Use of Artificial Neural Networks, Aided by Methods to Reduce Dimensions, to Resolve Overlapped Electrochemical Signals. A Comparative Study Including other Statistical Methods

Abstract.

 A method using artificial neural networks (ANNs) combined with Fourier Transform (FT) and Wavelet Transform (WT) was used to resolve overlapping electrochemical signals. This method was studied as a powerful alternative to traditional techniques such as principal component regression (PCR) and partial least square (PLS), typically applied to this kind of problems.

WT and FT were applied to experimental electrochemical signals. These are two alternative methods to reduce dimensions in order to obtain a minimal recomposition error of the original signals with the least number of coefficients, which are utilized as input vectors on neural networks. Tl⁺ and Pb²⁺ mixtures were used as a proof system.

In this paper, neural networks with a simple topology and a high predictive capability were obtained, and a comparative study using PLS and PCR was done, producing the neural models with the lowest RMS errors. By comparing the error distributions associated with all the different models, it was established that models based on FT and WT (with 11 coefficients) neural networks were more efficient in resolving this type of overlapping than the other chemometric methods.