Independent Component Analysis Combined with Neural Networks and Its Application to Raman Spectroscopy

A new method of model construction based on back-propagation artificial neural networks (BP-ANN) regression and independent component analysis (ICA) was proposed. In its application to Raman spectrum, the data of Raman spectrum were firstly compressed by wavelet transform, their independent components and the contribution matrix were then extracted by the independent component analysis, and finally, the model of the contribution matrix and concentration matrix was built by the artificial neural networks regression. The influence of the numbers of independent components and the neurons in the hidden layer on the properties of model was further analyzed. This new chemometric method has been applied to the determination of active substance content of four types of pharmaceutical tablet samples. The correlation coefficients (R) between the analytical values and the model predicted values of active substance contents are 0.995, 0.967, 0.976, and 0.982, respectively.