Artificial neural networks in classification of NIR spectral data: Selection of the input

Artificial neural networks (NN) with back-error propagation are used as a classifier to identify different strengths of drugs in formulations, different qualities of solvents and polymers through NIR spectra. To eliminate some irrelevant information and reduce the number of variables, several pretreatment methods based univariate feature selection, principal component analysis (PCA) and Fisher transformation (FIT) and some combinations are developed. Nine data sets are treated to study the effect of different data pretreatment methods to select the input of NN. Compared to pretreatment by PCA, univariate feature selection followed by PCA reduces somewhat the size of the structure of the NN for some data sets. PCA followed by FIT greatly reduces the architecture of NN for eight of the nine data sets. Our results suggest that PCA/FIT is useful way to pretreat the data as input of NN.