Synthesis of sEMG Signals for Hand Gestures Using a 1DDCGAN.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Modern prosthetics rely on AI trained with surface electromyography (sEMG) signals.
• Current sEMG data acquisition causes user discomfort and fatigue.
• Limited sEMG data sharing hinders prosthetic research and reproducibility.

Purpose of the Study:

• To propose and validate a novel method for generating high-quality sEMG signals using a 1D Deep Convolutional GAN (1DDCGAN).
• To address challenges in prosthetic control data acquisition and research limitations.
• To enhance the utility of sEMG data for prosthetic development.

Main Methods:

• Utilized a 1DDCGAN architecture to synthesize sEMG signals for hand gestures.
• Incorporated dynamic time warping, Fast Fourier Transform, and wavelets as discriminator inputs.
• Evaluated synthesized sEMG quality using two datasets, feature extraction via windows/increments, Mantel test, and classifier two-sample test.

Main Results:

• The 1DDCGAN successfully generated high-quality sEMG signals that preserved inter-feature correlations.
• Synthesized signals closely resembled original sEMG data.
• Classification accuracy improved by an average of 1.21-5% with the generated data.

Conclusions:

• 1DDCGAN is a viable solution for generating synthetic sEMG data, overcoming limitations of real-world data acquisition.
• The proposed method enhances data augmentation, reduces training time, and improves classification accuracy for prosthetic control.
• This approach promotes research reproducibility and facilitates the development of more sophisticated AI-driven prosthetics.