Continuous Prediction of Hand Joint Angles Based on sEMG and Tf-LSTM

To enhance the compliance control accuracy of myoelectric prosthetic hands, this study proposes a Transformer-LSTM joint angle prediction model based on surface electromyographic (sEMG) signals. Surface electromyographic signals were preprocessed using the DB8 subset of the publicly available Ninapro dataset, and three features—root mean square, mean absolute value, and waveform length—were extracted as inputs for the model. This model maps nine types of finger movements to predicted angles of five finger joints, and fits these predictions with actual values recorded by a data glove. The evaluation metric R ²was used to compare this hybrid model with Long Short Term Memory and Support Vector Regression models. The Transformer-LSTM model outperformed the latter two, showing an improvement of 10.69% and 29.97% in decoding performance, respectively. Thus, the model demonstrates a robust capability in estimating hand joint angles, providing a foundation for future multi-joint control in myoelectric prosthetic hands.