HD-sEMG Feature Decomposition via Muscle Synergy and Dissimilarity Metric Learning for Robustness Against Unknown Gestures

Most traditional myoelectric pattern recognition (MPR) systems are limited to recognizing a fixed set of gesture classes and are prone to performance degradation when exposed to unknown gestures. This study proposes a robust MPR framework that simultaneously enhances intra-class compactness and improves open-set rejection performance. Time-domain features of high-density surface electromyography (HD-sEMG) signals are first decomposed into pattern-specific and pattern-variant components, preserving essential muscle activations and reducing intra class variability. A unified model is then constructed by integrating dissimilarity metric learning with classification, enabling simultaneous estimation of an anomaly score and class label for the input gesture. For each known gesture, a pattern-specific decision boundary is defined based on the maximum anomaly score. This allows accurate classification of known gestures and effective rejection of unknown ones. The proposed method is evaluated on a self-collected dataset containing 17 gestures and a public benchmark dataset containing 65 gestures. In intra-session experiments on both datasets, it achieves over 99% classification accuracy for known gestures and more than 98% rejection accuracy for unknown gestures. Under challenging inter-session conditions, it still maintains over 77% open-set recognition accuracy, substantially outperforming existing open-set MPR methods. These results demonstrate the effectiveness of combining muscle synergy decomposition with dissimilarity metric learning to improve the robustness of myoelectric interfaces.