| Abstract
Surface electromyography signals (sEMG) are known to be nonstationary and
highly noisy [1]. Hence, classification of the signal in relation to human motion can be
difficult. Based on other studies (reviewed later), it is believed that there exists a
correspondence between sEMG signals and human limb motion. This research explores
this relationship.
The study goals of this work contribute to the funded research in which Idaho
State University (ISU) is participating, ARWED. The project targets stroke victims who
suffer partial loss in their motor abilities. The goal is to develop an augmented reality
device (ARWED) that helps to train and retrain patients.
A robust approach to identify the intended motion using only sEMG signals is
achieved. Previous research in the field succeeded in classifying sEMG in relation to
kinematical variables. Using kinematical variables provide more information, which, as a
result, makes classification and analysis an easier task. Instead, this study targeted the
mapping of sEMG signals without provided information about the kinematical variables.
Artificial Neural Networks (ANN), along with seventeen features, are able to map
out the intended motion. The data of twenty participants are collected for the purpose of
validating the achieved results. For each participant, a separate neural network is trained
to predict the intended motion based on information extracted from different sEMG
signals.
The proposed approach is tested on twenty individuals and five different finger
motions. It is valid in deducing the five motions. The performance (mean square error) of
xvi
the different training algorithms is compared. Levenberg-Marquardt (LM) Algorithm
with an average performance of 0.0461, Bayesian Regularization (BR) Algorithm with an
average performance of 0.03105, and Scaled Conjugate Gradient (SCG) with an average
performance of 0.1398 have all shown reliability in identifying the different types of
motion. The linear correlation between the predicted and the target outputs is also
computed. The scores are 0.9898 for LM, 0.99305 for BR, and 0.96753 for SCG. |