Knowledge Distillation for Efficient Human Activity Recognition on Resource-Constrained Smartwatches

Human activity recognition (HAR) using smartwatch sensor data is a crucial enabling technology for healthcare monitoring, fitness assessment, and assistive applications. Deploying advanced deep learning architectures on wearable devices with limited computational capacity, however, poses significant challenges. This work presents an efficient HAR solution based on knowledge distillation, enabling compact models to retain high classification accuracy while satisfying smartwatch hardware constraints. The proposed framework adopts a teacher–student paradigm in which high-capacity neural networks—specifically CNN, LSTM, and CNN-LSTM models—trained on multimodal sensor data transfer their learned representations to a lightweight CNN student model optimized for resource-limited environments. Multi-sensor fusion from accelerometer, gyroscope, and magnetometer measurements, commonly available on commercial smartwatches, is incorporated to enhance recognition capability. Evaluation is performed on two benchmark datasets: the CHA dataset, containing 13 daily activities from 10 participants, and the WISDM-HARB dataset, comprising 18 activities from 51 participants. Results indicate classification accuracies of 97.78–98.34% on CHA and 91.03–93.04% on WISDM-HARB, along with storage reductions of up to 85.6% and parameter count decreases of up to 56.9% compared with the original teacher models. These findings demonstrate the practicality of knowledge distillation for delivering accurate, real-time HAR on resource-constrained wearable platforms.