Pattern Recognition For Active Motion Of Ankle Joint Based On Dual Modal Feature Fusion

Stroke can cause the patient's abnormal walking gait.Independent walking is an important part to improve the quality of life and rehabilitation training.Being lack of the effective communication and control means to rehabilitation equipment,the traditional approaches still do not availably improve the patient's motor function,and it brings heavy burden to patients,families and society.Therefore,exploiting an intelligent system,which is suitable for the identification and control of human lower limb movement,has become a hot spot in the field of rehabilitation engineering.At present,the input signal in the most motion control system of the human lower limbs is the single mode EEG signal or the surface EMG signal of the corresponding muscles.Usually the single modal identification system may be more restricted in application.Since the multi-mode signal fusion has many kinds of input forms and can provide more extensive feature information for action pattern recognition,it is more effective to the identification and control of complex movements.At the same time,the multimodal physiological signals are complementary.Therefore,the information fusion and comprehensive utilization can eliminate the blindness of the brain computer interface(BCI)and improve the universality of the system.Based on the feature fusion of EEG and EMG signals,the paper explored the classification of four kinds of active action for human ankle joint by the methods of pattern recognition.The main work includes the following aspects:(1)The experimental paradigm of EEG and EMG fusion system was designed by E-prime software.Based on experimental paradigm,EEG and EMG signals were synchronously measured under the different action mode of the ankle joint by EEG and EMG equipments.(2)The original EEG signal and EMG signal were pretreated to obtain pure signals.For the EEG signal,applying Scan4.5 analysis software to remove the baseline drift and ocular artifacts,using AR model method to remove the spontaneous EEG signal,utilizing coherent averaging method to obtain P300 evoked potential signal.For the EMG signal,the adaptive canceller was designed to inhibit the 50 Hz power frequency interference,the wavelet de-noising method was used to eliminate the low frequency drift,high frequency interference and motion artifacts,and obtain the pure EMG signal.(3)Feature extraction of EEG and EMG signals.For EEG signal,the methods to the wavelet transform and the time domain energy entropy were designed to extract the EEG features with the best difference.For EMG signal,the methods to time domain and frequency domain analysis were used to extract the corresponding features of EMG signal.The wavelet packet transform was utilized to extract the feature vectors of wavelet packet coefficient energy and variance with good singularity.(4)Feature classification for dual-modal fusion system.Using the support vector machine classifier based on radial basis function kernel,the ankle motion modes were classified separately by P300 evoked potential features and surface EMG features.Based on the combination of the EEG wavelet transform approximation coefficient with the time domain energy entropy,the average classification accuracy of classifier is 88.2%.Based on the EMG feature vectors,energy and variance of the wavelet packet coefficients,the classification accuracy is 92.8%.Based on EEG and EMG fusion system,the four motion modes of the ankle joint were classified by Bayesian classifier.And the average classification accuracy reaches to 93.1%.The simulation results show the effectiveness of the fusion system in the motion pattern recognition of the ankle joint.The methods and results of this research can be applied to the fields of sports rehabilitation engineering,neuroscience,artificial intelligence,bionic robot and so on,and have the dual value of science and application.