Application of Surface Electromyography based Pattern Recognition for Efficient Control of Upper Limb Prostheses

Abstract

The main aim of the hand prostheses is to help people restore human hand functions using artificial limbs. Electromyogram (EMG) signals have been used as a control signal, and this control scheme is referred to as Myoelectric Control (MEC). The conventional prostheses use a proportional control scheme based on the amplitude of the EMG signal. However, these schemes cannot achieve more than two degrees of freedom. This limited functionality is the key reason for the rejection of prosthesis by the amputees. If additional degrees of freedom are required, then Pattern Recognition (PR) based MEC offers favorable control. This research work aims at improving the classification accuracy of surface EMG driven pattern recognition (PR) system. Many factors affect the classification efficiency of PR based MEC. Significant challenges and practical limitations need to be addressed before making the PR scheme commercially available. The goal is to tackle these problems and to provide a solution using novel strategies developed in this research work. Surface Electromyogram (sEMG) signals are contaminated with a wide variety of noise, and this causes problems in PR. Noise sources such as power-line interference, motion artifact, ambient noise, characteristic instability of the signal, and noise due to electronic and recording equipment could be present in the sEMG signal. Noises can be decreased but cannot be removed totally by using high-quality equipment and intelligent circuits. Conventional filtering methods are commonly used to remove noise. But, if the noise from the recording instrument lies in the usable frequency range, it becomes hard to eliminate noise using conventional filters. In the pre-processing of sEMG signals, the challenge lies in the suppression of noise associated with the measurement and signal conditioning. The first contribution of the thesis is overcoming this limitation by proposing a novel pre-processing method. The method differentiates the original sEMG from noise using higher order statistics such as kurtosis, which is the fourth moment of distribution. The effectiveness of the method is demonstrated in terms of the improvement in PR performance. A significant number of studies have been performed on the various stages iiiof sEMG-based PR. There have been problems during the clinical implementation of the system even though the previous studies have reported a high classification accuracy of more than 90%. PR has shown great promise in predefined settings in laboratory conditions. The real-time factors which affect the performance have to be taken into consideration for PR to be commercially available. There are various other factors that also affect the performance of the PR system, such as variation in limb position, variation in forearm orientation, variation in electrode position, variation in force level, and change in the characteristics of the sEMG signal. It is becoming crucial to test the PR with these various factors due to the difference between ideal laboratory conditions and practical application of the MEC prostheses. The second contribution of the thesis is to address the robustness aspect of the PR-based control by developing a novel classification scheme that can function well under such changing conditions. Specifically, the focus was given to variations in force levels and wrist orientations. The proposed scheme achieved a significant improvement in classification accuracy when compared to the traditional method. To demonstrate that this research can be translated to clinical applications, study has been conducted on sEMG data set of upper limb amputees. This distinguishes the study from most of the previous studies done on non-amputee subjects. The findings of this work could improve the quality of life of amputees with better interaction to the outer world.