Feature Extraction And Pattern Classification Of Electromyographic Signals

The electromyographic signal (EMG) is the record of the neuromuscular activity of human skeletal muscles. Such recorder provides a reliable window on the motorial system of human. The surface EMG (SEMG) signal, which is acquired using the surface electrode, is widely used in prosthetic control since it provides noninvasive access to the basic physiology of motor control and muscle contraction. In the last decade, EMG-controlled prosthesis has appealed great research interests from a wide range of areas such as clinical diagnosis, rehabilitation engineering, and the emerging area of biomechatronics.This thesis investigates the algorithms for EMG feature extraction and pattern classification, which are essential for the control of dexterous myoelectric hand with multiple degrees-of-freedom. The EMG signals are acquired by four surface electrodes placed on four different muscles from the forearm.The thesis presents a comparative study on different feature extraction algorithms, including the time-domain analysis, frequency-domain analysis, and time-frequency analysis (e.g., the wavelet and the wavelet packet).The primary contributions of the thesis are two novel algorithms for EMG feature extraction. The first is based on the vector multi-variants auto-regressive (MVAR) model. It has been shown that, with some appropriate algorithm for dimension reduction (e.g., Fisher Linear Discriminate Analysis), the MVAR model based feature achieves higher classification accuracy than the scalar AR model. This result indicates that efficient algorithms for capturing the spatial coherence between multi-dimensional EMG signals may improve the performance of the EMG decoders. The second algorithm is based on parameterized bispectrum, with the feature parameters are obtained from the axial and radial integration. The average classification accuracy of the algorithm for the eight motion(supination,pronation,fist clench,fist stretch,wrist up,wrist down,wrist inward,wrist outward) amounts to 98.8%, which is competitive to the best results reported in the literature. The algorithms have been implemented and will be incorporated into the biomechatronic interface of the SJT-2 prosthetic hand.