Quantitative multivariate analysis with artificial neural networks

Abstract

Quantitative interpretation of spectra can be achieved by using artificial neural networks with multi-layer architecture. Both back-propagation (BP) and radial basis function (RBF) are implemented and tested with raw absorption spectra and normalized spectra of glucose solutions in MATLAB. Simulation results showed that the partial least square (PLS) method can have a better performance with small number in the calibration set. However, with increasing size of data set, as in the cross validation method, RBF and BP have better performance. With optimal spreading factor, RBF can have the same degree of accuracy but significantly faster convergent speed comparing to BP. The normalization scheme can also significantly affect the performance of both RBF and BP.