Industrial process monitoring using nonlinear principal component models

Abstract

A new approach to identifying nonlinear principal component models is presented. This involves the application of linear principal component analysis (PCA) prior to the identification of a modified autoassociative neural network (AAN) that represents the required nonlinear PCA model. The benefits of this new approach are that (i) the size of the reduced set of linear principal components (PCs) is smaller than the set of recorded process variables, and (ii) the set of PCs is better conditioned as redundant and insignificant information is removed. The result is a new set of input data for a modified network. The usefulness of this approach is illustrated using a recorded industrial data that relates to crack detection in an industrial melter process.