Research On EEG Analysis Based On Multi-view Learning

Electroencephalograph(EEG)contains a large amount of physiological and pathological information,which helps humans understand the functional state of the subject’s brain,and further learn various cognitive and physiological processes.In recent years,many researchers have studied the automatic analysis and recognition of EEGs,aiming to use EEGs in task scenarios such as epilepsy detection and emotion recognition.However,EEG data are complex and high-dimensional,while most of existing works extract signal features from one dimension,which cannot fully utilize the information in various dimensions.This paper proposes an approach based on multiview learning to extract and fuse EEG representations from different views,which will help model focus on more important information to improve the accuracy of classification.The main work and contributions of the paper are as follows:First,to address the problem that existing automatic analysis models fail to fully utilize the multi-dimensional information of EEGs,this paper introduces multi-view learning technology to mine the features of EEG signals in both temporal and frequency dimensions.Two representations are constructed from EEG signals: one-dimensional time-series view and two-dimensional time-frequency view.We apply them as the input of multi-view encoders based on the Transformer architecture,in order to capture more important temporal and frequency components from different views with selfattentive mechanism.The experimental results show that the multi-view representation can improve the accuracy of the classification model compared to the single-view feature representation.Second,to address the problem of difficulty in fusion and information redundancy in feature representations between different views,this paper designs a multi-stage information fusion framework and builds an automatic analysis model for short-time EEG signals.A dual encoder structure is designed based on stacked attention mechanism,where the feature representation of time-frequency view is used to guide the learning of the time-series view’s representation for the initial information interaction.Besides,a gated attention unit is constructed for information integration between different views to make full use of the complementary information in the multi-view representation of EEG signals.Experimental results show that the EEG analysis model based on fusion of temporal-spectral feature achieves improvement in all the metrics on the epilepsy detection task.Thirdly,to address the problem that the lack of global temporal information leads to the degradation of accuracy in long EEG signal classification tasks,this paper adds a temporal aggregation unit based on dilated convolution,which make full use of context information in order to capture long-term temporal dependencies.Meanwihle,two spatial attention mechanisms are introduced inside the stacked dual-encoder to enhance the ability of learning spatial relationships in multi-channel EEG signals.The experimental results show that the EEG analysis model combining temporal-spectralspatial multi-domain features shows good accuracy and stability on the emotion recognition task.