Statistical projection methods and artificial neural networks

Abstract

There is an increasing strategic interest in the use of NIR to monitor and control material properties. The aim being to tightly control quality variables and reduce product variability. Given the number of academic publications in NIR spectroscopy the question arises as to why has industry been so slow in taking it up. The interpretation of spectral data is usually achieved using the linear projection techniques of principal components regression and projection to latent structures. However, nonlinear responses, which can be attributed to a whole range of sources, including physical, chemical, sensor and instrument, as well as asymmetric noise effects, etc., are evident in many chemical processes. Multivariate methods capable of detecting and modelling nonlinear features have received increasing attention, e.g. multivariate adaptive regression splines and neural networks. A study is carried out on industrial spectral measurements to assess the ability of these analysis methods as inferential estimators of material properties for process control.