Control of an Externally Powered Prosthetic Forearm Using Raw-EMG Signals

Abstract

This paper proposes a new motion discrimination method using raw EMG signals to improve control performance of a prosthetic forearm. This method uses a novel recurrent neural network based on a well developed hidden Markov model. The proposed network can model a time sequence of EMG signals using recurrent connections, and different two processes such as filtering and a pattern discrimination are unified together and realized in a single network. Weight coefficients of the network are regulated by the back-propagation through time algorithm.
In the experiments, five subjects which include two amputees performed control of the prosthetic forearm. We confirmed that the proposed method could cope with time-varying characteristics of EMG signals and could achieve considerably high discrimination accuracy compared with the previous methods. Response of the discrimination result to the input EMG pattern was also improved using the proposed method.