Online proportional myoelectric control of a humanoid shoulder motions using electromyogram signals

This paper proposes a two phase strategy for proportional myoelectric control of Surena 3 humanoid robot which benefits from strength of two common myoelectric control methods, Pattern recognition base and simultaneous proportional control, for improving joint angle estimation. The aim of this research is to present a human-robot interface to create a mapping between electrical activities of muscles known as electromyogram (EMG) signals and kinematics of corresponding motion. First phase concerns with motion classification using Quadratic Discriminant Analysis (QDA) and Majority Voting (MV). Several common motion classification algorithms and feature vectors including time domain and frequency domain futures were investigated which lead to QDA and a superior feature vector with more than 97% classification accuracy. The second phase concerns with continuous angle estimation of shoulder joint motion classes using Time Delayed Artificial Neural Network (TDANN) with overall accuracy of 90% R2. QDA serves as a high level controller which decides between four TDANN correspond to each shoulder motion classes. QDA and TDANN models trained with several sets of offline data and were tested with online dataset. Online and offline data estimation accuracy and model robustness against disturbances show a significant improvement compared to similar methods in this field.