Upper Limb Movement Intention Recognition Based On EEG And SEM

In recent years,the incidence of stroke has increased year by year,and the rehabilitation training of stroke patients after surgery has been widely concerned.At present,the rehabilitation training system is mostly mechanical,patients are prone to secondary injury of the affected limb in passive rehabilitation training.This method cannot achieve good rehabilitation training effect.In order to increase the initiative of patients in rehabilitation training,motor intention recognition based on sEMG and EEG signals has become the focus of research in the field of rehabilitation training.In this paper,upper limb action recognition was studied,which was based on feature fusion of EEG and sEMG.On the basis of preprocessing and feature extraction,three offline upper limb movement intention recognition models were constructed by EEG signal,fusion of EEG and sEMG signals.Then the online movement intention recognition was realized on the upper limb rehabilitation training platform.The main work was as follows:(1)For EEG signal high and low frequency noise,physiological artifacts,other interference and sEMG signal baseline drift,high and low frequency noise,power frequency interference,other interference,there were two combined filters designed to denoise EEG and sEMG signals.For the EEG signal,the wavelet transform method was used to remove the baseline drift,the 1-100 Hz Butterworth bandpass filter and the 50 Hz notch filter were used to denoise the EEG signal.Then the denoised EEG signal was referenced and removed artifacts by ICA.For the sEMG signal,the wavelet transform threshold method was used to remove high-frequency noise and signal burrs,the wavelet transform digital filtering method and notch filter were used to remove baseline drift and power frequency interference.According to the daily needs and the effect of movement rehabilitation training,six rehabilitation training movements were determined,including horizontal abduction of the left shoulder,horizontal adduction of the left shoulder,left elbow flexion,horizontal abduction of the right shoulder,horizontal adduction of the right shoulder,and right elbow flexion.The experimental paradigm of signals acquisition was designed,and the collected signals were preprocessed.(2)A method for compensating eigenvalues of sEMG signal in frequency domain based on fatigue estimation is proposed.Optimizing the traditional frequency domain features of sEMG signal to solve the problem of decreased recognition accuracy of classification models caused by muscle fatigue.Firstly,rectangular window was added to the traditional frequency domain features,and the optimal window parameters were found to improve the correlation between signal and time.In order to reduce the feature dimension,five statistical features are extracted from the windowed features,including mean,median,variance,root mean square,peak.According to the assumed value and correlation coefficient of the signal,the final mean-frequency domain feature was selected to replace the traditional frequency domain feature for sEMG signal analysis.The new mean-frequency domain feature was optimized according to the established fatigue rule table and compensation rule.Compared with the SVM,KNN,and BP models before feature optimization,the accuracy of upper limb movement intention recognition increased by 11.2%,4.2%,and 4.1%.(3)A classification model based on the decision-level fusion method of EEG and sEMG signals was constructed,in order to effectively fuse EEG and sEMG signals and further improve the accuracy of the model.SVM,KNN,and BP classification models were constructed based on EEG signal,and also were constructed based on sEMG signal.The accuracy of each action in each model was composed into a weight matrix.To build a decision-making voting mechanism,each weighted classification model is a decision-making member,and the final result is voted by 6decision-making members.This decision-level fusion method makesEEG signal and sEMG signal synergistic and complementary,and at the same time combines the advantages of the three classification algorithms to construct a new classification model that fusesEEG and sEMG signals.Finally,the accuracy of offline motion intention recognition reached 96.3%.Compared with the feature-level fusion model with the highest accuracy of EEG and sEMG signals,the accuracy rate has increased by 4.2%,and compared with the single-modal model with the highest accuracy rate,the accuracy rate has increased by 5.9%.(4)An online rehabilitation training system was built.The system includes hardware equipment such asEEG cap,EMG acquisition instrument,and upper limb rehabilitation training platform,which realizes online acquisition,processing,and classification of signals.Four models,with higher accuracy of single-mode and multi-modal signals in the offline upper limb movement intention recognition experiment,were written into the system.These four models were the BP model of EEG signal,the SVM model of sEMG signal,the feature-level fusion SVM model of EEG and sEMG signals,and the decision-level fusion model of EEG and sEMG signals.An online verification experiment was designed,and the average accuracy rates of online recognition of the four models were 56.6%,78.3%,80.8%,and 86.4%.The decision-level fusion model of EEG and sEMG signals achieved good results,and the model recognition time was only 0.1254 seconds.The classification results were converted into control commands to control the limb rehabilitation training platform.This method realizes the active control of the rehabilitation robot,improves the patient’s active participation and training effect,and prevents secondary damage to the patient.