AdaWGAN: Data Augmentation for Few-Shot HD-sEMG Gesture Recognition Using Single-Trial Data

Area of Science:

• Biomedical Engineering
• Human-Computer Interaction (HCI)
• Machine Learning

Background:

• High-quality labeled surface electromyography (sEMG) data is scarce, hindering human intention recognition and practical HCI applications.
• Existing sEMG augmentation methods inadequately exploit spatial information from high-density sEMG (HD-sEMG) recordings.

Purpose of the Study:

• To propose a novel dual-branch Adaptive Weight Wasserstein Generative Adversarial Network (AdaWGAN) for physiologically consistent HD-sEMG data augmentation.
• To improve the fidelity and category-alignment of generated HD-sEMG samples for enhanced gesture recognition.

Main Methods:

• AdaWGAN jointly learns spatiotemporal activation maps and frequency-band features from single-trial HD-sEMG data.
• An adaptive weighting mechanism balances adversarial and classification losses to improve semantic consistency.
• The model generates high-fidelity, category-aligned HD-sEMG samples for gesture recognition tasks.

Main Results:

• AdaWGAN synthesized samples demonstrated high fidelity, with Pearson correlation coefficients exceeding 0.95 for most gesture classes.
• Achieved superior gesture classification accuracies (88.2±6.3% and 94.3±4.4%) on benchmark HD-sEMG datasets, outperforming state-of-the-art methods.
• Attribution visualization confirmed that AdaWGAN captures physiologically meaningful muscle activation patterns.

Conclusions:

• AdaWGAN offers an effective solution for augmenting HD-sEMG data, addressing the data scarcity challenge in HCI.
• The developed model contributes to interpretable and physiologically plausible generative models for biomedical HCI applications.