EEG Generation of Virtual Channels Using an Improved Wasserstein Generative Adversarial Networks

Aiming at enhancing classification performance and improving user experience of a brain-computer interface (BCI) system, this paper proposes an improved Wasserstein generative adversarial networks (WGAN) method to generate EEG samples in virtual channels. The feature extractor and the proposed WGAN model with a novel designed feature loss are trained. Then artificial EEG of virtual channels are generated by using the improved WGAN with EEG of multiple physical channels as the input. Motor imagery (MI) classification utilizing a CNN-based classifier is performed based on two EEG datasets. The experimental results show that the generated EEG of virtual channels are valid, which are similar to the ground truth as well as have learned important EEG features of other channels. The classification performance of the classifier with low-channel EEG has been significantly improved with the help with the generated EEG of virtual channels. Meanwhile, user experience on BCI application is also improved by low-channel EEG replacing multi-channel EEG. The feasibility and effectiveness of the proposed method are verified.