Nonlinear process monitoring using bottle-neck neural networks

Abstract

In industry and laboratories, statistical process control (SPC) is often used to check the performance of processes. The need for multivariate statistical process control (MSPC) becomes more important as the number of variables that can be measured increases. It is common practice to use principal component analyses (PCA) or partial least squares (PLS) to construct multivariate control charts. However, PCA and PLS are linear in nature whereas many processes exhibit nonlinear relations between the process parameters and the quality parameters (i.e. the settings and the product). An example of such a nonlinear relation is the value of the pH as a function of the input flow of an acid and a base. In this paper the first approach of a novel method is presented which uses the centre hidden neurons of a bottle-neck neural network to perform nonlinear MSPC. The output of the bottle-neck network are the reconstructed input set and a predicted dependent set. Furthermore, a special case of a bottle-neck neural network (an auto-associative neural network) is also used for nonlinear MSPC. The output of auto-associative neural networks is a reconstruction of the input set.