Detecção de adulteração por adição de leite bovino ao leite bubalino utilizando redes neurais artificiais e outras técnicas de mineração de dados

Abstract

The growing demand for buffalo milk products in the Brazilian market in recent years has encouraged more producers to invest in this production area. The high nutritional value of buffalo milk compared to bovine milk has contributed to this increase in consumption, which leads to difficulties in acquiring milk during the off-season. Brazilian legislation allows the marketing of milk with a mixture of different species, as long as it is properly identified on the package label. However, this practice often happens without consumers awareness, or before the milk even reaches the industry. In order to improve fraud detection, this study aimed to perform physical-chemical analysis of cow and buffalo milk using infrared spectroscopy to compare the milk composition the two species and to develop models of Artificial Neural Networks and other data mining techniques from compositional data to detect the addition of cow milk to buffalo milk. For the physical-chemical analysis of buffalo, samples were collected from refrigeration tanks over 24 months, totalling 837 samples. Analyses were performed for the composition, fat, lactose, protein, total solids, and solids non-fat contents. Higher fat contents (6.19%) and protein (4.25%) were find in the spring and summer seasons, respectively. Regarding total solids, the highest values were found in the spring seasons (16.14%) while solids non-fat had higher values in the summer (10.03%), which is the rainy period of the region. For the studies involving the comparison of bovine milk composition with buffalo milk and artificial intelligence, 300 samples of buffalo milk samples and 300 samples of cow milk were collected during the months of October/2021 to March/2022, and mixtures with nine levels of bovine milk addition to buffalo milk were prepared, simulating adulteration (1%, 2%, 5%, 10%, 20%, 30%, 40%, 50%, 75%) and two levels without mixing (0% - buffalo milk without addition and 100% - cow milk without addition). The samples were analyzed by FITR and the results obtained were tabulated and statistically compared with higher components concentration for buffalo milk in all evaluated parameters, these results were used to set up the neural network architectures along with laboratory routine data. Multilayer Perceptron Networks with one and two hidden layers and Radial Base Function Neural Networks were tested using IBM SPSS Statistics® software. The same database was employed to test data mining techniques using the Rapidminer® software. Naive Bayes, Generalized Linear Model, Logistic Regression, Fast Large Margin, Deep Learning, Decision Trees, Random Forest, Gradient Boosted Trees and Support Vector Machine tests were used. The Multilayer Perceptron networks achieved classification results of 97.4% and 97.0% for one and two layers respectively, while the radial basis function obtained 97.3% of accuracy. The data mining models with the best results were Random Forest, Support Vector Machine and Decision Trees. The results showed good accuracy, sensitivity, specificity and precision for both the Neural Network tests and the other data mining tests, demonstrating that they are good analysis and adulteration prediction tools to be used as a screening test by the industry.