کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Cross-Subject Aligned Contrastive Learning for sEMG Gesture Recognition

نویسندگان :

Ali Akbari¹ Zahra Moradi Shahrbabak² Monire Ameri Haftador³ Mehran Jahed⁴

1- دانشگاه صنعتی شریف 2- دانشگاه تهران 3- دانشگاه صنعتی شریف 4- دانشگاه صنعتی شریف

کلمات کلیدی :

Surface electromyography (sEMG)،gesture recognition،supervised contrastive learning،cross-subject alignment،subject-invariant representations،representation learning،pattern recognition

چکیده :

In human-computer interaction and prosthetic control, surface electromyography (sEMG) is a potential sensing modality for gesture recognition. Nevertheless, it is still challenging to develop models that consistently generalize among users due to the significant inter-subject variability of sEMG data. In order to specifically target subject-invariant embeddings, this study suggests a two-stage representation learning method. In order to promote subject-invariant yet class-discriminative features, a cross-subject alignment (CSA) term is added to a supervised contrastive loss during training of a convolutional neural network (CNN) encoder. This term penalizes between-subject dispersion within each gesture class. A Random Forest (RF) classifier is then fed the resultant embeddings. Three protocols, strict Leave-One-Subject-Out (LOSO), 5-Fold Cross-Validation, and an Aggregated condition are used to test the method on our collected dataset of twenty participants executing four gestures. The baselines are an end-to-end CNN and two traditional feature-based models (SVM and RF). The suggested approach outperforms all baselines statistically substantially (p < 0.05, paired t-test) in the 5-Fold and Aggregated settings, achieving accuracies of 97.5% and 98.3%, respectively. Our system outperforms the feature-based RF and SVM baselines and quantitatively surpasses the end-to-end CNN in the difficult LOSO assessment, achieving a mean accuracy of 85% ± 13%. Explicitly aligning cross-subject embeddings results in more stable and robust sEMG gesture recognition, as demonstrated by the alignment term's 3% absolute improvement in LOSO accuracy and reduction in subject-to-subject variance when compared to an otherwise identical Supervised Contrastive learning model trained without CSA. These findings underline the practical significance of the proposed framework for building reliable, user-agnostic sEMG-based interfaces with improved cross-subject robustness.