Research On Surface EMG Signal Detection And Pattern Recognition For Bionic Manipulator

With the rapid development of man-machine interface technology,artificial intelligence and machine learning are widely used in various fields of civil life.In order to improve the quality of life of disabled limbs,the application and control of intelligent prosthesis has become the focus of current research in recent years.Using EMG signal to control prosthetic limb is a popular human-machine interface technology at present.EMG signal can best reflect the physiological changes of human action,which is of great significance in the field of prosthetic limb control.EMG control needs accurate signal acquisition,effective feature extraction,high recognition rate classification algorithm as the basis,and also needs to ensure the real-time of the system to reduce the control delay.At present,most of the research is based on EMG pattern recognition in the laboratory environment.There is no in-depth study on the real-time performance of the algorithm,and the acquisition signal is usually carried through commercial special equipment,which is not easy to carry.This paper will work from the following aspects:Firstly,aiming at the problems of high price and low accuracy of professional commercial acquisition equipment and most domestic EMG signal acquisition equipment,a portable and wearable EMG signal acquisition device is designed.ADS1299,which is specially used to collect bioelectric signals,is adopted to ensure the accuracy of EMG signal acquisition.At the same time,the traditional EMG acquisition equipment has fewer channels.This paper designs a signal acquisition system that supports up to 4 channels,and uploads data to the host computer through wireless transmission.Secondly,aiming at the problem of more interference noise in EMG signal,such as power frequency interference,baseline drift and so on,the filter design of 50 Hz notch filter,median filter and band-pass filter is adopted to filter it.Aiming at the problem of low accuracy of action classification with single feature,the feature selection of EMG signal is tested in detail in this paper.By comparing the results of feature combination in single time domain,frequency domain and each feature combination,the average absolute value,root mean square and waveform length are selected as the input of the classifier.Using BP neural network and support vector machine to classify the selected eigenvalue combination,the accuracy is 92.8% and 83.3% respectively.Finally,BP neural network is selected to classify the action of the three-time-domain feature combination extracted from the four-channel EMG signal,which improves the accuracy by 3.2% compared with the single feature classification.Finally,aiming at the lack of complete software training platform for acquisition equipment at present,this paper designs a hardware-based action recognition system to control the communication,sampling rate,acquisition mode and other related parameters of the equipment,which can accurately and effectively receive uploaded data and draw real-time waveform of data,and can carry out feature extraction,classification training,access features and training for data.Network operation integrates data activity segment detection,network training parameter display,serial communication and other functions to realize real-time control of the manipulator.