Self-Adaptive Non-Linear Methods for Improved Multivariate Calibration in Chemical Processes

In this paper we present new methods for non-linear multivariate calibration and their application under realworld conditions. The developed non-linear methods, which are applied to FT-NIR absorbance spectra recorded in-line in different industrial production processes (i-Red GmbH and RECENDT GmbH) result in enhanced model quality and robustness. We propose two new concepts for reducing the dimensionality of the calibration problems, which may get severe when several hundreds or thousands of wavelengths are contained in the spectra. One is based on a statistical approach using a modified variant of forward selection, but extending it to extract bands instead of single wavelengths. Thereby, the robustness with respect to noisy recordings is increased. This concept is termed as forward selection with bands; the other one is a wrapper method which is based on a global heuristic search process achieved through genetic algorithms. Internally, they employ a new fuzzified crossover operator in order to weight nearby-lying bands accordingly. The calibration phase is equipped with an own developed non-linear version of PLS (partial least squares) on the basis of Takagi-Sugeno fuzzy inference systems offering the possibility to define piece-wise linear predictors which are combined to a non-linear model through Gaussian kernels, representing fuzzy rules. We will further demonstrate methods how to incrementally adapt the non-linear version of PLS over time with new incoming samples in order to account for significant system dynamics with different outweighing strategies. This is essential to assure high stability and predictive performance of the calibration models in case of dynamic processes over a long timeframe. The application potential of these methods will be underlined by several results achieved from a viscose fiber production process (at Lenzing AG) and from a melamine resin production process (at Metadynea Austria GmbH).