A Study Of Multifunctional Human Machine Interface With Use Of Electromyographic Signal

Surface electromyographic (sEMG) signals, which record the electrophysiological activities of muscle fibers during voluntary contractions, are widely used as the inputs of human-machine interface (HMI) in electrically-powered prosthetic hands, referred to as myoelectric control. The primary function of HMI is to discriminate the patterns of sEMG signals and therefore to generate the commands for the myoelectric control of hand/wrist movements. In order to control a multi-functional and dexterous prosthetic hand, it is required that the HMI should be capable of decoding multiple patterns of hand/wrist-movement from sEMG signals with desirable accuracy rate. This dissertation investigates advanced algorithms for feature extraction and classification of sEMG signals. The main results are outlined as follows:In order to extract the higher-order statistical information existed in sEMG signals, we adopt a biseptrum transform to analyze the signals. The bispectrum integration and Fisher linear discriminant are used to extract features from bispectrum matrices. We use a Davies Bouldlin index and different classifiers to investigate the effect of higher order information on movement recognition. The experimental results demonstrate that the bispectrum feature is superior to other first or second-order features in terms of class-separating capability and classification accuracy.Based on the frequency information of sEMG signals, we propose a Fourier-derived cepstral (FC) feature to identify hand/wrist movent patterns, the computational process of which mainly consists of the fast Fourier transform and discrete cosine transform algorithms. Fisher-ratio feature selection is used for dimension reduction of FC coefficients. In addition, we propose a feature-level post processing method to improve separating capacity of features. The experiments show that the discriminant FC features have not only high classification accuracy but also fast computing speed. Aiming to analyze transient sEMG signals during the initial phase of muscle contractions, we propose a new sEMG signal processing method based on smooth localized complex exponentials ( SLEX ) and local discriminant basis. This algorithm is able to extract the time-evolutionary features from sEMG signals by dividing the signals into several localized segments.In order to adapt to the time-varying characteristics of sEMG signals, a self-enhancing classifier is investigated. An adaptive algorithm for updating the parameters of a classifier during testing time is incorporated in the stationary classifier to improve its classification performance. The primary advantage of the proposed classifier is that the updating algorithm makes the classifier self-evolutionary, which is able to enhance the adaptability of the HMI.In summary, the gist of this dissertation is to develop effective approaches to sEMG feature extraction and classification through the combination of theoretical studies and practical works. The experimental results demonstrate that the proposed EMG pattern recognition methods significantly improve the classification accuracy of sEMG signals containing multiple movement patterns. In addition, the proposed sEMG pattern recognition approaches are embedded into a DSP based hardware platform. It lays a foundation for the future application of myoelectric control.