Research On Myoelectric Pattern Recognition Based On Muscle Synergy Analysis And Exploration Of Its Applications

Electromyographic(EMG)signal generated by muscular activities is regarded as an important kind of bioelectric signals,which can reflect the movement intentions and states of body movements in real time.In such physiological context,myoelectric control using the surface EMG(sEMG)recorded from active muscles via sensors has recently become one of the hot research topics for human-computer interaction.In terms of control strategy,myoelectric control can be divided into on-off control,proportional control and pattern recognition based control.The first two provided too limited function to meet the needs of practical applications.Recently,due to the advantage in control of multiple degree of freedoms(DoFs)toward multifunction application,the control strategy based on pattern recognition has become the research focus.However,there are many considerable challenges in practical scene that severely restrict the development of the control strategy based on pattern recognition,which fails to translate into clinical and industrial application.Specially,the conventional myoelectric pattern recognition method only presents the motion pattern from data,and lacks physiologically relevant information about the underlying motor control processes generating the observable muscle activation patterns,resulting in the myoelectric control system without robustness and generalization ability.By contrast,muscle synergies are hypothesized to be basic building blocks employed by the central nervous system(CNS)to create complex movements,which can explains the control mechanism of CNS on muscles and provides a new approach to accuratly decode neural control information from sEMG generated by body movements.Given the important laws of motor regulation revealed in neuroscience research,this thesis incorporates muscle synergistic concept into the conventional myoelectric pattern recognition framework and proposes a series of myoelectric pattern recognition methods based on muscle synergies.Specially,we focused on exploring the robust myoelectric pattern recognition under varying muscle contraction level,and a universal myoelectric pattern recognition toward multiple users that can be quickly customized and calibrated for new user.Furthermore,an application of myoelectric pattern recognition in rehabilitation training towards individuals with neuromuscular injuries and a series of engineering problems for practical application are both explored.The primary work and achievements of this thesis can be summarized as follows:(1)Myoelectric pattern recognition robust for varying muscle contraction levels.A novel conceptual framework incorporated muscle synergies into conventional myoelectric pattern recognition was proposed to alleviating the effect of muscle contraction strength on classification performance.This is inspired by the assumption that the consistency of muscle synergy can be maintained at various muscle contraction levels.Specially,this study presented two myoelectric classification methods that incorporate muscle synergistic information:shared and task-specific muscle synergy-based classification(SMSC and TMSC respectively).Given the synergies specific to each task extracted by non-negative matric factorization(NMF),the TMSC tried to achieve the decision by computing similarity between the original data sample and its estimates rescontructed with different muscle synergies specific to individual taks.By contrast,the SMSC method utilized shared muscle synergies with NMF across all examined tasks as a set of global bases.In this regard,the muscle synergy analysis in the SMSC method worked as a feature transformation algorithm.Finally,a decision for the task discrimination was obtained by feeding the derived transformed feature into the conventional myoelectric pattern recognition classifier.To validate the effectiveness of our proposed classification methods,their performance in discriminating functional forearm tasks against varying muscle strengths was examined with data from a group of intact-limbed healthy subjects and compared with routine myoelectric pattern recognition classifier.To give a comprehensive performance evaluation,three classifiers were combined with multiple feature sets including the conventional time-domain(TD)feature set and other advanced feature sets recently developed to be robust to force variation.Different training or testing schemes were designed for conducting myoelectric pattern-recognition analyses using data from different combinations of force levels.Across all examined subjects,our results revealed that the average classification accuracies were substantially reduced when training with one force level but testing with others.This compromised performance was found to be remarkably compensated(7.53%-14.97%)by using any synergy-based methods.Furthermore,the TMSC method achieved comparable or even better performance than the other two classification methods and exhibited its usability as a robust classifier to varying muscle strength levels.(2)Multiuser myoelectric pattern recognition based on adaptive nonnegative matrix factorization.To alleviate the effect of repeated training burden on new user when using myoelectric control system,the purpose of this study is to explore a universal and general myoelectric pattern recognition method toward multiple users.Inpsired by the universal adavantage of muscle synergy model over EMG data represented by model capability,we extended the TMSC framework and proposed an adaptive learning based on non-negative matrix factorization algorithm.This algorithm reused the pretrained multiuser model parameters as the initial value of the new user model,and quickly adapted model parameters with calibration samples for new user.Two adaptive methods were presented in this study.In one method,the subjects-shared task-specific muscle synergies was pretrained,which was also regarded as the shared muscle synergies for new user.The new user’s personalized task-specific muscle synergies can be obtained through adapative learning and then are added to the subjects-shared task-specific muscle synergies obtaining the final task-specific muscle synergies for classification.In the other method,task-specific muscle synergies are independently pretrained for multiple users.Then for the new user,the model parameters are a linear combination of pretrianed models in the memory,while the weighted coefficients of these pretrianed models are learned through adaptive learning.To verify the effectiveness of our proposed methods,the high-density EMG signals of 21 finger gestures from 12 healthy subjects were analyzed and the leave-one-out cross-validation was employed to evaluate the classification performance.Experimental results showed that satisfactory results can be obtained with a few calibration samples.(3)Exploration of myoelectric pattern recognition toward motor rehabilitation application.With stroke survivors involved in this study,we explored the application of myoelectric pattern recognition control strategy to the issues and solutions of rehabilitation training toward neurological injury patients.Firstly,due to the inhibition of neuro-muscular transmission and limitation of effective neural control information extraction following stroke,we applied high-density electrode arrays to capture the most abundant muscle activity information and presented time-frequency representations of the wavelet packet decomposition technique for effective neural control information extraction.Secondly,due to the infeasibility of conventional EMG electrodes placement based on muscle anatomy analysis caused by various limb dysfunction,the wavelet packet feature selection approach was extended to channel selection.Such procedure can not only select effective myoelectric control sources from high-density electrodes,but also ensure the clinical utility of the myoelectric control system.Furthermore,we presented a pattern-recognition-based myoelectric control systems supporting multi-DOF and proportional control for hemiplegic upper limb rehabilitation training based on the aforementioned myoelectric pattern recognition incorporated with muscle synergy.The effectiveness of the proposed methods in hemiplegic movements recognition was validated on the sEMG dataset from stroke patients.Specially,the pattern-recognition-based myoelectric control systems supporting multi-DOF and proportional control can provide more nature motor feedback information of rehabilitation training for patients,which will help improve central nerve plasticity and rehabilitation treatment.