Aplicação do aprendizado de máquina sobre a qualidade da produção familiar do leite

Abstract

Paraná stands out in dairy production with significant growth in this activity. However, there is great heterogeneity in the profile of the producers. The Southwest region of Paraná is characterized by milk production mainly from family farming and is one of the most productive in the state. The objective of this research was to apply the data mining technique (machine learning) through decision tree models, composed of rules, to identify patterns to predict milk quality using the indicators of somatic cell count (SCC) and total bacterial count (TBC) in data collected (questionnaires and laboratory analyses) from milk producers in the Southwest region of Paraná. Models were created for the final classes SCC and TBC, resulting in two decision trees. A database was used for training (70%) and a database for testing (30%), totaling 221 records. In data categorization, the classification of milk for SCC was: A (up to 400,000 cells/mL), B (up to 500,000 cells/mL), or C (above 500,000 cells/mL), and for TBC: A (up to 100,000 CFU/mL), B (up to 300,000 CFU/mL), and C (above 300,000 CFU/mL). After these steps, data mining and knowledge extraction were performed. Regarding the decision tree results, the relevant attributes for generating the SCC model were lactose content and type of milking. For the TBC model, the main attributes were the number of animals and the use of bleach in cooling tank sanitation. For both models (SCC and TBC), the daily milk production in liters and the origin of the water used in the milking procedures were important factors. The efficiency of the generated models was 63.64% for SCC and 74.24% for TBC.