Cross-Species Behavioral Representation Learning Using Domain-Adversarial Adaptation on Wearable IMU Signals.

Animal locomotion exhibits highly structured temporal dynamics despite significant inter-species biomechanical and anatomical discrepancies. While wearable inertial measurement unit (IMU) sensors and deep learning have advanced animal activity recognition, existing systems remain largely species-dependent, requiring large-scale labeled datasets for each deployment. To address this, we propose a biomimetic cross-species behavioral representation learning framework that extracts transferable locomotor structures from heterogeneous IMU signals. The proposed methodology integrates behavioral ontology harmonization, imbalance-aware augmentation, and semi-supervised domain-adversarial adaptation to reduce inter-species distributional discrepancies. Unlike conventional classification models, our dual-head architecture enables simultaneous processing of multi-label and single-label behavioral structures across anatomically diverse species. Extensive experiments conducted on dog, goat, and horse datasets demonstrate that the proposed framework significantly improves cross-species transferability, achieving a mean Macro-F1 score of 0.711 compared to 0.345 for direct transfer learning. Furthermore, we show that sparse target supervision is critical for stabilizing adversarial adaptation. K-Means head adaptation partially mitigated fully unsupervised negative transfer. However, its performance remained below semi-supervised Domain-Adversarial Neural Network (DANN), indicating that sparse target supervision is still necessary for stable cross-species adaptation. These findings suggest that biologically motivated locomotor similarity can support cross-species behavioral transfer, although external validation on additional species, sensor placements, and deployment conditions is still required.