Surface electromyoagraphic (SEMG) signal acquisition-analysis-classification for the real-time control of a robotic hand

• Κύριος συγγραφέας: Νικολαΐδη, Βασιλική-Θεοφίλη
• Άλλοι συγγραφείς: Nikolaidi, Vasiliki-Theofili
• Γλώσσα: English
• Έκδοση: 2022
• Θέματα: Myoelectric signals; Robotics; Pattern recognition; Real time; Random forests; Code generation; Onset detection; Shifted skew log laplace; Features; Teager-Kaiser operator; Prosthesis; Arduino; Teensy; Ensemble classification; Genetic optimization; Feature evaluation; Dimensionality reduction; Ηλεκτρομυογραφικά σήματα; Ρομποτική; Αναγνώριση προτύπων; Πραγματικός χρόνος; Παραγωγή κώδικα; Κινητική πρόσθεση; Ελάττωση διαστάσεων; Πλειοψηφική ταξινόμηση; Γενετική βελτιστοποίηση; Αξιολόγηση χαρακτηριστικών

This thesis engages in a rapidly growing scientific topic - the modelling of a robotic arm that can be controlled by myoelectric signals from an individual's muscles, which could be used as an interface for robot teleoperation or motion control of a prosthetic limb. The purpose of this thesis is to achieve gesture recognition of finger gestures from sEMG signals and do a feasibility study on the real-time control of a robotic hand, using low cost off-the-shelf hardware. Myoelectric signals were collected from four healthy subjects, by placing the MyoWare boards on four muscles of the forearm to detect predetermined hand gestures. The subjects were then instructed to do the predetermined gestures: moving each finger separately, making a fist, two finger pinching and grasping. These data were first processed and then classified using different supervised learning classification algorithms. Emphasis was given to achieve real time response for the robotic hand controlled by the sEMG signals. For the acquisition, analysis and processing of the sEMG signals, as well as the control of the robotic hand, a single Teensy 4.0 board was used. A new family of Teager-Kaiser Operator (TKO) based algorithms is proposed that can enhance the onset(the sEMG signal start indicating muscle contraction) and offset(the sEMG signal end indicating muscle relaxation) detection. The tested algorithms are modified versions of popular onset detection algorithms that achieved also offset detection using the proposed TKO based probability and likelihood estimation. New features that can be used for the better real-time classification of sEMG signals are also proposed. These features proved useful and outperformed commonly used features in the literature. The recognition accuracy achieved for the training set reached up to 100% in simulations using a 500ms sample window. The recognition accuracy during the hardware tests with novel signals, using a 150ms sample window, reached up to 70%. The robotic hand implemented for the realisation of the thesis was taken from the InMoov open source project and was 3D printed using PLA filament. The hand is actuated by five Tower Pro MG 946R digital servos. The developed code can be split in two main parts. The first part that deals with the analysis, simulation and training of classification algorithms is developed mainly in Python and a small part in Octave. The second part involves the robotic application of moving a robotic hand in real time based on sEMG signals and was developed in C++ 11. Finally using the csnake and MicroML Python libraries the aforementioned two parts of the code were connected by using the first part to automatically generate C++ classifier modules to be used in the latter part. Up to five such modules can be seamlessly included in the robotic application and be used in an ensemble (voting) classification scheme.