Sound identification of abnormal pig vocalizations: Enhancing livestock welfare monitoring on smart farms

• Published in: Information processing & management 2024-07, Vol.61 (4), p.103770, Article 103770
• Main Authors: Xie, Yutao, Wang, Jun, Chen, Cheng, Yin, Taixin, Yang, Shiyu, Li, Zhiyuan, Zhang, Ye, Ke, Juyang, Song, Le, Gan, Lin
• Format: Article
• Language: English
• Subjects: Computational acoustics; Smart farm information system; Sound identification
• ISSN: 0306-4573
• DOI: 10.1016/j.ipm.2024.103770

•Proposed an acoustic recognition-based early warning system for animal welfare.•Processed and cleaned audio data; established and published implementation criteria.•AST framework elevates livestock vocalization identification with detailed global audio nuances.•AST framework demonstrates high accuracy in real datasets, ideal for dynamic farms.•System applied in two farms in China, the world's largest pig farming nation. In smart agriculture, analyzing animal vocalizations provides a non-invasive, continuous monitoring approach that directly correlates with specific animal conditions, enhancing welfare. Pig vocalizations, in particular, are critical for managing farm events and improving animal welfare. Yet, traditional methods, mainly involving Convolutional Neural Networks (CNNs), focus on local audio features and complex combinations, struggling with varying audio lengths and high computational costs. Addressing these issues, this study introduces a novel approach with an Audio Spectrogram Transformer (AST), designed to detect abnormal pig vocalizations. Our method involves a two-stage process: segmenting the audio to retain only valuable information, and classifying through an attention mechanism that analyzes both nuanced and global audio features. This technique significantly improves accuracy and efficiency in vocalization analysis. Tested on 7600 real-world audio samples, our method demonstrated a significant improvement, achieving an accuracy of 93 % and enhancing inference speed by 19 times compared to existing CNN-based techniques. Additionally, we conducted interpretability analysis and feature selection experiments to evaluate the efficacy of different feature combinations. These experiments verified that our attention-based approach not only simplifies the input features but also provides superior performance over traditional models. The findings of this study underscore the potential of AST in transforming livestock welfare monitoring by offering a more accurate, efficient, and scalable solution.