Research On Pattern Recognition Technology Of Upper Limbs Based On Surface EMG Signal

Surface Electromyography(sEMG)is a kind weak bioelectricity signal accompanied by muscle activity.It has the advantages of easy acquisition,strong anti-interference and high cost performance.Rich information contained in it can well characterize the degree of muscle fatigue and the specific action intention of human bodys.It is widely used in clinical testing,prosthesis control,rehabilitation engineering and other fields.While the number of limb dyskinesia people increasing rapidly due to mechanical injury,stroke and other causes,many causes such as other reasons labor shortage have forced the rehabilitation training as a social problem.In this study,the shoulder joint of upper limbs,which has more DOF(degree of freedom)and shorter communication distance with the brain,is studied.The key technologies of signal processing,feature extraction and motion recognition for upper limbs self-rehabilitative training using surface EMG were mainly studied,intended to provide a practical interactive program for upper limb rehabilitation training system based on surface EMG.The main research work of this paper is as follows:(1)Signal acquisition and processing.Based on the preliminary study of the preparation mechanism and characteristics of EMG signal,the construction of signal model and the selection of pick-up electrodes,combined with the knowledge of human anatomy,sport medicine,rehabilitation and so on,we have figured out the relationship among the designed six classic upper limb action modes,muscle groups and sEMG.Otherwise,we have determined the placement of eight electrodes.And,the collected eight-channel EMG signals are compared and analyzed.The four-channel signals(CH1,CH3,CH6 and CH8),which are distinguished from the rest,were selected as signal sources for subsequent research,such as filtering and sifted windowing,endpoint detecting and other pretreatments.(2)Feature Extraction and analysis.At first,the characteristics of EMG signals in time-domain,frequency-domain and time-frequency-domain are analyzed.After that,we extract four varieties of time-domain features(MAV,SSC,WL and ZC)and two types of frequency-domain features(MPF and MF),and do some statistic analysis of certain samples based on the filtered signals.The quality of the characterization is further studied.Finally,the excellent eigenvector groupn(MAV?SSC?WL?MPF?MF from CH3?CH6?CH8)between different action modes is selected according to the law of "high cohesion and low coupling" for the follow-up pattern recognition algorithm research.(3)Pattern recognition and exploration.Firstly,the classical statistical classifier(LDA linear decision,Boosting,Bayes decision method,SVM),fuzzy classifier,neural network classifier and other related methods for surface EMG signal are summarized.We focus on the two classification models of BP neural network and deep belief network(DBN),and compare the performances of them from the aspects of recognition rate,iteration times,stability of variance and so on.The average recognition rate for six motions is up to 92.55%based on DBN.At the same time,the influence of the related factors such as the composition of eigenvectors,the amount of modes and electrodes on the pattern recognition results is further explored.The experimental results show that the DBN performs better than BP neural network in various aspects.