There has been an increase in systems for precision livestock monitoring in the past decade.
Automated behavioural classification and identification through sensors has the potential to improve health and welfare of the animals as these behavioural monitoring systems can be used to detect changes in behaviour that are associated with disease. However, these monitoring systems generate huge amounts of information that require processing and subsequent communication. Therefore, to eventually deliver a complete precision livestock monitoring framework in veterinary epidemiology that can efficiently monitor in real-time and for long durations, factors that have an impact on processing, communication and power consumption have to be evaluated. Additionally, such long-term architectural systems should be able to cope with challenging new or changing conditions. Using sheep behaviour as an example, some of these factors are explored including sampling frequency, sensor position, and window sizes on the performance of automated classification. Moreover we propose a combined offline and online learning algorithm that can handle new changing conditions in the classification of behaviour in sheep. Hence, both the combined algorithm and the embedded edge processing Intel device provide a useful technology for real-life long-term monitoring systems in precision livestock farming.