Classification of Surface EMG Signals with Respect to Percent Maximum Voluntary Contraction Using Artificial Neural Networks

Primary Faculty Mentor’s Name

Rocio Alba-Flores

Proposal Track

Student

Session Format

Poster

Abstract

This paper presents a classification system based on Artificial Neural Networks (ANN) for the percent of maximum voluntary contraction (MVC) of surface electromyography (EMG) signals. Maximum voluntary contraction is the greatest amount of force a muscle can generate. EMG signals are electrical signals that are generated by muscle cells when the muscle is contracted. These signals are non-linear and susceptible to changes in the muscle therefore an adaptive system such as an artificial neural network is necessary to determine proper classifications. A MATLAB based simulator was used to generate the EMG signals. The simulator consisted of two models, one that replicates a single-fiber action potential to generate motor unit action potentials and a motor neuron pool to describe motor unit recruitment during isometric voluntary contraction. The system analyzes computer generated signals that replicate live signals from the bicep and then determines the percent of MVC. Nine different ranges were used to classify the signals. Three characteristic features such as; average rectified value, root mean square, and mean frequency were used as inputs to the system. These characteristic features were extracted from the EMG signals and were used as the input to different ANN structures. Several ANN structures were explored based on general guidelines for determining the number of neurons and layers. Also, two different training algorithms, scaled-conjugate gradient and Levenberg-Marquardt were tested with each structure for the optimum results. The systems were trained with a 455 sample set of data and then tested with two independent randomly generated sets of 50 samples. The ANN with a 16 neuron structure and a single hidden layer using the scaled-conjugate gradient algorithm proved to be the most effective in correctly classifying the signals. This process will be expanded to live signals with an end goal to develop an assistive robotic device for people with disabilities as well as the elderly.

Keywords

Electromyography, Artificial neural network, Maximum voluntary contraction

Location

Concourse/Atrium

Presentation Year

2014

Start Date

11-15-2014 2:55 PM

End Date

11-15-2014 4:10 PM

Publication Type and Release Option

Presentation (Open Access)

This document is currently not available here.

Share

COinS
 
Nov 15th, 2:55 PM Nov 15th, 4:10 PM

Classification of Surface EMG Signals with Respect to Percent Maximum Voluntary Contraction Using Artificial Neural Networks

Concourse/Atrium

This paper presents a classification system based on Artificial Neural Networks (ANN) for the percent of maximum voluntary contraction (MVC) of surface electromyography (EMG) signals. Maximum voluntary contraction is the greatest amount of force a muscle can generate. EMG signals are electrical signals that are generated by muscle cells when the muscle is contracted. These signals are non-linear and susceptible to changes in the muscle therefore an adaptive system such as an artificial neural network is necessary to determine proper classifications. A MATLAB based simulator was used to generate the EMG signals. The simulator consisted of two models, one that replicates a single-fiber action potential to generate motor unit action potentials and a motor neuron pool to describe motor unit recruitment during isometric voluntary contraction. The system analyzes computer generated signals that replicate live signals from the bicep and then determines the percent of MVC. Nine different ranges were used to classify the signals. Three characteristic features such as; average rectified value, root mean square, and mean frequency were used as inputs to the system. These characteristic features were extracted from the EMG signals and were used as the input to different ANN structures. Several ANN structures were explored based on general guidelines for determining the number of neurons and layers. Also, two different training algorithms, scaled-conjugate gradient and Levenberg-Marquardt were tested with each structure for the optimum results. The systems were trained with a 455 sample set of data and then tested with two independent randomly generated sets of 50 samples. The ANN with a 16 neuron structure and a single hidden layer using the scaled-conjugate gradient algorithm proved to be the most effective in correctly classifying the signals. This process will be expanded to live signals with an end goal to develop an assistive robotic device for people with disabilities as well as the elderly.