Interpretation of FTIR spectra by principal components-artificial neural networks

In order to improve the training speed and increase the predictive ability of artificial neural networks, principal component analysis (PCA) and partial least squares (PLS) were introduced to compress the original data. The principal components (PCs) of FTIR spectroscopic data matrix were obtained by PCA and PLS methods respectively, which were used as the inputs of neural networks. Results indicated that improvement was achieved in three aspects when the PCs instead of the original data were input to the networks. First, iterations were distinctly decreased from 8000 to less than 10. Second, computation time was shortened from 34.95 s to less than 1 s. Third, standard error of prediction (%SEP), mean relative error (MRE), and the root mean square error of prediction (RMSEP) decreased by 35% for the singular value decomposition-artificial neural network (SVD-ANN) and 80% for the nonlinear iterative partial least squares-ANN (NIPALS-ANN) or so, which means that the predictive ability was improved significantly. In addition, F-test was introduced to compare the performance of PCA and PLS for compression of original data, and it was shown that the latter model was more efficient. The presented methodologies of variable selection provide a simple and rapid technique for ANN to interpret FTIR spectra accurately and are advantageous to the widespread use of artificial neural networks.