Towards an Effective EMG-based Neuromuscular Interface for Human-robot Interaction

Abstract

In recent years, the requirements of individual assistant systems for elderly and disabled people are daily increasing, as well as the function expansion of prosthetic control, military, residential and commercial robots. In this case, human-robot interactions have become a popular research area. Since these robots are directly interacted with the users, there are several challenges in the design and control of such human-robot interaction technology. Electromyography (EMG) signal is the electrical signals of the human body, which contains a wealth of information on human action and can be used to determine the user's intent. The purpose of this thesis is to develop an EMG-based human-robot interface, which can identify the body's response by signal processing and model calculations, and can also transform the response into the motion control instructions, and control the robot to complete the body movement intentions. The existing physiological models have provided a continuous motion prediction method. This method of the 'simplified musculoskeletal model' took the mechanical revolute instead of human joint, the straight line instead of skeleton, and the straight segment between the muscle starting point and adhesion point instead of the muscle. During the complex motion of human body, the prediction accuracy of this model is greatly reduced since it is not close to the human actual physiological structure. Also, it cannot be used for the calculation when the muscular force line crosses the joint center. Currently, the studies of the impact of physiological model parameters to the sensitivity of interface have three problems: the amount of assessed parameters was few, the evaluation method was single, and the results of different researches had disagreement. Especially, the analysis of overall parameters in the neuromuscular model was less. The existing sensitivity evaluation was focused on the impact of musculotendon parameters sensitivity to the model. Through two cases study of elbow flexion/extension and forearm pronation/supination, this thesis overviews the new progresses that aim to address the existing gaps in this research field. The elbow joint was selected to implement a new method of muscle modeling, which could improve the accuracy of model during the complex motion of the elbow, while ensuring the real-time processing of the interface. The forearm rotation was chosen because of the weak EMG of forearm muscles, the short moving time and small changes in muscle length. The interface for forearm rotation has its particularity. A new EMG-driven elbow physiological model has been developed to predict the elbow flexion and extension. In the process of modeling, this thesis made assumptions based on the physiological properties of muscle. Through the elbow experiments from a plurality of subjects and a variety of movements, the model’s ability of accurately predicting different moving trajectories was verified. The model was also implemented and verified by a single degree of freedom (DOF) exoskeleton. A new EMG-driven physiological model for forearm pronation/supination has been established. It can predict the forearm continuous rotation movement by the EMG activations from the superficial part of three muscles. The model contained a unique physiology musculoskeletal model. The experiments from four subjects showed the effectiveness of this method. The establishment of this forearm physiological model has opened up a new way for the prediction of complex joint system with small amplitude motions. A new sensitivity assessment method of model parameters, three-step layered approach, has been established. By using this method, this thesis analyzed the characteristics of the model parameters. A relatively small subset of the parameters was generated for parameter tuning. This method provided a new way of thinking for the parameters sensitivity analysis. The purpose of parameter tuning is to make the model can precisely match every subject. This thesis programmed two kinds of evolutionary algorithm - Differential Evolution (DE) and Genetic Algorithm (GA), and experimentally compared their performances in three aspects. Because of the high accuracy and fast convergence capability, DE can be used for fast online tuning. And GA can only be used in offline tuning. A controller based on the fusion of EMG and force information has been proposed to validate the proposed models in real time control environment. A 5-DOF upper limb exoskeleton was developed by the Medical and Rehabilitation Research Group at the University of Auckland, the exoskeleton was used to evaluate the effectiveness of the EMG based controller (EBC). The results showed that the dynamic auxiliary effect of the exoskeleton is obvious (the decrease of muscle activation could be ensured above 52% when the assistance works), and the physiological model based EBC can adapt to different individuals. This also showed the effectiveness and online adaptability of the EMG-based Neuromuscular Interface proposed by this thesis.