Classificação do teor de biodiesel metílico de macaúba e níger na mistura com diesel, por meio da combinação da técnica de Espectroscopia MIR e dos métodos classificadores PLS-DA e SVM

Abstract

In recent decades, public policies have been developed to encourage the use of biofuels in Brazil, such as the establishment of a minimum mandatory content of biodiesel to commercialized diesel. In order to avoid adulterations in the composition of this blend established by law, a strict quality control is necessary throughout the entire production and distribution chain. Regarding this topic, many researches can be promoted, in order to develop methodologies that could give us quick and reliable answers for the quality control of this binary mix. In this work, analytic methods were developed to classify methyl biodiesel of macaúba and níger samples (both combined as a mixture) combined with diesel, by the combination of analytical/chemometric techniques of medium infrared spectroscopy (MIR), Discriminant Analysis by Partial Least Squares (PLS-DA), and by the use of support vector machines (SVM). For the construction of the discriminative models, a total of 74 combinations of BX diesel were produced, generated by the addition of methyl biodiesel of macaúba and níger to the pure diesel (B100), in a concentration range from 0,25% to 30,00% (v/v). The first group of samples was prepared using a proportion of 10% biofuel and 90% of pure diesel (B10), and the second according to different BX percentages (B3 to B9 and B11 to B30 containing, each of them, from 3 to 9% and from 11 to 30% of biodiesel). For each PLS-DA model were used 50 samples for the training sets and 24 samples for the testing. During the quality analysis of the models, the root mean squared errors of prediction were low, with percentages within the reproducibility of the standardized method. The performance criteria obtained for the developed models, such as sensitivity, specificity, precision and efficiency were of 100%. They also presented an excellent capacity to discriminate all the samples by its respective blends, with a prediction 100% accurate. The SVM algorithm also proved to be a great option to discriminate between the two classes of methyl biodiesel of macaúba B10 and BX. Regarding the performance analysis of the predictive models, the data base was divided in three subsets: training, test and validation, including 50/12/12 samples respectively. Different functions of the Kernel, such as the linear, polynomial, sigmoid and the radial base function (RBF) were tested, aiming to evaluate the generalization capacity of the classifier. Regarding the níger model, the RBF Kernel was the one that presented the best accuracy (0,92) in relation to the rest, with the hyperparameters defined as 𝐶=10 and 𝛾=10−4. For the macaúba models, a great precision was obtained for all the tested Kernels, with the hyperparameters defined as 𝐶=10 and 𝛾=10−5. The precision and the efficiency of classification of the SVM were higher than 70,0% in the test analysis and validation in both models. The results also revealed a value equal to 100% for the training data, which reinforces the discrimination capacity of the SVM models. In general, the good performance obtained in this study, suggests that the analytical methodologies are viable. The combination of analytical / chemometric techniques allowed to discriminate, quickly and efficiently, the samples of B10 and BX from the two methyl biodiesels in the mixture with diesel, also offering the possibility of minimizing the use of solvents and reagents. The good performance obtained in this study, indicates that this kind of methodology is viable, and could be used in the quality control of this fuel.