Passive brain-computer interface (BCI) based on electroencephalographic (EEG) signals show promise for monitoring human vigilance in critical settings that continuously strain human attention. However, the lack of extensive public datasets limits progress in artificial intelligence research. To overcome this problem, we implemented generative adversarial networks (GANs) to augment existing EEG datasets. We focused on two datasets: a publicly available resting-state dataset and a custom one simulating industrial activities. After extracting time-variant alpha indices using continuous wavelet transform, we compared the generated data with real data using the L2 distance metric and autocorrelation function. We additionally assessed two forecasting models trained on original and augmented datasets, comparing their predictions. The integration of synthetic data led to an improvement in signal behavior prediction, as evidenced by a 32% reduction in mean absolute error for the resting-state dataset. Furthermore, a metric inspired by the Frechét Inception Distance was computed using the forecasting model to discern the distributions of embeddings for the real and generated data, with results showing a strong resemblance between the two. This study challenges the limitation of acquiring domain-specific EEG data and aims to develop a robust signal generation framework.