Research On EEG Modeling Based On Covariance Feature And Its Application In Fatigue Driving Detection

With the vigorous development of domestic industrial modernization and the improvement of people’s living quality,the number of cars has been increasing year by year,but it is also the casualties and direct economic losses caused by traffic accidents.Fatigue driving has become one of the main causes of serious traffic accidents globally.The research on effective fatigue driving monitoring technology is of great significance to ensure road traffic safety.Fatigue driving refers to the driver in a monotonous driving environment,sustained high intensity and excessive physical and mental overdraw,resulting in physiological and psychological dysfunction,so as to objectively appear the decline of driving skills.During driving a car,the driver’s brain needs to coordinate and direct the body to conduct driving operations,and abundant time-frequency information in EEG can reflect the state of the brain under different physiological conditions.Therefore,this paper considers using EEG signal modeling to monitor the fatigue degree of drivers.Brain-machine Interface（Brain Computer Interface,BCI）in the related research results show that in SPD（Symmetric Positive Definite）space,Riemannian metric or other metric between the positive defifinite matrix is superior to the traditional Euclidean distance,and the second order statistics such as regularized covariance matrix can better capture the signal characteristics in manifold space,so as to achieve better classifification performance.However,the current researches focus on the spatial features of EEG covariance but ignore the temporal information,which leads to insufficient extraction of the effective information in the signal.This paper focuses on exploring the potential of temporal domain features in the covariance manifold space based on EEG signals,and further investigating the effective representation of EEG in the covariance manifold space after the spatial-temporal information.Main work and achievements of this paper include:（1）Proposed a spatiotemporal fusion model of EEG covariance based on S-devergence:The temporal-domain features of covariance were extracted and investigated by using sliding-window technique,which was called Temporal Relation（TR）domain features;In（Spatial-Relation,SR）domain,we first calculated the geometric centers of two classes respectively by iterative algorithm in the SPD space,and then took the SDivergence（Stein-Divergence）between a sample trialwise covariance matrix and two centers as features of the SR domain,then we fused features of two domains and then put forward a model for fatigue driving detection.（2）Put forward a Spatial-temporal Joint Optimization Network on Covariance Manifolds of Electroencephalography for Fatigue Detection: considering the learning ability of model that merely extract SR domain features of covariance by using S-Divergence is weak,in this paper,by introducing a SPDNet（Symmetric Positive Definite Network）to learn SR domain characteristics automatically,we put forward a Spatial-temporal Joint Optimization Network on Covariance Manifolds of Electroencephalography,which realize the more accurate detection for fatigue driving.In the process of network optimization,we used SPDNet and RNN to update the parameters of the model iteratively,and then learned the more separable Spatial-temporal distribution of the covariance matrix.（3）The data set based on this research is obtained by collecting the EEG signals generated by the stimulation of the driver receiving the system’s randomly turning over the car in the virtual reality environment.The experimental results show that the classification accuracy can reach 86.238% by using only the covariance TR domain feature,which strongly demonstrates the validity of the covariance temporal feature investigated in this study.The classification performance of spatio-temporal fusion features is better than that of any single domain features.The accuracy of S-divergence based covariance spatio-temporal fusion model can reach 89.28%,which is 3.042% higher than that of TR domain features alone.The classification accuracy of SPDNet can reach 91.024%,indicating that compared with S divergence,SPDNet can better learn the spatial domain representation of the covariance of EEG signals in waking and fatigue states,and has better complementarity in temporal domain,which is more conducive to improving the performance and robustness of the model.In summary,we explored the covariance temporal domain characteristics of EEG,and proved its effectiveness and good classification performance.And on this basis,by fusing the spatial-temporal information of EEG in the covariance manifolds,the spatial-temporal complementarity can be utilized to obtain more stable and effective characterization,by which we can realize the accurate detection of fatigue driving on the basis of EEG modeling.