Time Series Analysis of Calf Behaviour: Machine Learning Methods for Disturbance Detection

Advancements in Precision Livestock Farming (PLF) highlight the need for real-time animal welfare monitoring, especially in dairy calves, to detect disturbances early and optimize health and productivity. Calf behaviour has been widely recognized as a reliable welfare indicator, making its accurate classification essential. However, accurately classifying calf behaviours from time series data generated by accelerometer neck collars presents significant challenges due to data complexity and variability. Existing methods often lack the precision needed for effective real-time monitoring, leaving a gap in practical applications. This thesis introduces a scalable machine-learning framework for analyzing time series data from accelerometer sensors worn by calves. The main contribution lies in its ability to accurately classify calf behaviours and provide an application to help detect disturbances, offering a novel approach to improving management practices in dairy farming. The research uses the ActBeCalf dataset, developed during this PhD, alongside manually annotated video footage. This dataset provides a gold standard for behaviour classification and is a valuable resource for advancing calf behaviour analysis. The thesis outlines a process for converting raw accelerometer data into an annotated dataset. This process includes synchronizing data with video annotations, correcting time drifts, and conducting manual verification to ensure precision. Statistical features were used to classify pre- and post-dehorning periods in the first feature engineering stage. Feature selection was then employed using traditional statistical methods and Machine Learning techniques to identify informative features from the time and frequency domains. These features capture subtle patterns in calf activity, enabling accurate classification and detection of disturbances. The second feature engineering stage focuses on classifying specific calf behaviours critical to welfare monitoring, such as drinking milk, grooming, lying, running, and walking. Machine learning algorithms, including the ROCKET feature extraction mechanism and Deep Learning approaches, are tested for accuracy and generalizability across different calves and conditions. This classification provides a foundation for monitoring calf welfare by identifying deviations from normal behaviour patterns. Higher-level features are derived from these classifications, such as time spent on each behavior, frequency, activeness vs. inactivity, and behavioral transitions. These metrics offer a comprehensive understanding of calf welfare, enabling the identification of subtle behavioural changes that may not be apparent from raw data alone. This approach enhances the detection of disturbances and provides insights into overall calf well-being, supporting more informed and proactive management decisions. The research demonstrates that integrating well-chosen statistical features, effective behaviour classification models, and behavioural metrics can enhance real-time calf welfare monitoring. The developed framework offers valuable insights that can improve animal welfare and farm management practices. The thesis also highlights the potential for incorporating additional data sources, such as environmental sensors and video monitoring, to further refine disturbance detection and provide a more comprehensive understanding of calf welfare. Integrating multimodal data could lead to more accurate predictions, enabling farmers to make informed decisions and take proactive measures to ensure livestock well-being. In conclusion, this research significantly contributes to Precision Livestock Farming by advancing Machine Learning applications to analyze time series data in real-world livestock science. It paves the way for more sophisticated automated systems for livestock management, with important implications for the future of animal husbandry.