Key points are not available for this paper at this time.
To ensure effective and safe human-robot interaction in rehabilitation or performance enhancement using robotic devices, it is crucial to accurately estimate human motion. However, achieving high-precision joint angle measurement using surface electromyography (sEMG) has remained a challenge. In this study, a novel deep learning framework that combines Graph Convolutional Neural Networks (GCN) with Long Short-Term Memory (LSTM) networks to predict joint angles is proposed. The framework utilizes the correlation matrix to capture spatial correlations between nodes, and LSTM networks are employed to capture temporal correlations between nodes. The proposed deep learning network framework on a self-built dataset to assess its performance is validated. Experimental results demonstrate that the joint angle prediction model proposed in this study effectively predicts the elbow joint angle under different load conditions. Considering the spatial correlation between sensors significantly improves the accuracy of joint angle prediction. Compared with other network models, such as Bi-LSTM, TCN, and CNN-LSTM, the proposed elbow angle prediction algorithm achieves the highest fitting degree of 95.3%, with a minimum mean square error (MMSE) as low as 0.0197.
Teng et al. (Thu,) studied this question.