Combining artificial neural networks with data fusion to analyze overlapping spectra of nitroaniline isomers

Abstract

A novel method named DF–GRNN based on generalized regression neural network (GRNN) combined with data fusion (DF) was applied to enhance the ability of extracting characteristic information and the quality of regression for the simultaneous spectrophotometric determination of o-nitro-aniline, m-nitro-aniline and p-nitro-aniline. Data fusion is a technique that seamlessly integrates information from disparate source to produce a single model or decision. Wavelet representations of signals provide a local time–frequency description and are multiscale in nature, thus in the wavelet domain, the quality of noise removal are implemented by a scale-dependent threshold method. Information from different wavelet scales is just like different sources of information. Integrating the information from different wavelet scales to obtain a GRNN model belongs to the technique of data fusion. GRNN was applied for overcoming the convergence problem met in back propagation training and facilitating nonlinear calculation. In this case, by optimization, wavelet functions, decomposition level and thresholding methods and the width (σ) of GRNN for DF–GRNN were selected as Daubechies 18, 7, HYBRID thresholding and 0.4 respectively. The relative standard errors of prediction (RSEP) for total compounds with PLS, PCR, BP-MLFN, DF–LS-SVM, DF–GRNN and GRNN were 6.74, 7.40, 25.6, 4.47, 4.73 and 6.56%, respectively. Experimental results showed the DF–GRNN method to be successful for simultaneous multicomponent determination even where there was severe overlap of spectra and to be better than GRNN,PLS and PCR. The DF–GRNN method is a hybrid technique that combines the best properties of the two techniques, which makes this method attractive and promising.