Research On Sampling Rate And Feature Extraction Of Surface EMG Signal For Gesture Recognition

The Electromyopgraphy(EMG)is a biomedical signal,which is acquired by the electrical response generated in muscles throughout its contraction symbolizing neuromuscular activities(contraction/relaxation).The signal is closely related to the state of muscle activity,and it reflects the functional state of the nerve muscle to a certain extent.It is easy to gain from a noninvasive approach and has a good biomimetic performance.Besides,sEMG is widely used in the field of assisted diagnosis.However,sEMG signal is very susceptible to environmental influences,and there are some problems in the use of actual scenes,such as weak robustness and unstable performance.In addition,the cost of acquisition equipment is high and it requires long-term maintenance.Therefore,it still has a great limitation for clinical medical application.The aim of this study is to improve and popularize the pattern recognition application based on sEMG signal(such as prosthetic hand control).In order to reduce the cost of acquisition equipment,the influence of electromyogram sampling rate on gesture recognition is explored.In addition,this thesis proposes a new EMG feature and designs a network feature extraction method for surface EMG signals to improve the gesture recognition rate in different scenarios.The main works of this thesis are as follows:(1)The public dataset Elonxi DB is proposed,and the sEMG signal acquisition procedure is described in detail,including the selection of gestures,the arrangement of electrode positions,and the collection rules for myoelectric signals.(2)The traditional methods of sEMG feature extraction are studied,including time domain feature(TD),frequency domain feature(FD),time-frequency domain feature(TFD)and parametric model method.Based on these algorithms,an improved feature is proposed: Dynamic Amplitude Difference(DAD).Experiments show that this feature can reflect the signal change process to a certain extent.Therefore,it can reduce the error and improve the recognition effect.(3)In order to demonstrate the gesture recognition performance of the traditional classification model in the time-segment signal,this thesis uses K-Nearest Neighbor algorithm(KNN),linear discriminant analysis(LDA)method and support vector machine(SVM)to identify the acquired sEMG features in intra-session and inter-session.On this basis,we explore the influence of the sEMG sampling frequency on gesture recognition.Experiments show that the sensitivity of EMG characteristics subject to sampling rate is different.Besides,the low sampling frequency(about 200Hz)can still get a good recognition effect in different scenarios,which provides practical significance for reducing the cost of hardware acquisition.(4)In this work,a network feature extraction method for surface EMG signals is proposed.In the case of electrode position shift and data difference between different people,the network feature(CNNFeat)is compared with 26 traditional EMG features(including DAD feature),and it is evaluated from three indicators: recognition rate,safety distance and repeatability.Besides,this method is also used to disclose the EMG dataset(NinaPro)to verify its robustness.Experiments show that the CNNFeat can improve the gesture recognition effect and has strong robustness.The network structure proposed in this paper can obtain good performance in myoelectric applications.At the end,the thesis summarizes the whole paper and proposes the plans and prospects for future work.