Long-Sequence Lower Limb Action Recognition via Action-Specific Attention and Logic Correction for Daily Rehabilitation.

Accurate and robust intent recognition is the cornerstone of human-machine collaboration in rehabilitation, particularly for controlling Supernumerary Robotic Limbs (SRL). However, recognizing long sequences of continuous lower-limb movements remains challenging due to severe category imbalance and the absence of biomechanical logical constraints. This paper proposes a hybrid deep learning framework, named ASAM-BiLSTM-MC, which employs a multimodal fusion strategy combining surface electromyography (sEMG) and inertial measurement unit (IMU) data. First, a sliding-window-based Action-Specific Attention Mechanism (ASAM) is integrated into a BiLSTM network. This dynamic weighting strategy significantly enhances the model's sensitivity to sparse yet critical transitional actions, such as the sit-to-stand movement. Second, a Markov Chain (MC) based logic correction module is introduced to post-process predicted sequences, utilizing prior state transition probabilities to eliminate physiologically implausible transitions. Using offline subject independent evaluation, the proposed model achieved an overall accuracy of 96.17%(±1.27%) and a Macro F1 Score of 94.40%(±1.58%). Compared to baseline models, recognition performance for critical transitional actions improved significantly (e.g., squat-to-stand increased by 11.74%). Furthermore, in a preliminary out of distribution evaluation, the model achieved 87.32% accuracy on elderly subjects excluded from training and demonstrated enhanced safety and smoothness in SRL tracking control experiments. This study addresses the reliability limitations of purely data-driven models in complex rehabilitation scenarios, offering new insights for safe and precise human-machine collaborative control.