Optimising Multivariate Calibration by Robustness Criteria

Abstract

Regular maintenance of (multivariate) near infrared (NIR) calibration models is a crucial but time-consuming step to ensure a successful NIR application in industry. Naturally, robustness of these models is essential to minimise both maintenance time and cost. In this paper, a method combining Taguchi philosophy, experimental design and artificially-derived spectra, is proposed to evaluate and improve the robustness of NIR calibrations. This approach is based upon a typical industrial NIR application, the determination of hydroxyl value of ester products. Experiments have been designed to investigate which parameters (control and signal) influence the performance of the calibration. Two calibration models have been selected for the robustness investigation. One benchmark model was based on general criteria applied for NIR calibration and another based on Taguchi's criteria. Artificially-derived spectra were produced by adding severe fluctuations of simulated wavelength shifts into original spectra for both models, then, the models' performance was evaluated six months after the calibration. The model selected based on Taguchi's criteria, is clearly more tolerant to wavelength shifts and less sensitive for overfitting in comparison with the “benchmark” model.