Epilepsy Detection And Emotion Recognition Based On EEG

The brain is the most important and magical organ in the human body,but the exploration of its mechanism is only the tip of the iceberg.EEG signals record the overall reflection of the electrophysiological activities of brain neurons on the scalp,and reflect the physiological,pathological and even specific intention information of the subjects.It is an effective means for humans to explore the brain.At present,there have been a large number of researches and applications based on EEG signals,such as epilepsy detection,emotion recognition,sleep stage and fatigue detection,among which epilepsy detection for neurological diseases and emotion recognition for intelligent human-computer interaction are hot research topics.From a technical point of view,both of these studies involve fast and accurate feature extraction and classification of EEG signals.This thesis aims to improve the accuracy of epilepsy detection and emotion recognition,and the following researches have been carried out:(1)Aiming at the problem of high dimensionality and redundant information in the features of multi-channel epilepsy EEG caused by the phenomenon of "channel crosstalk",an epilepsy detection algorithm based on local preserving projection is proposed.Firstly,the non-linear components in the EEG signal are analyzed through the high-order spectrogram,which shows that with the intensification of the electrophysiological activities in the brain,the EEG signal presents a stronger non-linearity.Secondly,inspired by the manifold learning method used in face recognition,the EEG feature nonlinear mapping method based on LPP is designed,which reduce the dimensionality of EEG feature.The algorithm is verified on CHB-MIT and Bonn epilepsy dataset respectively.The experimental results show that inserting the proposed method into the typical EEG signal processing flow can effectively improve the classification performance and reduce the feature redundancy.(2)Aiming at the problem that the accuracy of the existing methods is difficult to improve due to the ignoring of multi-domain information,an emotion recognition algorithm that combines space-frequency information and feature fusion network is proposed with the help of the powerful nonlinear fitting ability of deep learning.Firstly,construct a three-dimensional tensor representation of the EEG sequence signal.The 3D tensor interpolates differential entropy(DE)features from five rhythms over the topology of EEG channels,which contains frequency domain and spatial domain information.Secondly,a neural network structure for fusion of space-frequency information is designed,which includes a frequency band weight module and a spatial feature learning module.The former is used to emphasize important frequency bands while the latter extracts deep semantic features.The algorithm is verified on the SJTU Emotion EEG Dataset.The experimental results show that the proposed method achieves an accuracy of 90.4% on emotion recognition task,which is better than the advanced methods.(3)Aiming at the problem of low migration accuracy of the "cross-subject" emotion recognition model,"cross-subject" emotion recognition algorithm based on collaborative and adversarial strategy is studied.Firstly,the distribution of EEG feature from different subjects was analyzed through visualization methods,and a hypothesis was put forward based on the obtained results: the edge distribution of EEG feature from different subjects is different,but the conditional distribution is the same.Then,the collaborative and adversarial strategy is introduced.The adversarial strategy is implemented by the discriminator and gradient reversal layer set in the feature extraction output part to align the distribution of different subjects’ EEG features;the collaborative strategy is implemented by the discriminator set between the convolutional blocks which can guide the network to extract more category information from the target domain.This strategy can be embedded in arbitrary EEG feature extraction network.This algorithm is verified by adopting Leave-one-subject-out cross validation strategy.The experimental results show that the proposed method achieves 84.4% accuracy on the task of cross-subject emotion recognition.The thesis combines the theories of signal processing and machine learning to analyze EEG signals.The proposed algorithms for epilepsy detection and emotion recognition have achieved high accuracy.The development of accurate epilepsy detection and emotion recognition models is useful for countering neurological diseases and establishing humanized human-computer interaction is of great significance.