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

In viscose production, it is important to monitor the concentration of several substances (H2SO4,Na2SO4 and ZnSO4) as part of the spin bath in order to assure a high quality of the final product. The acid and the two salts govern the precipitation and agglomeration of the cellulose from viscose solution and the formation of the viscose fibre. During on-line production, these process parameters usually show a quite high dynamics depending on the fibre type that is produced and on environmental influences. In such cases, conventional chemometric models, such as principal components regression, partial least squares regression, locally weighted regression and others [1][2], as well as non-linear techniques recently employed in calibration, e.g. [3][4], may show severe downtrends in performance when quantifying the concentrations of new on-line data. This is because they are established once based on pre-collected calibration spectra and kept fixed during the whole life-time of the on-line process, thus not being able to adapt to dynamically changing situations at the system. Recently, a new concept termed as eChemo (evolving chemometric models), was introduced in [5] to overcome these deficiencies of static calibration. It possesses the ability to self-adapt and re-calibrate based on newly recorded on-line spectra obtained through FT-NIR measurements, but it requires permanent supervision, i.e. real values measured by means of a titration automaton, which are time intensive and expensive from an industrial viewpoint.