Comparison of sEMG-Based Feature Extraction and Motion Classification Methods for Upper-Limb Movement.

Area of Science:

• Biomedical Engineering
• Robotics
• Human-Computer Interaction

Background:

• Surface electromyography (sEMG) is crucial for intuitive control of prostheses and robotic arms.
• Current motion recognition methods face challenges in real-time accuracy and computational time.

Purpose of the Study:

• To evaluate eight combinations of four feature extraction methods (RMS, DFA, WP, MM) and two classifiers (NN, SVM).
• To determine optimal sEMG signal processing for mapping upper-limb motions.
• To identify feature-classifier pairings for real-time and non-real-time applications.

Main Methods:

• sEMG signals were recorded from six upper-limb muscles of seven subjects.
• Four feature extraction techniques (RMS, DFA, WP, MM) were analyzed.
• Two classifiers, Neural Networks (NN) and Support Vector Machine (SVM), were compared.
• Performance was assessed based on motion recognition accuracy and time consumption.

Main Results:

• NN achieved 88.7% accuracy in training; SVM achieved 85.9% in real-time tests.
• WP showed 97.7% accuracy in training; MM achieved 94.3% in real-time tests.
• SVM was faster in training, NN was faster in real-time computation.

Conclusions:

• The combination of Muscular Model (MM) and Neural Networks (NN) is recommended for real-time sEMG control.
• The Muscular Model (MM) with Support Vector Machine (SVM) is suitable when computational time is not a primary concern.