Myoelectric Prostethic Hand Control Based On Neural Network Using Electromyography

Abstract

Through years, myoelectric prosthetics have been studied and developed with the aim of
controlling assistive devices for the user based on their intended movements. Pattern recognition and signal
classification based on electromyography signals (EMG) are promising for hand prosthetic devices.
However, the prediction and robustness offered by scientific researches is still not sufficient for many real
life applications. Inaccuracy in classifying the signals lead to a slow-response system which is unlikely
wanted by the end users. Thus, with the aim of having a highly accurate system, in this work, a prosthetic
hand is developed based on two algorithms models, which are feed forward- backpropogation and recurrent
neural network. Surface electromyogram (EMG) signals obtained from the skin surface represent the
muscles activities. They are used as the input signal in the developed models and the predicted results are
send to actuators in order to move and generate the desired movement for the user. In order to monitor and
evaluate the system, a GUI has been designed to showcase the predicted signals from the trained model. By
using 5 extracted features; MAV, RMS, LogD, SD, and Waveform, the accuracy of the developed models
are more than 70%. The predicted results show the accuracy of recurrent neural network is 94% while feed
forward - backpropogation is 73%. In conclusion, the developed model using recurrent neural network has
given higher accuracy and better performance in the system. Hence, this indicates that the recurrent neural
network model is more applicable for motion classification.