EMG methods for prosthesis ankle-subtalar free-space control

Abstract

EMG-based prosthetic joint controllers have been an active research field for more than fifty years. However, several challenges remain to be addressed[9]. Electrodes positioning, controllers calibration, and controllers’ linear approximation error are the most challenging problems among them.

This thesis introduces three methods to solve those problems respectively. They are 1. A non-negative blind source separation algorithm named non-negative orthogonal decomposition(NOD). This algorithm aims to replace non-negative matrix factorization(NMF) for muscle motion base extraction. NOD recovers the source signal by finding the borders of the input signal and translating the borders onto the coordinate axis. The translated signals are the recovered signals. 2. An unsupervised algorithm for generating joint trajectories from EMG signals in reciprocating movements. The EMG signal and trajectory can be used for EMG-based prosthesis joint controller calibration. And 3. an innovative EMG-to-joint-position controller. It uses neural networks to compensate for the nonlinearity of the well-known bilinear model[2].

The NOD algorithm successfully extracted motion bases from the EMG signals. Compared with NMF, the motion bases are more independent and stable. The minimum-jerk-based trajectory generator generated smooth and biomimetic trajectories on intact subjects. The trajectories are close to the ground truth collected from the goniometer. The third model also has considerable improvement in joint angle accuracy over the linear muscle model.