Towards Dexterous Manipulation through Motor Learning and Biomechanical Modelling

Abstract

Myoelectric controlled prosthetic hands represent an effective tool to restore functionality and enhance the quality of life for upper limb amputees. Such devices provide sensing, multifunctionality and more natural control. In the current state of the art solutions, the control is mainly accomplished through sophisticated motion encoding by using machine learning algorithms for residual forearm muscles. Offline analysis and evaluation of motion detection accuracy for such algorithms on data sets are the main focus of current studies. However, there is a significant gap between laboratory evaluations and system integration in the complicated real-time environment. Because a sufficient and comprehensive analysis of complete prostheses requires a sophisticated synchronisation of data acquisition, motion classification, and timely prosthetic actuation with a wearable compact system, most prosthetic control lack the robust interface to facilitate all required functionalities in an acceptable manner for the majority of users. Even if advancements in data integration and computational power enable high prediction accuracy, the practical implementation of such technology is still being challenged by various influences, particularly those related to the fact that the signal sources are biological signals that change considerably by limb position, variations on muscle contraction, electrode shifting and amputation level. Therefore, most of the existing prostheses are passive, and their dexterity properties remain fixed with limited object grasping and hand gestures. This research presents the design of a bypass socket and integrated real-time control system based on pattern recognition algorithms to control a prosthetic hand. This study covers a compact system development beginning with investigating the anatomy and natural dexterity of the human hand, motor control, and human-like physical manipulation for data collection, going through the sEMG feature extraction and finally implementing adequate embedded pattern recognition on a prosthesis prototype. A wide range of techniques such as sEMG signals, data gloves, and force sensors was employed to collect data from able-bodied subjects. Popular pattern recognition algorithms such as k-Nearest Neighbours (k-NN), support vector machines (SVM), linear discriminant analysis (LDA) and artificial neural network (ANN) were used to differentiate individual finger manipulations and hand motions. The performance of classifiers with different muscle observation approaches and a variety of feature extraction methods with two windowing sizes and the various number of the electrode was compared against the publicly available data sets and similar studies. The offline analysis results led to a novel bypass socket design to minimise electrode shifting, causing difficulty to use during model training and inconsistencies between users, which increases motion detection errors between the desired and performed motions. New electrode arrangement by socket prototype ensured the transmission of the most significant input from all muscles and standardised data acquisition between sessions, particularly considering the real-time conditions with a limited source of signals to stump and dynamic arm orientation. It provides a sufficient approximation to pattern recognition since it resists elbow rotation and provides immense practicality to achieve an intuitive embedded system. A combined dynamic data acquisition and control approach that yields high accuracy and robustness were implemented as the final strategy and tested in real-time with an able-bodied subject. The development of control architecture is based on how humans maintain control stability during dynamic arm orientation over time, particularly in different amputation levels. The system performance was tested with real-time evaluation metrics such as motion completion rate, motion detection accuracy, reach and grasp experiments and timing of the syste

Date of Award	1 Aug 2022
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Zhenmin Zou (Supervisor), Tingting Mu (Supervisor) & Lei Ren (Supervisor)