Rafael Ávila dos Santos
, Eric Ferreira Schmiele
, Leomar Santos Marques
, Ricardo Rodrigues Magalhães
, Danton Diego Ferreira
& Bruno Henrique Groenner Barbosa
Abstract: Electromyogram (EMG) signals have been extensively utilized for the control of robotic prosthesis movements. In this context, feature extraction and feature selection play a pivotal role in classifying EMG signals into different intended movements. Although time-domain descriptors, such as mean absolute value, zero crossing, slope sign change, waveform length, and sample entropy, are widely used, their discriminative capacity may be limited under noisy conditions. This study evaluates features extracted from higher-order statistics (HOS), discrete wavelet transform (DWT), wavelet packet decomposition (WPD), and time-domain (TD) descriptors for EMG-based movement classification. In addition, four feature selection techniques are compared: mutual information, ReliefF, mRMR, and Boruta. The experiments were conducted using MLP, KNN, SVM, and RF classifiers, considering both original signals and signals corrupted with Gaussian noise at different SNR levels. The results indicate that cumulant-based features are highly robust in noisy scenarios and that selecting the 25 most relevant features can substantially reduce dimensionality while preserving or improving classification performance.
Keywords: Electromyogram, Robotic Prosthesis, Cumulants, Feature Selection, Wavelet Transform, Higher-Order Statistics.
DOI code: 10.21528/lnlm-vol24-no2-art1
PDF file: vol24-no2-art1.pdf
BibTex file: vol24-no2-art1.bib
