An improved Generative Adversarial Network for Denoising EEG signals of brain-computer interface systems

Removing artifact contamination from electrooculogram (EOG) and electromyography (EMG) signals is a key preprocessing step for detecting and diagnosing Electroencephalography (EEG) signals. Previous traditional approaches to denoising depend on inherent experience and prior knowledge, which is clearly detrimental to the development of real-time brain-computer interface (BCI) systems. Several methods have been proposed for end-to-end denoising using deep learning networks recently, but effectively improving the accuracy and robustness of denoising remains a great challenge. Here, we carefully design a novel Generative Adversarial Networks (GAN) neural network framework including a generator (G) and discriminator (D) where the G is used to remove artifacts in EEG and reproduce the original signal. Through the relevant tests on the EEG data set, the experimental results indicate that our algorithm far outperforms the three benchmark models in the case of the worst signal contamination, especially showing converges quickly and rarely appears mode collapses during the training process which may better promote the development of real-time EEG denoising in the BCI.