SOLUTION OF INVERSE PROBLEMS OF RAMAN SPECTROSCOPY OF AQUEOUS SALT SOLUTIONS WITH THE APPLICATION OF WAVELET NEURAL NETWORKS

Read more This paper presents the results of solving the problem of determining the salt composition of multicomponent aqueous solutions by their Raman spectra using artificial neural networks and the method of projections on latent structures. Three methods of input feature transformation are considered: aggregation of adjacent spectral channels, discrete and continuous wavelet transforms. It is shown that all of them can reduce both the dimension of the input data and the error of determination of the salt concentrations in the problem under consideration. The most effective method for the solution of the considered problem was a continuous wavelet transform. The multilayer perceptron trained on the transformed by this method input features provided the average error of determination of concentration for all 5 salts 0.023 M that is 38 % less than the error obtained by artificial neural network used without data compression. Thus, Raman spectroscopy combined with the use of artificial neural networks trained on transformed input features demonstrated high efficiency in solving the problem of identification and determination of concentrations of 5 inorganic salts dissolved in water.

Keywords:
Raman spectroscopy, aqueous solutions of salts, projections on latent structures, artificial neural networks, wavelet analysis.
УДК 535.3; 504.4; 004.8

SOLUTION OF INVERSE PROBLEMS OF RAMAN SPECTROSCOPYOF AQUEOUS SALT SOLUTIONSWITH THE APPLICATION OF WAVELET NEURAL NETWORKS