Modified evolving window factor analysis for process monitoring

Reaction process monitoring and control are usually involved with direct measurement or indirect model-based prediction of concentration profiles of the constituents of interest in a chemical reaction at regular time intervals. These approaches are expensive, time-consuming and sometimes impossible. On the other hand, application of so-called 'calibration-free' techniques such as EFA and EWFA to spectral data usually provides important information regarding the structural variations in the chemical system without identification of the chemical components responsible for the variations. In this paper a novel spectral data pre-processing algorithm is presented which helps EWFA to extract the concentration trends of the components of interest within the reaction. The proposed algorithm uses the pure spectrum of the component of interest to develop a so-called 'weighting filter' which is applied to the input spectral information before EWFA. The algorithm was applied to a real Raman spectral data set obtained from a pre-treatment distillation column used for removing unwanted heavy/cyclic hydrocarbons from naphtha in an oil company. Comparison of the concentration trends resulting from the proposed algorithm with those obtained using conventional PLS1 models shows that the new calibration-free and on-line algorithm outperforms the calibration models obtained by difficult and expensive laboratory work. Copyright (c) 2005 John Wiley & Sons, Ltd.