Font Size: a A A

Complex Network And Deep Learning-Based Information Fusion Of EEG And Its Application

Posted on:2022-05-26Degree:DoctorType:Dissertation
Country:ChinaCandidate:Y X YangFull Text:PDF
GTID:1520307034962759Subject:Control Science and Engineering
Abstract/Summary:PDF Full Text Request
Studying the brain neural mechanisms and further realizing accurate assessment,of different physiological and psychophysiological states,are of great significance for human health management and the improvement of the quality of life.As one of the most reliable and effective clues to identify human brain states,EEG is widely used by researchers at home and abroad.However,due to the low signal-to-noise ratio of EEG and the complex signal-coupling among channels,it has brought great challenges to the accurate assessment of the physiological and psychological states.Since complex networks can fuse multivariate information,and reveal the complicated information interaction mechanism among different brain regions,in recent years,they have been gradually used in EEG-based brain neural mechanisms studies.At the same time,deep learning technology provides new technical means for fusing and analyzing multichannel EEG,but its application in EEG-based tasks involves a series of technical challenges,such as the difficulties of designing the model structure.To solve these challenges,this dissertation uses Recurrence Quantification Analysis(RQA)and complex network methods to optimize the inputs of DL models,uses the deep Q network(DQN)strategy to design and optimize the model structures,and introduces the broad learning system(BLS)to improve model training efficiency.This dissertation firstly proposes a multivariate motif entropy visibility graph network(MMEVGN)to study the brain neural mechanism.Through the mutual information of the motif sequences of the limited penetrable visibility graph,MMEVGN is constructed and then different network measures are extracted,to study the differences of network topological structure so as to further reveal the differences of brain neural mechanisms under different emotion states,and explore the key electrodes related to emotion process.Aiming at the neural mechanism of driving fatigue,a multivariate weighted ordinal pattern transition network(MWOPTN)is proposed.Through the Shannon entropy of ordinal pattern transition between pair-wise channels of EEG signals,the MWOPTN is constructed and its small-worldness measure is extracted.Then the difference of the small-worldness characteristic of the MWOPTN from alert to fatigue state are studied,and the key electrodes related to fatigue are revealed through nodal degree measure.Aiming at the design problem of CNN models,an input optimization method based on recurrence plot is firstly proposed to reduce the size of input.Fear,which is the same negative emotion as sadness,is included in the classification task to improve the difficulty of the classification task.A channel-frequency convolutional neural network(CFCNN)model is proposed to perform the three emotions classification task on the optimized input.Further,the multivariate weighted joint recurrence network(MWJRN)is used to fuse the spatial information of EEG signals,combined with the time-domain information extracted by the differential entropy,to achieve input optimization,which is used for the same emotion classification task.The results show that the performance of emotion classification is further improved.A DQN-based automatic ally-search scheme is proposed to realize the dynamic optimization design of CNN models.The prior knowledge from a manually-designed high-performance model is adopted to design and limit the search space,and DQN is used as the search strategy to automatically search the CNN model structure.The effectiveness of the searched CNN model is validated on sleep stage classification and driving fatigue prediction tasks,respectively.In order to further improve the scope of network search but guarantee search efficiency,a step-by-step optimization framework is proposed.The particle swarm optimization(PSO)algorithm is used to expand the hyper-parameter space of the convolutional layer of the baseline CNN structure,searched by DQN.The results show that the proposed framework can further improve the performances of the previous two tasks.Considering the difficulty of designing the structures of deep learning models,in EEG-based tasks,a broad learning system is introduced.Aiming at the problem of applying a broad learning system to EEG data,a complex network-based broad learning system CNBLS is proposed.At the same time,the effectiveness of the model is verified in the task of driving fatigue detection.The results show that the model can achieve better results than other advanced methods on the task while reducing training cost.
Keywords/Search Tags:Electroencephalograph(EEG), Complex network theory, Deep learning technology, Convolutional neural network(CNN), Deep Q network(DQN), Broad learning system(BLS)
PDF Full Text Request
Related items