A Physiological Model Driven Neuromuscular Interface for Exoskeleton Assisted Rehabilitation

Abstract

Exoskeletons are anthropomorphic robotic devices that move in concert with their users to enhance their power and endurance capabilities. They are actively involved in exercising, reinforcing and moving patient limbs for rehabilitation or to assist in activities of daily living. However, exoskeletons were originally designed for fully-abled users and patients have difficulty controlling exoskeletons because they lack the physical capability. Therefore, an interface that can determine the user's intent to move is needed and this thesis aims to create such an interface with the electromyography (EMG) signal. The EMG signal results from the electrical activity within muscle and can be measured from the skin surface, without physical movement. A neuromuscular interface (NI), is defined as a module that consists of all the hardware and software components required to obtain, filter, process and output predictions of intended joint movement based on EMG signals. Existing interfacing methods based on the EMG signal are currently dependent on pattern recognition techniques that are only able to identify discrete states of movement. Physiological model-based approaches offer a means of continuous movement prediction, but these methods commonly require additional mass to enhance the EMG signals, the use of additional sensor data, or the simplification of movements beyond functional practicality. The approaches are also centred around hinge joints or joints that have been simplified to a single degree of freedom (DOF). There have been few attempts to address the complexities of multiple DOF joint systems. Developing a model for multiple DOF joints requires a better understanding of the superficial muscles that actuate the joint because these limit the accessibility of EMG signal sites. Therefore, bettering the understanding of the relationship between the superficial muscles of a joint and the available DOFs is crucial for effective EMG-driven multiple DOF model development. Commonly used tuning methods and performance measures are also inadequate for EMG signal-based interfaces because the criteria for optimisation is not suited for movement prediction, and the best response does not always coincide with the most desirable response. This thesis outlines the developments that aim to address these gaps through two case studies: the elbow joint and the masticatory system. The elbow joint was selected because of its simplicity and the masticatory system was selected because it contains a highly complex joint system and a challenging structure. An EMG-driven physiological model was initially developed for the elbow joint. It uses only EMG signals and a unique neuromusculoskeletal structure to predict the continuous motion of the elbow joint. With tuning, the model was able to identify random movements from multiple subjects in offline analyses. The model was also realised as a physical interface and the concept of an NI was tested and proved in real-time through the control of a physical representation of the elbow joint. A new EMG-driven physiological model for the masticatory system was built that was able to accommodate two DOFs of mandibular motion. The determination of EMG channels and the appropriate DOFs was done through a specifically designed study to better understand the influences and characteristics of the superficial mandibular muscles. This resulted in a unique muscular arrangement and a model of the masticatory system that was capable of movement in the vertical and lateral directions. The physiological model of the masticatory system was also combined with an artificial neural network (ANN) to form a hybrid model. This addresses an identified problem with multiple DOF models that concern the multiple roles a single muscle plays in different movements. The purpose of the ANN was to identify the type of movement occurring and then enhance or suppress the output from the physiological model. This new dynamism allowed continuous movement prediction with less interference between the available DOFs. The physiological models were all validated experimentally using data from multiple subjects. However, existing tuning approaches were found to be insufficient and would produce results that were not always in the best interests of the intended application. Thus, a serial tuning approach was proposed that utilised a different performance measure, the correlation coefficient,in a sequential optimisation process. This new approach produced better movement trajectories than current popular methods. These contributions have led to three journal articles, three conference papers, one provisional patent, and have laid substantial groundwork for the development of effective physiological models and NIs. Future work involves refining and improving the modelling process, and beginning to focus on the hardware components to complete the realisation of an NI.