Research On Feature Extraction Method And Application Of SEMG Signal Using In Rehabilitation Robotic System

Surface electromyography(sEMG)signal was one kind of complex bio-electric signals,generated by the accumulation of the electrical nerve stimulation on the skin surface due to muscle contract.In order to introduce sEMG into the robotic system,feature extraction method specific for that usage was one of the key techniques.Focusing on the application and research of the sEMG feature extraction method,combining with the demand of the rehabilitation robotic system,the main work of the improved twice feature extraction method in this thesis was concluded as below.(1)In order to extract the feature of muscle force,aimed at the instability of the sEMG signal,this thesis established the phenomenon model of the sEMG signal and adopted the Kalman filter based variance estimation to extract the nonrandom features named muscle force related variable.(2)Aimed at the complexity of the force change,the combination of the time domain,parametric model and time-frequecy domain extraction methods were used to extract muscle force-motion features from the first features.Due to the complexity of the upper limb motion,muscle force-motion features were not suit for describing the motion pattern directly.Aimed at this problem,Nonnegative Matrix Factorization(NMF)was used to compress the original features by self-learning of feature basis.Aimed at the slow convergence and deformity point of zero,alternating nonnegative least square based NMF using projected gradient(PGNMF)was used to fasten the convergence and prevent the deformity point.(3)In order to recognize the upper limber motion,Extreme Learning Machine was used to build the motion-muscle force feature based classifier of the motions.Aimed at the uncertainty in the practical situation,TSSVD-ELM was used to improve the stability of the classifier.(4)Software for sEMG signal processing was programmed and related experiments were conducted.Aimed at the continuous recognizing,the active segmentation method was proposed.At last,experiments of offline data simulation and online test showed a satisfied result of the feature extraction structure and methods proposed by this thesis.