Learning system for myoelectric prosthetic hand control by forearm amputees

This paper presents a novel learning system for myoelectric prosthetic hand control by forearm amputees using estimations of continuous joint angles. Wavelengths calculated using surface electromyogram (sEMG) signals of forearm amputees are input into a neural network (NN); past inputs are also used to take finger dynamics into consideration when estimating the metacarpophalangeal joint angles of each finger and wrist joint angle of pronation/supination and palmar flexion/dorsiflexion. The learning system has a three-step learning dataset generation process: (1) continuous motion of a virtual prosthetics hand (VR-hand) and motion timing bar are displayed to a subject; (2) the subject contracts his/her muscles following the VR-hand motion; and (3) sEMG signals and joint angles of VR-hand are measured and saved as the learning dataset. This system does not need to measure actual joint angles. To demonstrate the effectiveness of this learning system, RMS error of joint angle estimations are presented in cases of a motion set with 8 patterns for a healthy subject, and a motion set with 4 patterns for a right forearm amputee.