AWARE-Net: A Lightweight Joint Optimization Framework for Robust Sensor-Based Human Activity Recognition

Sensor-based human activity recognition (SHAR) serves as a core research direction in pervasive computing, mobile health, and related fields. Although existing deep learning methods have achieved promising progress in SHAR tasks, most optimize from a single dimension only. They struggle to simultaneously balance recognition accuracy, noise robustness, adaptation to class imbalance, and lightweight deployment requirements, leading to performance bottlenecks in real-world scenarios. To address these challenges, this paper proposes a lightweight joint optimization framework named AWARE-Net. Leveraging the lightweight TS-ResNet as a backbone encoder, the framework integrates spatiotemporal dynamic convolution feature encoding with a global loss function that fuses class-balanced loss, contrastive learning auxiliary loss, and temporal smooth regularization to achieve multi-objective joint optimization. Extensive experiments on three widely used SHAR benchmark datasets, namely OPPORTUNITY, PAMAP2, and USC-HAD, demonstrate that the proposed AWARE-Net achieves competitive performance compared with representative state-of-the-art HAR methods.