Kinaesthetic learning through sEMG signal classification

Abstract

To learn kinaesthetically how to grasp or manipulate objects, a key part is to understand the mechanisms involved in the performance of movements of the digits. A human being has an elaborate system that involves the brain sending electrical signals to the skeletal muscles of the body which activate the fibres of the muscles in the body and flexion, or extension of the limb can be performed. These electrical signals allow a researcher to extract the intentions of the human in such skills as grasping or object manipulation and use the biological data to train contemporary neural network algorithms to classify the electrical signals extracted during the movements and potentially design a novel system of control.

 This thesis presents a comprehensive exploration of the potential of the utilisation of the myoelectric signals captured from the surface of a human’s skin above a muscle while it is being activated to generate a movement. A popular technique known as surface electromyography (sEMG), has been employed in this thesis. sEMG is a non-invasive method of measuring the electrical activity of the muscles. Using sEMG signals and recurrent neural networks (RNNs) are demonstrated to perform the classification of finger movements and grasps with accuracies of 90% and above. This thesis employs a low-cost wearable sensor that allows for raw sEMG data to be extracted in real world environments and allow for more realistic experiments to be carried out.

 Significant contributions in the field of sEMG signal classification are presented in this thesis. These contributions exhibit improvements in the state-of-the-art in this field by employing novel neural network architectures and hybrid algorithms that have vii shown to improve the process of signal classification. These type of improvements to the performance of sEMG signal classification could have potential benefits for kinaesthetic teaching of robotic systems or teleoperation of robotic manipulators. An example application would be if someone wanted to teach an assistive robot that operates in a domestic environment how to manipulate or grasp objects, they could don a wearable sEMG sensor i.e., Myo gesture control armband, and by simply picking the object up and relay the sEMG information to the robot it could learn how to recreate the movements without the need for manual programming.

Thesis is embargoed until 31 August 2026

Date of Award	Aug 2024
Original language	English
Supervisor	Emmett Kerr (Supervisor) & Nazmul Siddique (Supervisor)