Robust Features And User Adaptation For Myoelectric Control

Research on myoelectric controlled interfaces has a history of more than 60 years.Compared to traditional amplitude-based methods,the pattern recognition-based(PRbased)scheme provides intuitive control of prostheses,and it is a promising way for the control of the prostheses with multiple degrees of freedom(DOFs).After decades of research,the state-of-the-art algorithm could achieve the accuracy above 90% when classifying 10 classes of motions in the laboratory.Contrary to its success in academia research,PR-based algorithms have not made any noticeable impact in multiple DOFs commercial prostheses,all of which still use the amplitude-based method.One of the key reasons for this gap is that experimental conditions in the laboratory are significantly different from conditions in daily use.In activities of daily living,the myoelectric signals of a given motion could be changed due to many factors from the human(user)and machine(physical system),which could greatly decrease classification performance of the PR-based scheme compared to that in the laboratory.In this dissertation,the influences of force variation,non-stationary of of myoelectric signals,and electrode shift on PR-based interface are all independently investigated.Force variation has demonstrated a substantial impact on classification performance.Based on the finding that myoelectric signals of multiple hand muscles scale linearly with force,the normalization process is used to decrease the differences among different contraction levels and two robust features are proposed.Experiment results show that when the classifier is trained on data of one force level,and tested on data of three force levels,i.e.20%,50%,80% maximum voluntary contraction(MVC),the performance of the two proposed features are both significantly better than that of the traditional time domain(TD)feature.This method makes the PR-based system robust against force variation without the need for using training data from multiple force levels.It has been shown in the literature that the classification performance of myoelectric signals degrades over time for the non-stationary of myoelectric signals.We collected myoelectric signals from 8 able-bodied subjects and two amputees over 11 consecutive days,and systematically investigated the changes of two types of classification errors(the training and testing data are from different days,and the separation is one day and three days,respectively).The experimental results show that errors both decrease exponentially and are gradually close to the within-day results(the training and testing data are from the same day).Feature space analysis shows that subjects produce more consistent motions over time.Results indicate that while using the control interface,the quality of myoelectric signals is improved by user adaptation(learning).When the classifier is trained after user adaptation,the classification performance degradation over time can be mitigated.Electrode shift is inevitable in dons and doffs.In this dissertation,gray-level cooccurrence matrix(GLCM)calculation is improved and combined with high density(HD)electrodes to exploit spatial distribution information as features.Experimental results show that when the electrodes are worn as a band around the circumference of the forearm,the proposed feature is invariant against the perpendicular shift where the shift distance is equal to the inter-electrode distance.The effect of the perpendicular shift is controlled within the value where the shift distance is half the inter-electrode distance.When the shift distance is equal to half the inter-electrode distance,classification accuracy is significantly increased compared to the three robust features,i.e.time domain autoregressive(TDAR),common spatial pattern(CSP),experimental variogramv(Variog).Combined with the small inter-electrode distance of HD electrodes,this provides an effective way to mitigate the effect of perpendicular shifts,which has shown to be the main source of performance degradation due to electrode shift.In summary,through proposing robust features and using user adaptation,this dissertation aims to improve classification performance of PR-based myoelectric control for daily living,facilitating the transfer of knowledge in academic research into commercial prostheses.