SenSwine: Monitoring a Swine Farm through the Adoption of Multiple Sensors and Machine Learning Algorithms

• Grant number: 25/05293-3
• Support Opportunities: Scholarships in Brazil - Doctorate
• Start date: December 01, 2025
• End date: July 31, 2029
• Field of knowledge: Agronomical Sciences - Animal Husbandry - Animal Production
• Principal Investigator: Ricardo Vieira Ventura
• Grantee: Bruno Braga Carnino
• Host Institution: Faculdade de Medicina Veterinária e Zootecnia (FMVZ). Universidade de São Paulo (USP). São Paulo , SP, Brazil

Abstract

The global increased demand for food due to population growth will pressure the livestock farming sector to enhance its efficiency and production. Swine farming will be particularly impacted, as pork is the second most consumed meat worldwide. However, the increase in the number of animals will not be accompanied by a proportional rise in labor availability, making it more challenging to monitor animal health and welfare and, consequently, compromising production efficiency. In this context, automated monitoring systems that combine sensors and artificial intelligence (AI) algorithms can continuously track animals, aggregating information on their health, welfare, and productivity. Moreover, these systems allow for the early and accurate detection of problems, reducing the need for human intervention. With the software output, on-farm decision-making is streamlined, and productivity losses can be controlled. The accuracy of AI-based detection can be further improved by integrating multiple sensors that generate different types of data, an approach that remains underexplored in precision livestock farming. Thus, the present study aims to apply machine learning techniques for the integrated processing of data from multiple sources (collected via sensors) to detect pig behaviors related to feeding and welfare during the growing-finishing phase. Additionally, the developed algorithms will be optimized to predict zootechnical parameters such as live weight (LW) and feed intake (FI) of the herd. To that end, three growing-finishing pens, each housing three pigs, will be equipped with RGB cameras, microphones, accelerometers, RFIDs, and thermometers and monitored for 90 days. During this period, feed intake will be controlled, and pigs will be periodically weighed. After data collection, researchers will manually identify the relevant features of each evaluated behavior. These labeled data will then be used to train neural networks for the automated classification of such behaviors. The feed intake prediction algorithms will utilize features extracted through machine learning from different sensors (microphones, cameras, and environmental condition monitors), which will then be integrated with the classified behavioral data to minimize prediction errors. Once the algorithms are developed, a new data collection phase will be conducted to apply and validate the developed pipelines in a practical setting. (AU)