EEG signal enhancement based on 2K-cGAN

Brain-computer interface technology provides a new connectivity pathway between the human brain and external computers, which can recognize desired commands by decoding task-related neural responses from electroencephalographic (EEG) signals. EEG is a unique information factor that correlates with the subject’s genetic information, and thus is considered to be robust against forgery, demonstrating great potential in identification, and EEG-based biometric identification is being progressively applied. However, while EEG exhibits significant pattern variability among subjects bringing advantages for authentication, it also adds challenges to its data acquisition, and how to acquire training data in a targeted manner is the current challenge. In this paper, we propose a generative adversarial network (2K-cGAN) EEG augmentation method based on 2K standard classifiers, which utilizes a model to generate data in the form of a tensor of EEG signals, and incorporates category labels into the network structure to regulate the training of the generator and the discriminator, while the discriminator is considered as a 2K dimensional standard classifier to categorize the output data, so as to enable the generator to target the generation of data. The proposed network model is trained and generated on two tensor representations of the BCI Competition-IV 2b dataset, and the accuracy of augmentation by different multiples is compared using the LetNet classical classification network, and the experimental results show that the data generated using this paper’s method can maintain feature consistency with the original data as well as capture the unique features of different classes.