Incremental and Decremental Active Learning for Optimized Self-Adaptive Calibration in Viscose Production

In viscose production, it is important to monitor three process parameters as part of the spin bath in order to assure a high quality of the final product: the concentrations of H2SO4, Na2SO4 and ZnSO4. NIR-spectroscopy is a fast analytical method applicable to conditions of industrial production and is capable of determining those concentrations. The collective composition of the spin bath varies in the industrial process, which implies changes in the matrix of the aforementioned analytes. Thus, conventional static chemometric models, which are trained based on collected calibration spectra from Fourier transform near infrared (FT-NIR) measurements, show a quite imprecise behavior when predicting the concentrations of new on-line data. In this paper, we are presenting a methodology which is able to cope with on-line self-calibration and -adaptation demands in order to compensate high system dynamics, reflected in conceptual changes in the mappings between NIR spectra and target concentrations. The methodology includes intelligent strategies for actively selecting those samples which should be accumulated into and excluded from the current data window in order to optimize the generalization performance of calibration models (thus termed as incremental and decremental active learning stages) while keeping the number of update cycles (and thus required target measurements) as low as possible. This follows the company requirements in terms of necessary cost reduction. Experiments on real-world data streams from viscose production process show that the new self-calibration methods are able to significantly reduce the number of update cycles while still keeping the predictive quality of the calibration models high (below 5% errors) for H2SO4 and Na2SO4. Incremental active learning is able to smoothen and improve the overall quality of the predictions, while decremental active learning achieves a lower number of medium to large prediction errors.