A Novel Framework for Cross-User Open-Set Myoelectric Pattern Recognition.

Abstract

This study is aimed to develop a robust myoelectric pattern recognition method for simultaneously alleviating cross-user variability and outlier motion interference. In the proposed method, a convolutional neural network (CNN)-based feature extractor is pre-trained using the data from a set of existing users. Next, a few labeled data of inlier motions recorded from a new user are utilized to implement model transfer and adaptation, while the prototype representation of each inlier motion is calibrated. In this process, a Euclidean metric-based prototypical loss is adopted to facilitate inter-class separability and intra-class compactness. Subsequently, any inlier/outlier motion is tested and identified based on a prototype matching procedure. The proposed method was evaluated on surface electromyogram signals recorded by an 8-channel armband from twenty subjects, including six inlier motions and ten outlier motions. When testing with each subject following a leave-one-out testing strategy (the remaining subjects were considered to form a set of existing users for pre-training a model), the proposed method achieved average accuracies of 82.37 ± 1.21% for the inlier motion recognition and 97.21 ± 2.65% for the outlier motion rejection, respectively, and it outperformed the existing methods with statistical significance (p < 0.05). The proposed method yielded excellent performance in cross-user open-set myoelectric pattern recognition with only a short and simple calibration routine. Our work offers a valuable solution for improving the robustness and usability of myoelectric gestural interfaces.