Data Rate Optimization for Power-Efficient Animal Behavior Classification
Board Location: #98
Discipline: Computer Sciences and Information Management
Subcategory: Computer Science & Information Systems
Session: 2
Eddy Figueroa Picon - University of Puerto Rico - Arecibo
Co-Author(s): Huiying Chen, New Mexico State University, Las Cruces, NM;Andres Perea, New Mexico State University, Las Cruces, NM;Mehmet Emin Bakir, New Mexico State University, Las Cruces, NM;Santiago Utsumi, New Mexico State University, Las Cruces, NM;Huiping Cao, New Mexico State University, Las Cruces, NM.
In extensive livestock production systems, the detection and classification of animal behaviors are crucial to monitor animal health and welfare. IoT devices, equipped with embedded accelerometers, capture motion data as ‘activity count’, which can be used to feed Machine Learning (ML) models for animal behavior prediction (Kahan et al., 2020). However, the trade-off between data volume required for accurate predictions and sensor power consumption poses significant challenges to battery lifetime, especially in outdoor production systems with limited infrastructure to manipulate animals and replace sensor batteries frequently.This study evaluates the performance of ML models in predicting animal behavior at different data rates while optimizing power efficiency by reducing data volume. We mounted 24 sensors on rangeland beef cows, transmitting one activity count per minute. Over a three-month observation period, we collected approximately 50,000 data packets and labeled 9,222 of them as grazing, walking, or resting. We performed simulations to quantify the reverse correlation between data collection frequency and sensor battery life: 289 days at one minute, 338 days at two minutes, 359 days at three minutes, 370 days at four minutes, and 391 days at five minutes. We assessed ML performance on the complete one-minute interval dataset and progressively subsampled datasets obtained by selecting every second, third, fourth, and sixth-minute data points, representing 50%, 33%, 25%, and 10% of the original dataset, respectively. Each dataset was partitioned into an 80% training set and a 20% testing set. We tested various popular ML algorithms, including Logistic Regression, Support Vector Machines, Random Forest, and XGBoost. XGBoost consistently outperformed other algorithms, achieving the highest F1 scores. Despite systematically reducing the training data size from 100% to 10%, the F1 score remained resilient: 0.94 with the entire dataset, 0.89 with 50% data, 0.87 with 33%, 0.84 with 25%, and 0.83 with 10%. Downsizing data to 10% results in roughly a 35% increase in battery life, with about a 12% decrease in accuracy. The results underscore the potential for downsizing data to reduce device power consumption and data transmission while preserving the robust predictive capabilities of ML algorithms. These findings hold particular significance for resource-constrained environments.In conclusion, our findings indicate a promising path towards optimizing data collection frequency, thus reducing power consumption while maintaining robust prediction accuracy. This research lays the groundwork for future studies to expand this balance to encompass a wider range of data types and ML approaches.
Funder Acknowledgement(s): This study was funded by the USDA-NIFA-AFRI’s Sustainable Southwest Beef Coordinated Agricultural Project grant #12726269 and contributions from the National Science Foundation (NSF). The research has been approved and endorsed by New Mexico State University Computer Science Department.
Faculty Advisor: Dr. Huiping Cao, hcao@nmsu.edu
Role: Take ownership of testing and implementing the different Machine Learning algorithms and reducing the sample size to reduce power consumption while maintaining robust prediction accuracy.