Research On Instruction Gesture Set Determination For Myoelectric Control Applications

The user dependence problem of electromyography(EMG)pattern recognition has seriously hindered the implementation of high-accurate and robust myoelectric control systems.Unlike most current researches which try to improve the robustness of userindependent EMG pattern recognition from the perspective of classification system design,this paper provides a new idea to solve the user dependence problem,which is from the perspective of instruction gesture set selection.Specifically,a general instruction gesture set determination scheme based on t-distributed stochastic neighbor embedding(t-SNE)is proposed.The scheme can be used to select the gesture combinations with low user dependence and high separability from large-scale gesture sets,which provides a technical basis for the determination of instruction gesture set for myoelectric control human-computer interaction applications.The main work carried out in this dissertation is as follows:(1)Two large-scale high-density surface electromyography(HD-sEMG)gesture databases were created and selected for the study.First,database DB_U was designed,containing 17 target gestures and 11 subjects.128-channel array electrodes were used for data acquisition.Meanwhile,database DB_Hyser_PR,which is a publicly available HD-sEMG gesture database,was selected for the study.DB_Hyser_PR contains 34 target gestures and 20 subjects.256-channel array electrodes were used for data acquisition.Sample augmentation and preprocessing operations were performed on the sEMG gesture signals from both databases to generate gesture samples.(2)Aiming at searching for gesture sets with low user difference and high separability for EMG pattern recognition in user-independent mode,a gesture separability evaluation scheme based on t-SNE was proposed.T-SNE was firstly performed on high-dimensional sEMG gesture samples of multiple users and multiple gestures for dimensionality reduction.Overlap degree parameter was defined to evaluate the overlap degree of two-dimensional feature vectors in the space.The value of overlap degree parameter was the measure of the gesture set’s separability.The scheme was applied to two large-scale databases.It was verified that there is a great correlation between overlap degree parameter and the separability of gesture set.The results verify the effectiveness of the proposed gesture separability evaluation scheme.(3)The optimal/inferior gesture sets with different sizes were determined for the two databases based on the gesture separability evaluation scheme.Specifically,to determine the candidate gesture sets of a certain size,the overlap degree parameters of all gesture combinations of this size in the large-scale databases were calculated.The gesture sets with small overlap degree parameter were regarded as optimal gesture sets,and those with large overlap degree parameter were regarded as inferior ones.The"optimal" and "inferior" gesture sets mentioned in this paper are just judged from the technical point of view:The separability of optimal gesture set is high and that of inferior gesture set is low.Therefore,the optimal gesture set can achieve good performance in user-independent recognition while the inferior one cannot.In this study,the optimal/inferior gesture sets containing(2～8)gestures were determined for DB_U_UI,and the optimal/inferior gesture sets containing(2～4)gestures were determined for DB_Hyser_PR.(4)Taking each optimal/inferior gesture sets as target gesture set,EMG gesture recognition was carried out using support vector machine(SVM)and bidirectional long-short term memory(Bi-LSTM)as classifiers.The recognition experiments include high-density sEMG gesture recognition in user-independent mode,low-density sEMG gesture recognition in user-independent mode and high-density sEMG gesture recognition in electrode offset mode.In all recognition tasks,the recognition accuracies of optimal gesture sets are higher than those of the inferior gesture sets by[12.57%,36.92%].Therefore,the optimal gesture sets determined by this study have better performance and can be used as candidate instruction sets for robust myoelectric control systems.The results verify the effectiveness of the instruction gesture set determination scheme proposed in this study.The study innovatively provides a new perspective to solve the user dependence problem in EMG pattern recognition,and the results are of certain application value to promote the development of myoelectric control technology.