Monitoring production processes of pharmaceutical and farm chemical with near-infrared spectra coupled with interaction moving window partial least squares

Abstract

On-line near infrared spectroscopy is utilized to monitor production processes of pharmaceutical and farm chemical. To improve the model performance, interval interaction moving window partial least squares (iMWPLS) is used for wavelength selection. As a modified version of moving window partial least squares (MWPLS), iMWPLS may accurately screen intervals out and optimize combination of the spectral intervals by iteratively accomplishing MWPLS with the selected intervals. Two on-line NIR datasets generated during the processes of aspirin synthesis and Dichloroacetophenone production are applied for multivariate calibration with partial least squares (PLS) and wavelength selection with iMWPLS. Results show that partial least squares models built with wavelengths selected by iMWPLS exhibit significant improvement with low root mean squared error of calibration and prediction, comparing with models established using the original and preprocessed spectra, as well as using selected wavelength spectra by MWPLS. This study has reflected that iMWPLS has ability to improve model performance by accurately screening spectral intervals to construct an optimized combination of the intervals.