IMPROVEMENT OF PLS MODEL TRANSFERABILITY BY ROBUST WAVELENGTH SELECTION

Frequently, a calibration model is adapted after being transferred to another instrument by, e.g., direct standardization (DS) or piecewise direct standardization (PDS). For this, a subset from the calibration set should be measured on both instruments. Usually, however, the cali bration samples cannot be measured on both instruments. Another approa ch is to make the model robust with respect to transfer to another ins trument during the development of the model, by data preprocessing. In this paper, the robustness of the calibration model is enhanced by us ing variable selection as data preprocessing. In the case under consid eration, variable selection consists in the calculation of a calibrati on model with a subset of the original wavelengths (of spectroscopic d ata) that retains its predictive ability when it is transferred to ano ther instrument. Both approaches (variable selection and (P)DS) are ap plied to the transferability of a PLS model which determines the water content in tablets. To this end, 140 tablets were measured on 2 near- infrared (NIR) reflectance instruments. It has been found that variabl e selection by simulated annealing (SA) enhances the model's robustnes s with respect to model transfer and also improves its ability. (C) 19 98 Elsevier Science B.V. All rights reserved.