ANALYSIS AND INTERPRETATION OF ELECTROMYOGRAPHIC SIGNALS

Abstract

This study aims to improve the classification of surface electromyography (sEMG) signals by applying advanced denoising and feature extraction techniques. Initially, the raw sEMG dataset undergoes wavelet denoising using Symlet family (level 1, sym4) with universal thresholding to enhance signal quality and reduce noise. The denoised data is then processed to explore their efficacy in sEMG classification. In this study, a limited EMG dataset comprising 11 abnormal and 11 normal individuals is segmented using windowing techniques (i.e. adjacent and overlapping) which includes adjacent windows with sizes of 250 milliseconds, 1 second, 2 seconds, and 3 seconds, as well as overlapping windows with the same sizes and overlapping percentages of 25%, 50%, 60%, 70%, 80%, and 90%. Subsequently, eleven-time domain features—such as Mean Absolute Value, RMS, Variance, zero crossings, Slope Sign change, Willison Amplitude, Myopulse Percentage Rate, Difference Absolute Standard Deviation, Average Amplitude Change, Skewness & Kurtosis, are extracted from each window, resulting in comprehensive feature matrices. These matrices are then classified using various machine learning classifiers, including support vector machines (SVM), k-nearest neighbors (k-NN), decision trees, Ensemble Methods (Bagging, Boosting), and Naive Bayes, to evaluate their performance and accuracy. This approach yields accurate results concerning optimal overlapping percentages, window sizes, suitable machine learning algorithms, visualization of relevant features, and detection of the abnormalities using EMG Signals.