Continuous Estimation of Joint Angle with Surface Electromyography (sEMG) and Deep Learning

In this paper, we presented a novel attention-based convolutional neural networks (CNN)-long short-term memory (LSTM) models using surface electromyography (sEMG) signal to estimate continuous joint angle of upper limb. 7 subjects were recruited to participate in the experiment. Joint angles were generated using the simulation software-OpenSIM as the training label. Then, the model's performance was evaluated under different motions by using Root Mean Square Error (RMSE) and Cross Correlation (CC). Our study shows that the proposed model can achieve the excellent performance, with a good agreement with experimental values for all cases. After that, the prediction performance of the model was compared with and without the preprocessing of window approach. It is found that, using window approach leads to a reduction in RMSE of 4.7-7.2 deg and an improvement in CC of 1.7-3.8%. In conclusion, our proposed model has an accurate continuous estimation of hand movements, offering a promising solution for the human-machine interaction and cooperation.