Comparative study of linear and nonlinear calibration algorithm for extrapolation ability of near infrared spectroscopy quantitative analysis

Abstract

The determination of the o-nitrotoluene (o-MNT) content in separation process of mononitrotoluene (MNT) is of interest, since it affects the purity of m-nitrotoluene (m-MNT) and p-nitrotoluene (p-MNT). In real-world applications, the calibration model inevitably requires dealing with complex extrapolation problems. Therefore, this study extracted the spectral features of the o-nitrotoluene based on the interval selection algorithm. The linear calibration method (partial least squares (PLS)) and nonlinear calibration methods (support vector machine (SVM), back propagation (BP), random forest (RF), extreme learning machine (ELM)) were used to build the calibration models based on o-nitrotoluene samples in different concentration ranges, and the prediction accuracy and robustness of the calibration model were compared. The results indicate that the effectiveness of different calibration methods is different when going from prediction accuracy to robustness. The prediction accuracy and robustness of RF models are not satisfactory. BP models, which are capable of producing very accurate results in terms of prediction accuracy, are not able to solve extrapolation problems. PLS model has more advantages in model prediction accuracy. ELM has shown the best behavior in terms of robustness of model, but is inferior to PLS in terms of prediction accuracy.