Development of robust calibration models in near infra-red spectrometric applications

Abstract

When spectral variation caused by factors different from the parameter to be predicted (e.g. external variations in temperature) is present in calibration data, a common approach is to include this variation in the calibration model. For this purpose, the calibration sample spectra measured under standard conditions and the spectra of a smaller set measured under changed conditions are combined into one dataset and a global calibration model is calculated. However, if highly non-linear effects are present in the data, it may be impossible to capture this external variation in the model. Recently, a new technique, which is based on selection of robust variables, was proposed for constructing robust calibration models. In this technique, a calibration model is developed which uses a subset of spectral values that are insensitive to external variations. This new technique is compared with global calibration models for constructing robust models in spectrometric applications. Both techniques are applied to two different near infra-red (NIR) spectroscopic applications. The first is the determination of the ethanol, water, and iso-propanol concentrations in a ternary mixture of these components; the second is the determination of the density of heavy oil products. In both applications, the calibration set spectra have been measured at a standard sample temperature and a subset has been measured at sample temperatures deviating from the standard temperature. It has been found that models based on robust variable selection are similar or sometimes better than global calibration models with respect to their predictive ability at different sample temperatures.