On the application of neural networks modelling to a wet end chemical process in paper making

This paper presents a novel approach and its application to the modelling of non-linear unknown systems via B-spline neural networks. The system is assumed to be described by a NARMAX model, which is subjected to coloured noise. A new regression formula is constructed and it has been shown that the generalised RLS algorithm can be directly applied to train the weights and produce a desirable estimate on the various modes of the noise. The selection of the knots distribution of B-spline neural networks is also considered. An algorithm has been developed which produces an optimal knot distribution for an input axis using training data. Finally, the method developed is applied to build up a local model for a paper machine wet end, which represents the relationship between added chemicals (rosin and alum) and the sizing of the resulting paper. Desired results are obtained