Domain-Adversarial Approach to Neural Networks Training to Determine the Composition of Wines Using Various Techniques for Measuring IR Absorption Spectra

Abstract

The study is devoted to solving the problem of simultaneous determination of the concentration of five components of wines (ethanol; a mixture of glucose, fructose and sucrose; a mixture of tartaric, malic and citric acids; glycerol; sulfur dioxide) with the IR absorption spectra of wines using neural networks. The use of domain-adversarial training of neural networks for quantitative analysis of the wine components concentration enabled overcoming the problem of IR spectra registering technique influencing their shape and ensured simultaneous determination of the concentration of the desired components with satisfactory accuracy.