Non-linear projection to latent structures revisited (the neural network PLS algorithm)

Abstract

Projection to latent structures (PLS) has been shown to be a powerful linear regression technique for problems where the data is noisy and highly correlated and where there are only a limited number of observations. However, in many practical situations, industrial data can exhibit non-linear behaviour. A number of methodologies have been proposed in the literature to integrate non-linear features within the linear PLS framework and thus provide a non-linear PLS algorithm. This paper presents an approach to the development of neural network PLS algorithms where either a sigmoid neural network or a radial basis function (RBF) network is fully integrated within the PLS algorithm using weight updating in the PLS input outer models. The potential improvements in modelling capability provided over the existing neural network PLS algorithms is assessed through comparisons on a simulation of a pH neutralisation process.