A Motion Intent Recognition Method Based on Dynamic Feature Fusion and Multi-Stream TCN

Clinical motor relearning theory emphasizes the active participation of patients, so the recognition of patient's motor imagery plays a key role in robot-assisted rehabilitation exercise training. We constructed an upper limb rehabilitation sEMG (Surface Electromyography) dataset by capturing 18 upper limb rehabilitation movements using a wireless sEMG acquisition system and proposed a deep learning model based on dynamic feature fusion with a multi-stream Temporal Convolutional Network (MS-TCN), which identified the patient's motor intention from the captured sEMG by end-to-end learning. On the one hand, the method can extract the temporal features of single-channel signals through the multi-stream TCN structure, and on the other hand, dynamic feature fusion can effectively utilize the primary-secondary relationship of multiple muscles in completing different motions, achieving high recognition accuracy and low response time. Through simulation analysis and experimental testing, the results show that this method can achieve higher classification accuracy than mainstream motor intention recognition algorithms on different subjects, and the effective use of this method will promote the recovery of patient's limb function and significantly improve the efficiency of rehabilitation training.