Estudo de misturas de café arábica e robusta usando FTIR e redes neurais artificiais

Abstract

Coffee is one of the most accepted and appreciated beverage by many countries in the world, being a natural product with distinct aromas and flavors. The main species are Coffea arabica (arabica) and Coffea canephora (robusta). They have a very different chemical composition, the arabica coffee provides a beverage with higher quality and aroma than the robusta coffee. Therefore, best quality beverages use only arabica coffee, due to the intense aroma, greenish grains, high acidity and lower amount of caffeine. The coffee blends are commonly used when you want to maintain a uniformity of flavor. Therefore, is necessary the development of reliable analytical methods to indicate the quantity of each type of coffee in a mixture. Artificial neural networks (ANN) are a set of principles based on mathematical and statistical techniques; it has now gained space to perform tasks of regression and pattern recognition. The ANN are capable of performing the mapping of complex and non-linear multivariate relationships between input and output. In this work two types of artificial neural network; the multilayer perceptron MLP, with supervised learning, and radial basis network RBF, with hybrid learning process, were used. The spectra were obtained in the spectroscopy equipment in Fourier transform infrared (FTIR), and appropriately pre-processed (normalization, baseline correction and smoothing). The results showed that multilayer perceptrons (MLP) and radial basis function networks (RBF) showed a similar performance with a mean absolute error of approximately 7% for test samples. So it appears that it is necessary to refine the technique to get minor errors. It is suggested the use of near infrared spectroscopy and/or analysis of extracts of samples.