APLICAÇÃO DE MÉTODOS DE SELEÇÃO DE VARIÁVEIS POR INTERVALOS NA QUANTIFICAÇÃO DE BANHA DE PORCO EM MANTEIGA DE CACAU

Abstract

Cocoa butter is considered a product with high added value, being the fundamental ingredient in the production of chocolates. Due to its high cost, lower value raw materials can be added to cocoa butter, such as hydrogenated vegetable margarine and lard. Therefore, it is necessary to develop analytical methods for the quality control of this type of food product. In this perspective, this work evaluated the use of spectroscopy in the medium infrared (MIR) combined with the regression by partial least squares (PLS) and methods of selection of variables by intervals to quantify lard in cocoa butter. In this study, 68 samples in the range of 5.0 to 30.0% of lard were adulterated in cocoa butter, spectra were obtained in the Middle Infrared region and all models were built using the MATLAB 6.1 software. The PLS models were built by centering the data on the mean and employing the cross validation process by venation blinds. The methods of selection of variables used were regression by Intervals Partial Least Squares (iPLS), regression by Backward Interval Partial Least Squares (biPLS) and Synergy Interval Partial Least Squares (siPLS), and we sought, among them, the model that presents the lowest mean cross-validation square error (RMSECV) and the lowest number of variables compared to the global PLS model, that is, the model containing all spectral variables (PLS). The PLS model was built with 6 latent variables (VL), having 0.7756 of RMSECV, 0.5555 of mean square error of calibration (RMSEC) and 0.6917 of mean square error of forecast (RMSEP). The best PLS model obtained from the variable selection methods was the iPLS in which the spectral data were divided into 25 intervals. And, the best interval indicated was the X interval containing Y variables. This model was built with 5 VL and presented the following parameters RMSECV of 0.4410, RMSEC of 0.3745 and RMSEP of 0.3805. Therefore, it was found that the iPLS25 model had better predictive capacity than the global PLS model, therefore, this model can be used in the quantification of lard in cocoa butter.