Alarming Pig Vocalization Based Prediction using Self Supervised Model BEATs

This thesis presents an advanced approach to classifying pig vocalizations into meaningful categories, focusing particularly on distinguishing pig sounds from non-pig sounds, as well as classifying pig sounds as alarming or non-alarming. Leveraging the robust BEATs model architecture, this work employs both supervised and semi-supervised learning techniques, significantly enhancing performance through the integration of pseudo-labeled data. The supervised baseline phase demonstrated strong initial classification capabilities, with accuracy reaching approximately 90.89% for pig sound classification and approximately 96.8% for alarming sound detection. Subsequent semi-supervised fine-tuning further improved the model’s performance, increasing accuracy to 95.1% for pig versus non-pig classification, and about 96% for alarming versus non-alarming sound detection. The semi-supervised approach introduced additional challenges, notably fluctuations in validation metrics due to the integration of pseudo-labels. However, these were effectively explained by the competing pressures inherent in semi-supervised training. Additionally, to validate the practical applicability and real-world relevance of the developed model, it was implemented in an online real-time application designed for live testing and demonstration. This application successfully demonstrated the model’s robustness and reliability in real world scenarios, emphasizing its potential for integration into automated monitoring systems in livestock farming to enhance animal welfare and productivity through the timely detection of alarming vocalizations. The results underscore significant advancements in audio classification methodologies and demonstrate the potential of semi-supervised learning frameworks in improving classification accuracy, offering valuable insights for future research and practical applications in automated livestock monitoring with the help of Artificial Intelligence.