Research On GAN-based Data Augmentation For Brainprint Recognition

Due to the concealment,stress resistance,vividness,and revocability of EEG signals,they are considered a new type of biometric factor.Identity recognition based on EEG signals(brainprint recognition)has made progress.However,obtaining EEG data requires a lot of time and effort,which poses a great challenge for the practical application of brainprint recognition and is also difficult for users to accept.Therefore,it is very meaningful to perform data augmentation on EEG signals for brainprint recognition scenarios.Compared to traditional data augmentation methods,generative adversarial network(GAN)models have stronger generalization and learning generation capabilities.Therefore,the EEG data augmentation methods in this article are all based on GAN.This article expands the training dataset of the brainprint recognition model by performing GAN data augmentation on EEG topographic map data with time-frequency spatial features and GAN data augmentation on temporal EEG data.Design and implement an integrated verification system for brainprint recognition data augmentation,and propose a brainprint recognition algorithm that can span multiple paradigm datasets.The main research content of this article is as follows:1.In image based brainprint recognition scenarios,the GAN model in the field of image generation also faces the problem of unstable training process due to the difficulty in characterizing the spatial features of EEG signals in existing image forms.This article proposes a data augmentation method for EEG Spatial Features GAN(ESF-GAN)based on spatial features of EEG images.The time-frequency characteristics of EEG signals are mapped to the cerebral cortex space and processed into EEG topographic maps,which to some extent represent the brain region activities highly related to the paradigm tasks when generating EEG signals.ESF-GAN quickly locates the key areas of real images through attention mechanisms and feature enhancement modules,improves the quality of generated images,and accurately characterizes EEG spatial features.And use gradient penalty functions to stabilize the training process of the network model.Experiments were conducted on the BCI Competition IV 1 dataset,and the proposed method was demonstrated to have advantages over traditional methods through visual comparison of generated images for qualitative analysis and calculation of FID scores for quantitative analysis.The usability of the method was verified through downstream brainprint recognition task experiments.2.In the context of temporal feature based brainprint recognition,the temporal resolution of EEG is much higher than the spatial resolution,but the existing GAN models have limited ability to generate temporal EEG data.This article proposes an attention based time series GAN(ATGAN)data augmentation method based on attention mechanism.Based on the randomness and high temporal resolution of EEG data,high-dimensional temporal features of EEG are mapped using an LSTM network structure based codec,generator,and discriminator.Attention mechanism is used to fully learn temporal features of EEG data and the correlation between electrode channels to generate temporal EEG data for expanding the training dataset of recognition models.Experiments were conducted on the BCI Competition IV 2a dataset,and qualitative analysis using PCA and t-SNE visualization,as well as quantitative analysis of discriminant and predictive scores,demonstrated that the method has more advantages compared to traditional methods.The usability of the method was verified through downstream brainprint recognition task experiments.3.In order to verify the usability of the two data augmentation methods proposed in this article in the brainprint recognition system,an integrated verification system for brainprint recognition data augmentation was designed and implemented,integrating two data augmentation methods and corresponding brainprint recognition algorithms for different scenarios to choose from.In order to verify the availability of ESF-GAN based data augmentation methods in brainprint recognition,a brainprint recognition algorithm based on EEG topographic maps that can cross multiple normal form datasets was proposed.The system has verified the availability of ESF-GAN based data augmentation methods and ATGAN based data augmentation methods in different recognition scenarios.