EMG signal analysis with machine learning, topological data analysis and deep learning models

In this thesis study, it is aimed to analyze Electromyography (EMG) signals effectively. In this regard, three different approaches were used, including traditional machine learning algorithms, topological data analysis and deep learning-based algorithm methods. In the first method, different features are extracted from EMG signals. Classification was carried out using individual, double, triple and quadruple combinations of these features. Additionally, among the 128 sensors, sensors with low electrical activity information were identified and eliminated. 99.75% accuracy was achieved in the selected 120 sensor and binary feature combinations. In the second method, the topological structure of the EMG data was analyzed by calculating the barcode diagram, persistence diagram and simplex values. When 0th dimension lifetimes were used as features, 70.26% accuracy was obtained. In the third method, EMG data was converted to images using the continuous wavelet transform method and feature extraction was performed using pre-trained deep learning models. The obtained features were given as input to the capsule network and the classification process was carried out. A success rate of 97.32% was achieved with this method. At the same time, correlation heat maps obtained from EMG data were used as images and given as input to deep learning models and subjected to classification.