QUANTITATIVE-ANALYSIS OF PAPER COATINGS USING ARTIFICIAL NEURAL NETWORKS

This paper describes a neural network approach to the quantitative ana lysis of paper coatings. infrared spectra of samples of coated paper w ere used as input data for the different types of artificial neural ne tworks. Unsupervised learning was applied to obtain the clustering of samples with respect to similarities in the spectra. The self-organizi ng artificial neural network of Kohonen type produced a visual represe ntation of the discovered groupings on a two-dimensional plane. Such m apping provided the expert a possibility to analyze the mutual arrange ment of samples and to predict the properties of the test samples usin g their relative position with respect to existing clusters. Supervise d learning with a multilayer feedforward network was used to construct the non-linear models that relate the spectral information and concen trations of three basic components of paper coating - styrene, butadie ne, and carbonate. These models were used for prediction of concentrat ions of paper coating components for the test data set. The results of modeling demonstrate that accuracy of classification and prediction i s better than those obtained with traditional methods like principal c omponent analysis or partial least squares (from 4% to 2% for differen t components). According to our experience, the modeling with artifici al neural networks is intuitively clear for the expert. This method al lows to construct complex multivariable and multiresponse models in un ified style. Causal relationships between inputs and outputs can be an alyzed and explained.