sEMG-based recognition of wrist-hand movements using a composite transformer-LSTM model

Surface electromyography (sEMG) signals are commonly utilized as a control source for upper-limb rehabilitation robots. Fast and high accuracy recognition of sEMG-based movements is crucial for the real-time control of robots. Here, we report a composite transformer-long short-term memory (LSTM) model for recognizing wrist-hand movements from sEMG signals. The sEMG data were collected to construct the training dataset. The consistent dataset was generated by a sliding window method to evaluate the model performance. We investigated the impact of window size and window increment on the sEMG-based recognition of wrist-hand movements. The results indicated that the recognition accuracy of the minimum window of 150 ms and the maximum window of 300 ms reached 94.68% and 99.35%, respectively. The recognition speeds were less than 0.3 ms. Finally, the trained transformer-LSTM model was combined with a majority voting strategy to determine the control outputs. The results confirm the model's capability for real-time and high-accuracy control with minimal delay. This method holds substantial significance for rapid feedback control of rehabilitation robots.