Trans-cVAE-GAN: Transformer-Based cVAE-GAN for High-Fidelity EEG Signal Generation.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Signal Processing

Background:

• Electroencephalography (EEG) signal generation is complex due to non-stationarity and spatiotemporal coupling.
• Existing generative models like VAEs and GANs struggle with EEG fidelity, spectral stability, and semantic control.

Purpose of the Study:

• To develop an advanced generative model for high-fidelity EEG synthesis.
• To enhance the controllability and stability of generated EEG signals for affective computing.

Main Methods:

• A Transformer-based conditional variational autoencoder-generative adversarial network (Trans-cVAE-GAN) was proposed.
• The model incorporates Transformer-driven temporal modeling, label-conditioned latent inference, and adversarial learning.
• A multi-dimensional structural loss was employed to preserve temporal, spectral, and statistical properties.

Main Results:

• The Trans-cVAE-GAN achieved high similarity to real EEG signals across multiple datasets (Pearson correlation ≈ 0.84).
• Generated signals exhibited low spectral divergence (KL divergence ≈ 0.39) and improved stability over baseline GANs.
• Augmenting emotion recognition tasks with generated EEG improved accuracy from 86.9% to 91.8%.

Conclusions:

• The proposed Trans-cVAE-GAN effectively generates high-fidelity, stable, and controllable EEG signals.
• This approach offers a scalable solution for affective computing and brain-computer interface applications.
• The model demonstrates significant improvements over conventional generative methods for EEG.