Hand and Finger Movements Detection by time-frequency Electromyography (EMG) Signals Processing and Machine Learning.

Abstract:

Detecting the movement of various body parts, including finger movements, requires modeling natural body movements. Since finger movements cause the contraction of arm muscles, electromyography (EMG) sensors can be used to measure these contractions, due to the ability of EMG to reflect human motor goals. Recently, human-computer interfaces have gained considerable popularity in various fields. These interfaces help computer systems convert biological information from the human body into understandable data. Among the signals used in human-computer interfaces are surface electromyography (sEMG) signals. This technology can be useful in various areas, including assisting individuals with motor disorders who may have difficulty interacting with their environment. Additionally, designing upper limb prosthetics using the output from this processing allows amputees to use this capability as part of their body to interact with the surrounding world.This thesis utilizes two databases available in the EMG DATASETS REPOSITORY. This database was compiled by Khoshaba and colleagues and contains hand-finger movements in fourteen and fifteen different classes. They recorded this database using 8 electrodes with a sampling frequency of 4000 Hz. This research aims to present an efficient method for classifying finger movements using EMG signals. In this work, time-frequency decomposition methods based on three approaches, Discrete Wavelet Transform (DWT), Maximum Overlap Discrete Wavelet Transform (MODWT) , and Short-Time Fourier Transform (STFT) with a minimal number of features. In the proposed algorithms for each of the three approaches, commonly used statistical features in this area were extracted from the decomposed components. Two feature selection methods, Fscnca and Relieff, were used for optimization and classification with fewer features. Then, for finger movement classification, SVM classifiers with various kernels were employed. The results of the classification show that the proposed approach based on feature extraction using DWT in the 14 and 15 class icons with 25 top features selected by the Fscnca method achieved an accuracy of 98.20% and 99.27%, respectively. These values indicate promising results for improving the process of detecting hand and finger movements.