A Multi-Attention Fusion Deep Network for Human Activity Recognition in Motor Rehabilitation

With the rapid development of wearable sensors and intelligent rehabilitation devices, sensor-based human activity recognition (HAR) has become an essential technology for motor rehabilitation assessment and assisted therapy. Accurate recognition of patients’ movement states and activity patterns during rehabilitation training is crucial for personalized rehabilitation planning, rehabilitation outcome evaluation, and remote rehabilitation monitoring. However, due to large inter-patient variability, irregular movement execution, and complex sensor noise, existing deep learning-based HAR methods still suffer from limited recognition accuracy and generalization capability in rehabilitation scenarios. To address these challenges, this paper proposes a deep learning-based human activity recognition framework for patient-oriented motor rehabilitation, incorporating a multi-attention fusion mechanism. The proposed model adaptively captures the importance of different sensor channels, temporal dependencies, and feature representations through hierarchical attention modules, thereby enhancing the discriminative capability for key rehabilitation-related motion patterns and improving recognition performance under complex movement conditions. Extensive experiments are conducted on multiple public inertial sensor-based HAR datasets to validate the effectiveness of the proposed approach. The experimental results demonstrate that the proposed method consistently outperforms existing state-of-the-art approaches in terms of classification accuracy and overall recognition performance, exhibiting strong robustness and generalization ability. The results indicate that the proposed model provides an effective technical solution for motion perception and intelligent assessment in motor rehabilitation processes, offering significant potential for practical applications in intelligent rehabilitation systems and remote rehabilitation monitoring.