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

Despite recent advances in human-computer interaction (HCI) based on physiological signals like surface electromyography (sEMG), the scarcity of high-quality labeled data remains a major barrier for human intention recognition and practical HCI deployment. Existing sEMG augmentation methods often use temporal or frequency domain representations from sparse electrode recordings, while the spatial muscle activation patterns in high-density sEMG (HD-sEMG) remain insufficiently exploited. To address this challenge, we propose a dual-branch Adaptive Weight Wasserstein Generative Adversarial Network (AdaWGAN) for physiologically consistent HD-sEMG augmentation. AdaWGAN jointly learns RMS-based spatiotemporal activation maps and frequency-band features, enabling the generation of high-fidelity, category-aligned HD-sEMG samples from single-trial training data for gesture recognition. Furthermore, an adaptive weighting mechanism is incorporated into the generator objective to dynamically balance adversarial and classification losses, thereby enhancing the semantic consistency between real and generated samples. Compared with state-of-the-art generative methods, the synthesized samples of AdaWGAN exhibited high fidelity to real data, achieving Pearson correlation coefficients exceeding 0.95 for most gesture classes. Evaluations on two benchmark HD-sEMG datasets demonstrated that AdaWGAN achieved the best gesture classification accuracies of 88.2±6.3% and 94.3±4.4%, representing an improvement of approximately 5% over the baseline and outperforming other augmentation strategies. Furthermore, attribution visualization analysis indicated that AdaWGAN captures physiologically meaningful features consistent with genuine muscle activation patterns. This work provides an effective solution for sEMG data augmentation and contributes to the development of interpretable and physiologically plausible generative models for biomedical HCI.