GP-Regression for more stable predictions of an EMG-based HMI

Abstract

Prostheses and robots can be controlled using electrical signals in the muscles which is called myoelectric control. Pattern recognition in myoelectric control is a challenging field, since the underlying distribution of the signal is likely to change during the application. Various reasons like slight changes of the arm positions or different levels of muscular activation often lead to significant instability of the control. This work tries to overcome these challenges by enhancing the myoelectric human machine interface (mHMI) of the mobile robot EDAN, which is developed at the German Aerospace Center (Deutsches Zentrum für Luft- und Raumfahrt; DLR). This is achieved through a modification of the underlying regression model and by the introduction of novel adaptive models based on unsupervised as well as semi-supervised incremental learning approaches. The initial model of the mHMI of EDAN is build upon the commonly used Fully Independent Training Conditional sparse Gaussian Process approximation, which is exchanged in this work by the Variational Free Energy approximation. On this foundation the novel adaptive models integrate an interclass and intraclass distance to improve prediction stability under challenging conditions. A decrease in computational time as well as an increase in performance of the modified sGP is achieved both in an offline as well as an online experiment. The unsupervised adaptive model demonstrates the successful incorporation of incremental updates which is shown to lead to a significantly increased performance and higher stability of the predictions in an online conducted user study. Finally, the work tackles multiple challenges of myoelectric control, including simultaneous control, closed loop control as well as adaptability to the user.