Feature Extraction And Classification Recognition Of SEMG For Myoelectirc Prostheses

SEMG is the ideal control signal source with the advantage of non-invasive measurement and good bionics. Because the methods of feature extraction and classification recognition aren’t mature, SEMG isn’t widely used. The important problem is that how to extract the effective features of SEMG and achieve high accuracy classification of movements in practical process of myoelectirc prostheses. SEMG is collected from upper extremity movements with biceps brachii, triceps brachii and deltoid muscle in this paper. Though the wavelet threshold de-noising, feature extraction and classification recognition, the classification rate of four movements is achieved high accuracy with crank web, arm thrust, vertical arm internal rotation and vertical arm external rotation.Firstly, the wavelet threshold de-noising method is analyzed. However, there are discontinuous points and constant deviations in the hard and soft wavelet threshold methods, according to those problems, the improved threshold method is used a transition of the nonlinear function and added a control parameter to deal with the threshold function. Using three threshold de-noising methods to de-noise SEMG, experimental results show that the improved threshold method possesses merits of hard and soft threshold methods. It increases signal to noise ratio. Meanwhile, after the control parameter of improved threshold de-noising method is tuned by numerical analysis method, the higher signal to noise ratio of SEMG is achieved. This step is prepared for feature extraction and classification recognition.Secondly, the time-domain feature, AR model coefficients and the maximum value of each level wavelet coefficients modules are extracted and compared after de-noising and active segment testing. Experimental results show that the latter two features have good separability for SEMG. Using the method of single feature combination in this paper, AR model coefficients and the maximum value of each level wavelet coefficients modules are combined to be feature vector and used the PCA to reduce dimension. The classification rate of combined feature which is proposed by this paper is compared with the classification rate of single feature and other combined features by the BP neural network classifier and the support vector machine classifier, experimental results show that combined feature which is proposed by this paper has better characterization capabilities for SEMG and the correct classification rate is higher than single feature and other combined features, and it is up to97%.So, SEMG is de-noised by improved threshold method and tuned by numerical analysis method, the signal to noise ratio of SEMG is higher. Then extracting the combined feature including AR model coefficients and the maximum value of each level wavelet coefficients modules and using BP neural network and the support vector machine to classify, the aim is achieved with high accuracy rate of different movements.