Methods Of Feature Extraction And Gait Recognition For Lower Extremity Based On Electromyographic Signals

Surface electromyography (sEMG) is generated by active motor units during skeletal musclecontraction at random and can be picked up by the surface electrodes. With the correlation betweenmuscle activity status and motion functional status, the EMG signals contain the combination anddecomposition related movement units, the movement intentions and the wealth of information ofphysiological state of the neuromuscular system, which is widely used in prosthetic control,human-computer interaction, bone parameter prediction, and diagnosis of neuromuscular diseasesand other aspects. Because the differences between different walking movements can be reflectedby the EMG signal characteristics, the study of these features can help to identify and analyze thegait state.Based on the experimental sEMG data from the human lower limbs, the methods of the gaitfeature extraction and recognition are studied in this dissertation. By using the methods of the signalde-noising, feature extraction and classification, pattern recognition, the effective recognition of thelower limb gait are achieved. The main contents are as follows:(1) For the nonlinear, non-stationary and non-periodic characteristics in EMG, a featureextraction method is designed by the multiscale entropy, and a measure of the complexity of surfaceEMG is achieved on different scales. Meanwhile, the sample sets of the feature vector are obtainedby extracting the time-domain characteristics of the integrated EMG and variance values.(2) In view of the support vector machine (SVM) parameter optimization problem, a particleswarm optimization-support vector machine (PSO-SVM) classifier is constructed, with the PSO forSVM parameter optimization choices, the SVM parameters of the penalty and the kernel functionare optimized. By comparing the results of the gait recognition for before and after optimization, theresults show PSO-SVM optimization methods are significantly improved gait recognition rateswhile taking into account the accuracy of the classification and adaptability.(3) For the specific and individual differences in gait problems, by the statistical analysis fromthe collected actual gait data, the characteristic of the non-uniformity (different time-length) gaitcycle for the human is studied, and through data groups with the different gait cycle length, theeffects of recognition rate are discussed in detail.The problems of the gait feature extraction and the pattern recognition for the lower extremityare systematically studied in this dissertation. By the collection of EMG experimental data andanalysis of the identification algorithm, the time domain and nonlinear features in EMG are efficiently extracted; by the parameter optimization method, the gait recognition rates of SVM aresignificantly improved; the characteristics of the non-uniformity gait cycle and its effects on therecognition results are also discussed. These conclusions have significant application importance indetection and evaluation of the human lower limb motor function, as well as in research anddevelopment of the intelligent rehabilitation and assistive equipments.