Study On Adaptive Algorithms For Myoelectric Control

Myoelectric pattern recognition(MPR)technology provides an efficient approach for amputees to control the dextrous prosthetic hand,which has been an active research topic for two decades.By applying sophisticated signal processing and machine learning algorithms,the recognition accuracy can be as high as >90% when classifying >10types of hand/wrist motion on the state-of-the-art(SOA)academic work.However,in contrast to such nearly perfect performance,few commercial prosthetic systems employ the MPR technology.The key reason for this phenomenon is that the experiment condition in laboratory is much different from the daily usage environment of prostheses.In practical usage,the electromyography(EMG)signals' characteristics will variate due to many factors such as shift of electrodes,muscle fatigue and sweat,thereby compromising the classification performance.Recently,the research community has shifted the focus to improve the adaptability and robustness of MPR.On account of the requirement for re-calibration of classifier in long-term/across-day usage of MPR and the effect of arm movement on MPR,the following research work is developed.The strong non-stationarity of EMG in across-day usage leads to the requirement for re-calibration of classifier in a new day.Nevertheless,we hypothesize there still exist similarities among the EMG of the same motion from different days.Therefore,the models(classifiers)trained on previous days may be reusable for the current day to some extent.We propose a domain adaptation(DA)framework,which reuses the previous models into the training(calibration)process of the current day,to reduce the re-calibration time.The model with domain adaptation is produced from the combination of previous models by certain weighted coefficients and the model trained on the new day.Two corresponding algorithms are implemented to deicide the weighted coefficients for two types of classifier,polynomial classifier(PC)and linear discriminant analysis(LDA),respectively.The experiment results show that the proposed algorithms are able to significantly improve the classification accuracy when the training data set of the new day is small.Furthermore,based on the assumption that the EMG signals of the same motion from different days contain certain invariant characteristics,a common model component analysis(CMCA)framework is proposed to fulfill the zero re-training in acrossday usage of MPR.CMCA is developed based on LDA.It aims to find a projection matrix which minimizes the dissimilarity among the several previous LDA models.The optimal projection matrix is used to extract the common model component.The experiment results demonstrate that CMCA can improve the robustness of MPR in across-day usage,thus increase the classification accuracy of previous model when it is used in a new day directly without re-training.In addition,a cascaded adaptation(CA)framework is proposed to achieve the fast calibration of MPR.The CA integrates the two algorithms,DA and self-enhancing(SE).In the across-day situation,the previous model is reused(DA)to reduce the retraining time of the new day;during the within-day usage,the testing samples and their recognized labels are utilized to automatically update the model parameters(SE)to follow the slow change of EMG signals,thereby avoiding the re-calibration of classifier.Both off-line and online experiments are conducted to comprehensively evaluate the proposed framework.The results demonstrate that the CA is able to decrease the training time of MPR.The system can be used for a day without re-training when the initial training time is not more than 1 minute,significantly reducing the training burden for prostheses users.We also investigate the issue about the robustness of MPR to arm movement.The experiment results show that the variation of arm position has significantly adverse effect on MPR.In addition to providing the classification accuracy,the effect of arm movement on the EMG feature distribution is quantified via three metrics,i.e.,repeatability index,mean semi-principal axis and mean centroid bias.Finally,the training strategy which is able to improve the robustness of MPR meanwhile requires as little training time as possible is proposed.In summary the aim of this dissertation is to improve the adaptability and robustness of MPR.These efforts have the potential for advancing the myoelectric control based on pattern recognition from laboratory to clinical application and helping to provide a ‘good commander' for the existing commercial dextrous prosthetic hand.