GADF-enhanced global-local feature network for wearable sensor-based cow behavior recognition

Continuous and reliable monitoring of dairy cow behaviors is a key component of data-driven herd management systems. Although wearable inertial sensors provide a practical alternative to camera-based monitoring, effectively representing complex motion patterns from raw inertial signals remains a challenge.

This work presents an Inertial Measurement Unit (IMU)-based cow behavior recognition approach that emphasizes multi-scale motion representation using nose-mounted sensing. Acceleration data acquired from a nose-attached inertial measurement unit are utilized to capture characteristic head movement patterns associated with feeding, rumination, and locomotion activities. Instead of directly processing one-dimensional time-series signals, the inertial sequences are restructured into two-dimensional pseudo-color matrices through Gramian Angular Difference Field (GADF) transformation, allowing temporal correlations to be explicitly preserved. To exploit motion information at different levels of granularity, a multi-scale feature modeling network is developed, wherein complementary motion cues are extracted and integrated via a cross-attention-driven interaction mechanism.

Experimental results demonstrate that the proposed framework achieves competitive recognition performance, with an overall accuracy of 90.12%, and maintains robustness across different behavioral categories.

These findings confirm that transforming inertial signals into GADF-based pseudo-color images, coupled with multi-scale feature modeling, offers an effective and robust solution for automated cow behavior recognition.