Desenvolvimento de metodologias para a detecção do comportamento ingestivo de bovinos

Abstract

Precision Livestock Farming has revolutionized animal production by integrating sensors, data science, and machine learning algorithms. However, some methodological foundations have not yet been fully explored. With this in mind, this study aimed to develop and validate methodologies for automating the detection of cattle ingestive behavior. This work used wearable sensors, including a triaxial accelerometer and gyroscope, fixed on the animals' neck and collar. Data collection took place between 2023 and 2025 across eight distinct production scenarios. Continuous visual observations were made, which were then manipulated to simulate time-interval observations. The data were organized into nine distinct databases and subsequently unified. Holdout validation and external validation were employed. Four algorithms were compared: Random Forest (RF), Light Gradient Boosting (LGB), K-Nearest Neighbors (KNN), and Artificial Neural Networks (ANN). Three hyperparameter search strategies were tested (Random Search, Bayesian Search, and Evolutionary Algorithms), as well as balancing techniques (SMOTE, ADASYN, NearMiss, and Random Undersampling). The impact of different data capture intervals (1s to 60s) and sensor attachment locations was also assessed. The performance metrics were Accuracy, Precision, Recall, and F1-Score. The predictive models performed well in classifying feeding, rumination, and idling behaviors, with RF standing out, especially in binary datasets. The generalization of the databases improved accuracy in external validation, confirming the hypothesis that algorithms trained on heterogeneous data are more robust. Although Holdout validation yielded better metrics, it overestimated the results. Balancing strategies impacted performance when considering precision, recall, and F1-Score, but not accuracy. Regarding sensor placement, neck-mounted sensors outperformed collar-mounted ones, particularly in detecting feeding behavior. Larger intervals between data captures (up to 60s) maintained good performance, indicating practical feasibility. Two-minute interval-based visual observations proved effective, saving observer effort without compromising accuracy. The combined use of these methodologies allows for progress in creating solutions applicable to the field reality of livestock 4.0. It is concluded that the integration of sensors, optimized algorithms, and heterogeneous datasets offers high potential for the automated monitoring of cattle ingestive behavior.