Continuous estimation of wrist torques with stack-autoencoder based deep neural network: A preliminary study

Abstract

The continuous estimation of kinematics or kinetics from electromyography (EMG) signals is essential for intuitive control of prostheses and other human-machine interfaces based on bioelectrical signals. In this preliminary study, we concentrate on the continuous estimation of wrist torques under isometric contraction of three separate degrees-of-freedom (D-oFs) with a stack-autoencoder based deep neural network. With this kind of deep neural network, features used for regression could be extracted autonomously other than in hand-crafted manner. Five subjects participated in the experiment under a visual feedback guide interface, in which surface EMG signals and wrist torques were concurrently recorded. It is shown that a promising estimation performance is achieved in all three DoFs. The outcomes of this study demonstrate the feasibility of this method on continuous estimation of wrist torques and reveal the potential for further being extended into continuous and simultaneous myoelectric control.