Aplicação de técnicas espectroscópicas, métodos quimiométricos, fusão de dados e seleção de variáveis no controle de qualidade de blends das espécies de café arábica e robusta

Abstract

The main objective of this thesis was to develop multivariate models to quantify and characterize mixtures of Robusta and Arabica coffees. For this purpose, 120 blends of ground coffees (0.0-33.0% m/m), prepared with coffee samples originated from ten different farmers, were formulated at three different degrees of roasting: light, medium and dark. Different instrumental techniques were used: attenuated totalreflectance Fourier transform infrared (ATR-FTIR or MIR) spectroscopy, near infrared (NIR) spectroscopy, paper spray ionization mass spectrometry (PS-MS) and total reflection X-ray fluorescence (TXRF). Models using partial least squares regression (PLS) were built individually for the spectra from each technique. In the sequence, datafusion models (different combinations of techniques) were also built at low and medium levels, in order to take advantage of the synergy between the datasets. The models were optimized by variable selection methods, such as genetic algorithm (GA) and ordered predictors selection (OPS). In general, the smallest prediction errors wereprovided by the low-level data fusion models. In all the cases, the variable selection methods significantly reduced the mean square errors of prediction (RMSEP) and the number of variables, increasing the correlation coefficient values between predicted and reference values. PLS models were interpreted through informative vectors andspecific coffee components were detected as marker species, such as trigonelline, sugars and chlorogenic acids. For the atomic data, the elements Mn and Rb were mostly detected as possible markers of the coffee species. The best models (MIR and MIR-PSMS) were validated and proper figures of merit were estimated, corroborating their accuracy, linearity, sensitivity and absence of bias.