Evaluation of feature extraction techniques and classifiers for finger movement recognition using surface...

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Electromyography (EMG) signals are crucial for controlling bio-driven systems like prosthetic hands.
• Effective signal processing, including preprocessing, dimensionality reduction, and classification, is essential for accurate control.
• Classifying complex EMG signals from multiple finger movements presents a significant challenge.

Purpose of the Study:

• To propose and evaluate a novel system for classifying six-channel EMG signals corresponding to 14 distinct finger movements.
• To compare the efficacy of various feature extraction techniques and classifiers for EMG signal analysis.
• To identify the optimal combination of feature extraction and classification for high-accuracy prosthetic hand control.

Main Methods:

• A system was designed to classify six-channel EMG signals from 14 finger movements, generating a 66-element feature vector.
• Six feature extraction techniques were evaluated: PCA, LDA, ULDA, OFNDA, SRLDA, and SRELM.
• Seven classifiers were assessed: SVM, LC, NB, KNN, RBF-ELM, AW-ELM, and NN.

Main Results:

• The combination of Spectral Regression Linear Discriminant Analysis (SRLDA) for feature extraction and Neural Network (NN) as the classifier achieved the highest classification accuracy.
• This optimal SRELM-NN combination demonstrated a classification accuracy of 99% for the 14 finger movements.
• The SRELM-NN approach significantly outperformed all other tested feature extraction and classifier pairings.

Conclusions:

• The SRELM-NN system provides a highly accurate method for classifying EMG signals for prosthetic hand control.
• This research advances the development of sophisticated bio-driven assistive devices.
• The findings highlight the potential of advanced machine learning techniques in improving human-machine interfaces.