Research On Myoelectric Pattern Recognition Based On Adaptive Learning And Its Human-machine Interactive Control For Prosthetic Hand

The bidirectional human-machine interface(bHMI)is the bottleneck of the state-of-the-art prosthetic hand,which directly determines the user experience.The bHMI of the prosthetic hand includes two parts,namely,the forward human motion intention decoding approach and the backward sensory feedback approach.How to achieve the reliable decoding of users' motion intentions and how to realize the user's intuitive perception of the working information of the prosthetic hand are the two major problems in the current bHMI research.The paper thoroughly reviewed the state-of-the-art bHMI of the dexterous prosthetic hand at first,then investigated the bHMI based on the Electromyography(EMG)pattern recognition and electrotactile sensory(ES)feedback.The objectives of this paper included improving the robustness of multiple motion pattern recognition based on EMG signals,and achieving the bidirectional interface compatibility,the reliability and the intuitive perception of the electrotactile sensory feedback.In this way,a human-machine interactive system with reliability and intuitive perception was established to improve the control and the perception of the prosthetic hand.Because of the transient or continuous changes such as electrode shifts after donning and doffing,sudden jitter of unrelated limbs,muscular fatigue,electrode displacement due to operating load,skin-electrode impedance variation due to environment temperature and humidity change,the multiple motion pattern recognition based on EMG signals suffers the performance degradation along time.To address the problem,most studies deployed adaptive learning algorithms to track the change of EMG features.However,current studies are lack of theoretical foundations,and thus the designing and the evaluation of the adaptive learning algorithms are based on experience or phenomenological analysis.This paper proposed the theoretical analysis method for the EMG pattern recognition based on the statistical learning theory.A model quantitatively describing the relationship between the recognition misclassification risk variation and the factors including the updating times,updating frequency,the generalization ability of the predictive model,and the ratio of samples with supervised labels was established.With the model,a class of standard adaptive learning algorithms were built up as a scale to evaluate different adaptive learning algorithms in general.Based on the proposed EMG pattern misclassification risk variaition analysis,the paper proposed the representative particle adaptive learning strategy(RPALS),which took advantages of the inherent characteristics of EMG signals to increase the equivalent ratio of samples with supervised labels.In order to reduce the computational cost during updating the predictive model,the paper proposed the universal incremental least square support vector machine(LS-SVM)and the universal incremental linear discriminant analysis(LDA)algorithms.By combining the RPALS and the universal incremental algorithms,the representative particle adaptive classifiers(RPACs)were established.Based on the simulated EMG data sequences,the robustness of the adaptive classifiers were proved.Meanwhile,the recognition results of myoelectric code control,the Electroencephal-ograph supervised myoelectric control,and the methods combining them with the proposed RPAC were evaluated to validate the misclassification risk variation theory.When building up the bHMI with ES feedback,it is essential to address the safety and the compatibility problem.The paper proposed the optimization method for ES electrode by modeling the electrical field of the stimulation to address the compatibility problem,and validated the model with nonwoven electrodes and flexible printed circuit electrodes respectively.The paper designed an array electrical stimulator with real-time electrode impedance measuring module and proposed the adaptive stimulating strategy to ensure the safety and spatial sensation unity during the sensory feedback.At last,the paper proposed a novel ES feedback coding strategy based on dynamic spatial modulation,which was proved to be more reliable and intuitive than conventional gradational frequency modulation and static spatial modulation.In order to validate the reliability of the bidirectional interface,the experimental platform based on HIT-V Hand was established.Noise cancellation method based on the least mean square(LMS)adaptive filter was deployed to eliminate the ES noise.With whole-day EMG signal acquisition,the standard adaptive learning algorithms,the RPACs,the code control,the EEG supervised EMG recognition method were validated.The experimental results showed the reliability of the adaptive learning methods,with which the misclassification risk variation theory was also validated.According to the interactive experiments for multiple motion recognition based on adaptive learning,the recognition result feedback was proved to be able to improve the online recognition results of adaptive learning.Furthermore,compared with single visual feedback,the sensory feedback combining ES feedback and visual feedback improved the online recognition performance,which showed the effectiveness of the ES feedback.