On-line Outlier/Redundancy Filtering and Semisupervised Incremental Calibration Modeling in Melamine Resin Production using FT-NIR Spectra

In the considered batch process of melamine resin production, the essential parameter to be monitored is the cloud point in the condensation process. It indicates the optimum moment to start the cooling process in order to stop the condensation and is of importance to assure high product quality. The current manual controlling of the condensation process is time-consuming and includes some uncertainty in the determination of the condensation end-point. Standard linear models have been trained on an initial calibration set and applied to a separate in-line test set. However, the obtained results leave much room for improvement, especially in case of lamp changes and intensity downtrends of the halogen light source, and some dynamics which cannot be fully controlled by static models. Furthermore, some internal dynamics, mainly non-linear, such as changing compositions of the educt, are assumed. In the data acquisition process, three consecutive spectra are recorded and matched to one single target measurement, and many other spectra, without target information, is stored. Theoretically, the latter could be considered as repetitions, thus averaged, but the presence of outliers and some other factors, produce artifacts that influence severely the posterior model accuracy. The non-linear behavior requires the usage of non-linear models, preferable those ones in which the degree of non-linearity can be adjusted on demand and on the fly.