É POSSÍVEL IDENTIFICAR TIPOS DE FRAUDE NO LEITE POR MEIO DE REFLECTANCIA HIPERESPETRAL APLICADA A APRENDIZAGEM DE MÁQUINA?

Abstract

Milk can be adulterated with different solutions, in order to generate a greater amount of milk, with a lower content of pure milk. Milk adulteration is a dangerous practice in the dairy industry, extremely harmful to consumers, as milk is one of the most consumed food products. Thus, the use of spectral data together with machine learning can be an essential tool in the detection of milk fraud by different sources of products used for this purpose, especially because so far it is an innovative methodology, since there are no reports of something similar in the literature. The objective of this work was to identify if it is possible to distinguish spectral adulterations in milk with different adulterants and if with the spectral information applied to ML models to find the best ML model for this distinction. . Samples of pure raw milk were used, to which samples of the same milk, adulterated with different substances, were compared. Specifically, raw milk was adulterated with whey, water, and starch at concentrations of 5%, 10%, 15%, 50%, and 75% for each adulterant. The samples were read in a spectroradiometer twice each. Subsequently, the data were tabulated for analysis with machine learning (ML). The following algorithms were used: artificial neural networks (ANN), decision tree J48 (J48), REPTree (DT), random forest (RF) and support vector machine (SVM) and logistic regression (RL) which was used as a control model. To evaluate the accuracy of the algorithms' classification, the percentage of correct classifications (CC), kappa and F-score were estimated. The methodology proposed in the work was that it would be possible to identify a difference in the spectral behavior of adulterated milk samples from pure milk samples, making it possible to identify more clearly the distinction between curves in the VIS and NIR regions. Another objective of the work was to verify the accuracy of machine learning algorithms in the classification of milk adulterations, in which the support vector machine (SVM) stood out, reaching accuracies of 80 for correct classification (CC), close to 1 for F-score and above 0.8 for kappa coefficient.