Publicação
Extraction of animal welfare indicators based on machine learning techniques
| Resumo: | This dissertation addresses the challenge of automatically detecting lambing events in real-time by leveraging sensor data gathered from collars worn by ewes. The primary motivation stems from the labor-intensive nature of manual observation and the need for prompt interventions around lambing, which can significantly affect animal well-being and farm efficiency. Fifty-four pregnant ewes from a norwegian farm were equipped with collars that recorded accelerometer X, Y, and Z axes and temperature data around the clock. Key preprocessing steps included thresholding and Interquartile Range (IQR) filtering to remove invalid or extreme values, Median aggregation to downsample the raw data (reducing size by over 95%) while preserving critical trends, scaling through min-max normalization to ensure consistency across features and files, and sampling strategies like undersampling and oversampling to mitigate severe class imbalance between the relatively brief lambing window and the much larger null periods. Using this refined dataset, multiple machine learning approaches were explored. Ensemble-based classifiers, specifically Random Forest and Extra Trees, consistently outperformed alternative algorithms, achieving high accuracy and Mathews Correlation Coefficient (MCC). Incorporating a short-term rolling window, like 60 seconds of past data, further boosted performance by capturing temporal patterns in sensor signals leading up to lambing. The top model achieved 0.83 MCC and 0.85 accuracy in distinguishing lambing states with sub-second inference speeds. Additionally, lightweight model variants were tested to measure suitability for on-collar microcontrollers. Although restricting model depth and estimator counts did reduce computational load, it also lowered overall accuracy, achieving MCC 0.58, suggesting further optimization or specialized hardware might be needed for fully embedded, real-time detection. Overall, the integrated pipeline and spanning data collection, rigorous preprocessing, and ensemble-based modeling shows promising potential for collar-based lambing detection. Nonetheless, false positives remain an area of concern, and further refinements are required to achieve the level of real-time, high-precision performance needed in practical farm settings. By reducing reliance on labor-intensive observation, this framework holds promise for advancing animal welfare and management efficiency across modern sheep farming operations. |
|---|---|
| Autores principais: | Arrais, Simão Teles |
| Assunto: | Animal welfare Machine learning Lambing Accelerometer Collar Sheep |
| Ano: | 2025 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade de Aveiro |
| Idioma: | inglês |
| Origem: | RIA - Repositório Institucional da Universidade de Aveiro |
| Resumo: | This dissertation addresses the challenge of automatically detecting lambing events in real-time by leveraging sensor data gathered from collars worn by ewes. The primary motivation stems from the labor-intensive nature of manual observation and the need for prompt interventions around lambing, which can significantly affect animal well-being and farm efficiency. Fifty-four pregnant ewes from a norwegian farm were equipped with collars that recorded accelerometer X, Y, and Z axes and temperature data around the clock. Key preprocessing steps included thresholding and Interquartile Range (IQR) filtering to remove invalid or extreme values, Median aggregation to downsample the raw data (reducing size by over 95%) while preserving critical trends, scaling through min-max normalization to ensure consistency across features and files, and sampling strategies like undersampling and oversampling to mitigate severe class imbalance between the relatively brief lambing window and the much larger null periods. Using this refined dataset, multiple machine learning approaches were explored. Ensemble-based classifiers, specifically Random Forest and Extra Trees, consistently outperformed alternative algorithms, achieving high accuracy and Mathews Correlation Coefficient (MCC). Incorporating a short-term rolling window, like 60 seconds of past data, further boosted performance by capturing temporal patterns in sensor signals leading up to lambing. The top model achieved 0.83 MCC and 0.85 accuracy in distinguishing lambing states with sub-second inference speeds. Additionally, lightweight model variants were tested to measure suitability for on-collar microcontrollers. Although restricting model depth and estimator counts did reduce computational load, it also lowered overall accuracy, achieving MCC 0.58, suggesting further optimization or specialized hardware might be needed for fully embedded, real-time detection. Overall, the integrated pipeline and spanning data collection, rigorous preprocessing, and ensemble-based modeling shows promising potential for collar-based lambing detection. Nonetheless, false positives remain an area of concern, and further refinements are required to achieve the level of real-time, high-precision performance needed in practical farm settings. By reducing reliance on labor-intensive observation, this framework holds promise for advancing animal welfare and management efficiency across modern sheep farming operations. |
|---|