Behavior classification and rhythm analysis of pigs based on wearable sensors

Assessing the health status of pigs is critical for the swine industry, as the timely detection of abnormal behaviors can improve both production efficiency and animal welfare. To address the current challenges associated with wearable sensor data in pigs, such as indistinct behavioral features, irregular behavior durations, and the difficulty of manual classification, this paper proposes a behavior classification method based on ear tag data to detect three primary behaviors in pigs: feeding, movement, and sleeping. First, we collected 3-D accelerometer data from the ear tags of 40 pigs over a 60-day period. Then, using a sliding window approach, we transformed the accelerometer data for different behaviors into short-time Fourier transform (STFT) spectrograms, time-amplitude graphs, scatter plots, and density plots at various time scales (0.5 seconds, 1 second, 10 seconds, 30 seconds, 60 seconds, 120 seconds, and 180 seconds). Finally, we employed a spatial attention-enhanced ResNeSt-101 model to classify the images representing pig behaviors. The experimental results showed that time-amplitude graphs provided superior behavior classification performance compared to other image types, with the 60-second time-amplitude graphs achieving the highest accuracy at 91.23%. Statistical analysis of pig behavior using ear tag data indicated that the primary feeding times were between 6:00-10:00 and 14:00-18:00, with a total daily feeding duration of approximately 1.5 hours. This study introduces a novel technological approach for pig behavior detection and health assessment, with promising potential to enhance production efficiency and animal welfare.