Robust animal activity recognition using wearable sensors: A correlation distillation-based information recovery method toward data having low sampling rates

Automated animal activity recognition (AAR) has succeeded dramatically due to recent sensing technologies and deep learning advances, enhancing animal health and welfare. Since animals need to be monitored over a long period, factors influencing the energy consumption of sensing devices must be considered carefully. As a critical factor, the sampling rate greatly affects energy usage, battery life, and data storage. To reduce energy costs, existing works often lower sampling rates. However, when the sampling rate falls below a limited threshold, the recognition performance would degrade rapidly due to missing many relevant signals. Therefore, this study proposed a novel correlation distillation-based information recovery (CDIR) method to improve the performance of AAR at low sampling rates. Specifically, we took two convolutional neural networks trained using data having higher and lower sampling rates as the teacher and student models, respectively. The CDIR enabled the student to mimic correlations within teacher features, facilitating missing information recovery. To evaluate its effectiveness, we conducted experiments on a public dataset acquired from six horses using tri-axial accelerometers and gyroscopes with 100 Hz. Data having low sampling rates were obtained by down-sampling the original data at different frequencies (e.g., 50, and 25 Hz). The experimental results demonstrated that our CDIR remarkably boosted the model trained on low sampling rate data (e.g., percentage-point increments in the precision, recall, F1-score, and accuracy of 2.98%, 3.22%, 3.05%, and 1.89%, respectively, for the 12.5-Hz data) while outperforming the existing KD algorithms. This inspired the development of energy-efficient animal monitoring systems.