44 / 2025-03-31 20:41:57

A Time-Frequency Feature Fusion Model for Cattle Behavior Recognition Using Transformer and Wavelet Transform

Time-frequency domain,Wearable sensor,Behavior recognition

3. 智慧牧场与动物福利 > （2）智慧养殖的大数据、机器学习和物联网应用

Automated animal activity recognition has advanced significantly through deep learning approaches, offering critical support for Precision Livestock Farming by enhancing production efficiency and reducing operational costs, yet existing methods face challenges in accurately identifying fine-grained behaviors and generalizing across individual variations due to their insufficient integration of temporal and frequency information along with limited temporal modeling capabilities. To address these limitations, this study proposes a novel framework integrating wavelet-based time-frequency analysis with temporal dependency modeling, where continuous wavelet transform captures localized dynamic patterns from wearable device-derived multi-axis acceleration data while a transformer architecture with learnable positional encodings models global temporal relationships, specifically designed to overcome the critical defect of separate processing of time-domain and frequency-domain features in conventional methods through coordinated time-frequency feature fusion. The framework is rigorously validated through leave-one-out cross-validation to ensure robust generalizability across individuals, with comprehensive experimental evaluations against state-of-the-art approaches conducted in the study, including detailed comparisons of classification performance and adaptability metrics, all thoroughly analyzed and discussed in the paper to demonstrate the effectiveness of the proposed methodology for automated behavioral monitoring in livestock management.