Comparação entre algoritmos de seleção de variáveis em regressão PLS para determinação de parâmetros de qualidade de biodiesel por espectrometria NIR

Abstract

The Firefly Bioinspired Algorithm (FFiPLS) is based on the attractiveness of fireflies as swarming behavior when looking for food. The variable selection technique is able to reduce the number of variables by intervals in order to obtain better predictions from the chemometric model. In this sense, the present study aims to compare the performance of PLS regression chemometric models using FFiPLS interval selection algorithms against the iSPA-PLS and iPLS to determine biodiesel quality parameters by NIR spectrometry. The development of this work involved obtaining samples of biodiesel from soybean oil via methyl route. Subsequently, the density of biodiesel was measured and finally added to diesel in proportions 5, 10, 15, 20, 25, 30, 35, 40, 45 and 50 (% v / v), the biodiesel index in diesel, both carried out in triplicate. A database was built from the measurements of the absorption spectra in the near infrared with 100 samples of biodiesel, in the range of wavelengths (λ) from 410 to 2500 nm. From this range, two spectral bands were chosen, 441-1551 nm and 1100-1600 nm. The regions were chosen because they are in the first and second overtone regions of the C – H. The chemometric treatment consisted of selecting the samples from the calibration and prediction sets using the SPXY algorithm, and to remove outliers the Hotelling T² test. The Savitzky-Golay derivation, a 2nd degree polynomial and a 17- point window, was used as spectral pre-processing. The software used for pre treatments and for removing outliers was The Unscrambler® version 9.7. Matlab® software version R2016a was used to select samples and variables as well as to build the models. Two other algorithms for selecting variables by intervals were compared to the FFiPLS algorithm: iSPA-PLS and iPLS. For the diesel biodiesel index, models with a range of 441-1551 nm with spectral pre-processing showed low values of latent variables, with better values for the prediction step (RMSEP, R²prediction and REP) for the three selection algorithms of variables, iSPA-PLS, iPLS and FFiPLS. As for density, the model with a range of 410-2500 nm with spectral pre-processing, the FFiPLS algorithm presented better values in the prediction step (RMSEP, R²prediction and REP).