Neural-Driven Simultaneous Recognition of Gestures and Forces via Feature Enhancement for Prosthetic Control

The multi-degree-of-freedom prosthetic control based on surface electromyogram (sEMG) is critical to enable flexible interaction between upper limb amputees and the environment. However, gestures and force levels are variables of each other and interfere with the accuracy of simultaneous recognition. In this letter, we propose a neural-driven simultaneous recognition approach of gestures and forces based on the motor unit (MU) discharges. The high-density surface electromyogram (HD-sEMG) signals were decomposed into the motor unit spike trains (MUSTs), and then the neural-driven features were extracted from the MUSTs by calculating the discharge count of each MU. The geodesic flow kernel (GFK) algorithm was proposed to enhance the neural-driven features and improve the accuracy of simultaneous recognition. Two random forest (RF) classifiers were used to predict gestures and force levels from the enhanced neural-driven features. The proposed MU-GFK approach was tested on eleven gestures and three force levels. The average accuracies of gesture and force recognition were 96.26% and 99.92%, respectively, superior to the conventional sEMG method. Moreover, the GFK algorithm improved the classification accuracy of gestures and forces by 0.98% and 3.5%, respectively. These results demonstrate the feasibility and superiority of the proposed MU-GFK approach, promoting the intelligent upper limb prosthesis to have the ability to precisely and simultaneously control gestures and forces, thus achieving more natural and intuitive human-machine-environment interaction.