Method And Algorithm Implementation Of Double-brain EEG Data Analysis

The human brain is an extremely complex system.The two main research contents of brain science are the exploration of brain machanisms and calculation,while the traditional brain science mostly focuses on single brain,lacking of inter-brain interactions.Brain computer interface(BCI)is a research hotspot in the filed of brain science.Multi-mind hybrid BCI provides a new idea for designing and developing more realiable,practical and multi-degree-of-freedom BCI technology.EEG(Electrocephalogram)-based hyper-scanning is an emerging technology that can simultaneously record multi brains’ EEG signals with the advantagies of higher temporal resolution and lower cost.Therefore,this dissertation studies the data analysis method and algorithm impletations for cross dual brains using EEG-based hyperscanning.Specifically,this dissertation systematically studies the feature extraction and selection and classification of dual brain EEGs and focuses on the applications and extensions in dual social emotion computing and motor imagery resting state recognition.The main contributions and innovation works of this dissertation are summarized as follows:1.EEG feature extraction and its application in hyper-scanning:the EEG feature extraction methods are analyzed systematically including PSD(Power Spectrum Density),complexity,brain network and the extention processing in hyper-scanning.For inter-brain PSD calculation,we introduced the cross-PSD;for inter-brain complexity,we used the transfer entropy;brain connectivity can be used to calculate the relationship between channels.Based on the measures,we proposed the inter-brain connnetivity computing and brain network analysis.2.Feature fusion algorithm based on Canonical Correlation Analysis(CCA):the algorithm includes two strategies of combination and summation.Fusing different features can-avoid over-fitting and enhance feature representation.The improved frog Leaping algorithm based on brainstorm(B-SFLA)is proposed,which modifies the leaping rules,increasing the search space and pointing out the rapid convergence direction based on the ’host’ role in brainstorm.It is verified that the algorithm has better optimization performance,which could be applied to parameter optimization of SVM(Support Vector Machine)and LDA(Linear Discrimination Analysis)classifers.3.The PLV-based(Phase Locked Value)brain network classification method:this method is proposed to solve the EEG cross-datasets classification.First,to calculate the PLV on preprocessed EEG and then transfer into a binary netork with the setting threshold;Secondly,to extract the feature parameters of network,degree,clustering coefficient and global efficientcy;Finally,the method is veried across four EEG datasets including two BCI open datasets,one neuroeconomics and fatigue driving datasets.In further,this method is extended into inter-brain EEG data analysis.4.The cross brain coarse-grained emotion recognition computing method is proposed using dual EEG-hyper-scanning signals:traditional EEG-based emotion recognition is using single brain modal without considering the interactions among multiple brains.This is the first study of cross-brain emotion computing using simultaneously reorded EEG signals.Firstly,a cross-dual-brains coars-grained(negative and positive)classification method based on PSD and inter-brain PLV is proposed,including experimental design,intra-and inter-brain PLV feature extraction across different frequency bands and comparisions on different classifiers.The other method based on the inter-brain phase-lag index(PLI)model for fine-grained emotion recognition(four categories,’anger’,’disgusting’,’neutral’ and ’happy’)is proposed which covers experimental design,Event Related Potential(ERP),intra-PLI and interPLI connectivity calculation,hPLI-Conv deep convolutional network(inter-,intra-PLIConv)construction and classification performance analysis.The results indicate that the ensemble tree with fusion PSD and inter-brain PLV made the best performance in the coarse-grained emotion recognition computing;the four emotional ERPs are significantly different and the response order ERP has no significant difference;the inter-PLI values are lower than intra-PLI and inter-PLI-Conv shows the best classification performance,which is up to 83.3%.5.The cross-brain caculation methods for motor imagery(MI)idle state recognition using dual EEG signals are proposed.The brain computer interface includes two categories:asynchronous and synchronous.The asynchronous brain computer interface can realize free control.The difficult problem is to recognize the task and idle states.This is the first study of dual-brain calculation methods in motor imagery idle state detection.The method includes dual-brain MI experimental design,common spatial model(CSP)and inter-brain network feature extraction based on minimum spanning tree(MST),a deep convolutional neural network(CNN)model combined inter-PLI and joint-brain PLI.In each analysis method,we designed several fusion two brains features strategies.In the first method,we fusion two brains to extract CSP features by averging the values of the two EEGs and combine the two subjects’ EEG electrodes respectively.The results show that the combined features with inter-CSP and inter-brain networks has the best classification performance;In the inter-PLI CNN recognition method,we constructed a CNN network and compared the classification performanece among intra-brain,inter-brain and joint-brain.Results show that the interPLI CNN model has the best classification performance,and the alpha band is significantly higher than other bands.The joint-brain model outperforms the signal brain.