Abstract:

[Objective] Feeding behavior is widely used to monitor the physiological status of dairy cows and supports early disease warning in herd management. Among various behavioral indicators, jaw movement stands out as the most direct external sign of feeding, providing a physically interpretable signal for automated recognition. Traditional recognition models for dairy cow feeding behavior, however, suffer from low accuracy and high hardware overhead. These limitations make them difficult to deploy on resource-constrained edge devices commonly found in pasture environments. To address these issues, a classification and monitoring approach is proposed that pairs a custom wearable device with lightweight ensemble machine-learning models. The goal is to build a lightweight model that balances recognition accuracy with on-device deployability, so that the system can run in real time on the hardware available at typical livestock farms. [Methods] A compact wearable unit integrating a tension sensor and a six-axis inertial measurement unit was fixed to the lower jaw of each cow. This device enabled synchronous acquisition of jaw opening-closing dynamics and head posture variation throughout natural feeding bouts. The feature extraction process was organized around the physical properties of the captured signals. Specifically, features were computed across four complementary dimensions. Time-domain statistics characterized amplitude and temporal distribution. Frequency and time-frequency representations captured rhythmic chewing patterns. Spatial correlation metrics encoded the inter-axis coordination of head motion. Trajectory-synthesis descriptors summarized the geometric path of jaw movement over time. The resulting 110-dimensional feature space preserved biologically meaningful signal structure across all four dimensions. Because many of the 110 features were redundant, a performance-feedback-driven restricted backward elimination strategy (PF-RBE) was designed and validated for dimensionality reduction. Conventional recursive feature elimination (RFE) ranked features by a single importance score and discarded the lowest-ranked ones in bulk, which risked removing features with limited individual contribution but meaningful joint effects. PF-RBE addressed this problem by using feature importance only as a search heuristic. The actual retention or elimination of each feature was determined by measuring the change in classification performance after each ablation step. Through recursive evaluation, this strategy retained key features more precisely and effectively balanced classification performance against feature dimensionality. The process yielded a compact 59-dimensional core feature subset. Four ensemble algorithms, Random Forest, XGBoost, LightGBM, and CatBoost, were then evaluated under multiple time-window lengths to determine how temporal context length affected recognition accuracy across different model architectures. [Results and Discussions] Recognition performance improved in a stepwise fashion as the time window was extended. The XGBoost model was trained on the PF-RBE-selected 59-dimensional feature subset with a 60-second time window. Through systematic comparison across different time windows, the 60-second window was established as the optimal behavioral perception window size. Notably, this window was able to completely describe the behavioral habits of dairy cows. Full feeding cycles including transitions between sub-behaviors were captured, and recognition accuracy improved substantially. Among all candidate models, CatBoost and XGBoost achieved the highest accuracies of 93.71% and 93.64%, respectively. Further comparison showed that, under the 60-second window configuration, the XGBoost algorithm maintained competitive classification performance while requiring fewer computational resources. Its model file was only 4.48 MB and its total inference time was merely 3.925 6 s. XGBoost offered an advantage over the other algorithms by combining competitive classification accuracy with low computational cost. The F₁-Score reached 92.81%, and was adopted as the core evaluation metric under class-imbalanced conditions. [Conclusions] Under conditions of limited computing power and storage, XGBoost still maintained high recognition accuracy. Its compact model size and low inference cost were well suited to the hardware constraints typical of edge-wearable devices deployed on livestock. Overall, the proposed approach provides a feasible technical reference for the deployment of edge computing and real-time behavioral monitoring in smart pasture applications.

Key words: dairy cow feeding behavior, behavior monitoring, lightweight machine learning models, feature dimension reduction, time window optimization