in the calibration spectra that are known to be not related to the property of interest. This can lead to multivariate calibration models that require fewer latent variables and are easier to interpret. This method requires the spectra of a sample to be measured under the different conditions that modify the spectra (for example, at different temperatures). These variations are not related to the concentration of the analyte and can therefore be removed before modeling with an orthogonalization step. The method has been used to remove the effect of temperature in the determination of NaOH in aqueous solutions by using near-infrared (NIR) spectra and partial least-squares (PLS) regression. This approach reduced the number of latent variables of the final model and made the interpretation of the PLS scores simpler.