NUST Institutions Library Catalogue

This research presents a two-stage approach to investigate the use of generative deep
learning models to synthesize realistic surface electromyography (sEMG) signals for
improved gesture recognition. In the first study, a Conditional Variational Autoencoder
(CVAE) was developed to generate synthetic gesture-conditioned EMG data. Experiments
conducted on a self-collected dataset and the publicly available Ninapro DB3 dataset
demonstrated that classifiers trained on hybrid datasets (real+synthetic) achieved higher
accuracies compared to real-only training with gains in some scenarios. The second study
introduced a novel CVAE-TCN architecture integrating temporal convolutional networks
to learn sequential dependencies and to enhance temporal realism of synthetic EMG
signals. Evaluation metrics including Pearson correlation, Dynamic Time Warping (DTW),
and Wasserstein distance confirmed improved signal fidelity and better class separation in
the latent space. Across both studies the generative models proved effective in addressing
data scarcity boosting classification performance and enhancing the robustness of sEMG
based gesture recognition systems.