Driving Fatigue Detection Method Based On EEG Core Brain Network

In recent years,with the increasing number of accidents caused by driver fatigue,the issue of driver fatigue detection has become an important topic of safety research,and electroencephalography(EEG)has been widely used in the study of driver fatigue assessment with the advantages of high temporal resolution and non-invasive signal acquisition.The common practice of EEG-based fatigue assessment is to directly extract and analyze the time-frequency features in the signal for fatigue classification and prediction.But this method can only analyze individual EEG channels independently,which is difficult to reflect the functional connections between different regions of the brain and portray the overall functional state of multiple regions of the brain.Thus,it is necessary to introduce a brain function network to analyze the brain as a whole.In fact,too many channels increase the probability of introducing noise and increase the computational burden,so it is necessary to extract the core network for subsequent analysis and classification of brain functional networks.For fatigue detection based on EEG brain functional networks,this thesis proposes three methods:(1)the core network extraction method based on inter-state differences,(2)the method of constructing discriminative features from the core network using tensor decomposition and feature selection,(3)the method for integrating core network and time-frequency features of corresponding channels for driver fatigue detection using multilayer convolutional neural networks.The specific research work is as follows:(1)Based on the principle of graph centrality and correlation analysis,the two state networks are considered at the same time,and the core network is identified based on the difference between states,so that the identified core network has more identification ability.Then six graph metrics recommended by references are extracted from the core network as classification features.The average of the highest accuracy on the core network is 12.08% higher than the result of the full node network(no core network extracted).Compared with the core network extraction methods in the current state-ofthe-art literature,the core networks identified in this thesis not only achieve the optimal classification results with a smaller number of key nodes,but also the distribution curve of the accuracy is significantly better than that of the core network extraction methods in the literature(4.03% improvement in the average of the highest accuracy).(2)Researchers usually extract the brain functional network graph indicators in different frequency bands and linearly splice them into feature vectors for subsequent classification,and this method is difficult to extract the correlation between the functional networks in different frequency bands.To address this problem,this thesis uses tensor decomposition and feature selection to extract effective classification features from multilayer functional networks for driver fatigue detection.The constructed discriminant features not only contain structural information about the network of individual frequency bands,but also additional relational information between different frequency bands.Under the fair premise that the feature dimensions are consistent,the highest classification accuracy of the tensor decomposition extracted features increases more than 5% on average over that of the graph-theoretic features on the same core network.The feature extraction method based on core network and tensor decomposition has strong information extraction and anti-interference ability,and can be directly applied to the initial network without binarization,and the classification accuracy can reach up to 91.14%.(3)In this thesis,the core brain network is combined with the time-frequency characteristics of the corresponding channel to improve the recognition rate of driving fatigue detection.We investigate the changes of brain network and channel timefrequency features when driver fatigue level changes and find the potential complementarity of the two features and the greater differentiation of the brain network features.Therefore,based on the core brain network and the time-frequency features of the corresponding channels,this thesis uses multilayer convolutional neural networks to integrate two types of features for driving fatigue detection.The final experiment shows that the multilayer convolutional neural network model integrating the core brain network and the corresponding channel time-frequency features has the best classification results,and the highest accuracy can reach 93.03%.