Research On The Classification Of Surface EMG Based On Online Learning Algorithms

The surface EMG(sEMG)signal is a kind of bioelectric signal which is the manifestation of skeletal muscle contraction and can be captured by a series of surface electrodes.The signal is easy to gain from a noninvasive approach.sEMG is widely used in the fields of biomedical engineering for rehabilitation and intelligent robot control.Unfortunately,it is far away from the clinic use because of some limitations in the artificial limb control field.The aim of this study is to promote the use of sEMG driven biomimetic limbs for clinic applications.This thesis applies the state-of-the-art off-line learning methods and perceptual model based online learning algorithms to classify the signal under different sampling periods.The main contents of this thesis are:(1)The mechanism of the generation of sEMG signals was analyzed.The method of surface electrode placement was studied,and the sEMG signal acquisition procedure was described in detail.(2)Several methods of feature extraction for sEMG signals were discussed,including time domain analysis,frequency domain analysis,time-frequency domain analysis and parametric model method.The variability of the characteristic elements obtained by different feature extraction methods in different actions was analyzed.Finally,the eigenvector of the surface EMG signal was composed from the time domain characteristic signal and the autoregressive model coefficients.(3)In order to demonstrate that the traditional classification model had unstable performance in the time-segment signal,this thesis designed BP neural network,support vector machine and linear discriminant analysis method,as the classifiers to discriminate the extracted eigenvectors,and statistically compared these models in the same period.It was identified that there existed difference in terms of classification accuracy and was aware that traditional classification model cannot achieve high accuracy in inter-session test.(4)This thesis discussed six kinds of online learning algorithms based on perceptual machine model,and proposed an on-line motion recognition framework for surface EMG signals.By collecting the surface EMG signals collected in multiple time periods,it was concluded that the on-line learning method based on the perceptron could solve the situation of accuracy degradation over time.And it was also found that the soft confidence-weighted learning algorithm could achieve outstanding performance.At the end,the thesis summarized the whole paper and prospected future works.