A teacher-to-student information recovery method toward energy-efficient animal activity recognition at low sampling rates

Abstract

Automated animal activity recognition (AAR) has advanced greatly through recent advances in sensing technologies and deep learning, and improved livestock management efficiency, animal health, and welfare monitoring. In practical automated AAR systems where animals need to be monitored over a long period, the sampling rate dramatically affects the energy consumption and battery life of sensing devices due to continuous data collection and transmission. Considering real-world benefits, existing works have often lowered the sampling rate to reduce energy costs. However, when the sampling rate falls below a threshold, the AAR performance degrades rapidly due to many relevant signals being missed. Therefore, this study proposed a novel method, dubbed teacher-to-student information recovery (T2S-IR), to improve the performance of AAR at low sampling rates. This approach effectively leverages the knowledge obtained from high-sampling-rate data, to assist in recovering the missing information in features extracted by the classification network trained on low-sampling-rate data. The workflow of the T2S-IR contains two main steps. (1) we utilize high-sampling-rate data for training teacher classification and reconstruction networks sequentially. (2) Then, we train a student classification network using low-sampling-rate data, while promoting its performance by exploiting the knowledge learned by trained teacher networks via two novel modules, namely the reconstruction-based information recovery (RIR) module and the correlation-distillation-based information recovery (CDIR) module. Specifically, the RIR module employs the pretrained teacher reconstruction network to enforce the student classification network to learn complete and descriptive features. The CDIR module enforces the feature maps of student network to mimic internal correlations within feature maps of pretrained teacher classification network along temporal and sensor axes directions. To validate our proposed T2S-IR, we conducted experiments on two public datasets acquired for horses and goats using triaxial accelerometers and gyroscopes with an initial sampling rate of 100 Hz. Data having low sampling rates were obtained by downsampling the original data at different frequencies (i.e., 50, 25, 12.5, 10, 5, and 2 Hz). The results demonstrated that our method remarkably boosted the classification network trained on low-sampling-rate data (e.g., percentage-point increments in the precision, recall, F1-score, and accuracy of 3.33%, 3.58%, 3.45%, and 2.19%, respectively, for the 12.5-Hz horse data and 7.6%, 4.44%, 6.9%, and 0.79%, respectively, for the 5-Hz goat data) while outperforming existing knowledge distillation methods. The enhanced classification network can be directly applied in practical AAR tasks with low sampling rates, significantly beneficial for scenarios with constrained energy sources for wearable devices.