Studies On Spatio-temporal Analysis Methods For EEG Signal Based On Regularization

Electroencephalography(EEG)is a commonly used tool for recording functional brain activities.With the relatively high time resolution,EEG can effectively track the dynamic changes of brain activities.The oscillatory signals and event-related potentials(ERP)are two types of EEG signals that used in brain-computer interfaces(BCI),and the corresponding temporal regularities and spatial regularities of them are both very rich.Within the framework of regularization,this thesis systematically investigated the spatio-temporal analysis methods to solve the existing problems of analyzing these two different types of EEG signals.For the oscillatory signals,in order to extract the spatio-temporal features more effectively,and implement high-performance classification for the typically high-dimensional features,this thesis proposed an algorithm termed regularized spatio-temporal filtering and classification(RSTFC).Specifically,in the feature extraction module,the spatial filters and channel-specific high-order temporal filters were optimized simultaneously.As the number of unknown parameters would increase substantially with the embedding of temporal filters,a 7)2-norm regularization was enforced in the objective function to address the over-fitting issue.In the feature classification module,a convex optimization algorithm named sparse Fisher linear discriminant analysis(SFLDA)was proposed for simultaneous feature selection and classification of the typically high-dimensional spatiotemporally filtered signals.Experimental results showed that,compared with the stateof-arts spatio-temporal filtering algorithms,RSTFC was advantageous in terms of classification performance and computational complexity.As EEG feature extraction algorithms with high performance often require placing a large number of channels,which greatly impedes the daily convenience of BCI systems,therefore in this thesis,an algorithm termed combined channel selection and spatiotemporal filtering(CCSSTF)was proposed for resolving the channel selection issue.In order to select the most discriminative channels without sacrificing the performance,CCSSTF simultaneously selects the target channels and optimizes channel-specific temporal filters,Then spatio-temporal filtering method was applied on the optimal channel subset to extract features for classification.Experimental results showed that,the number of channels was greatly decreased with CCSSTF implemented,while it can be guaranteed that the performance was not reduced or even improved.Therefore the preparation time of subsequent experiment can be greatly reduced in practical applications.The ERP signals are often disturbed by background EEG noise,so the signal-tonoise ratio(SNR)of single trial signal is very low.Moreover,the ERP signals at adjacent time points had redundancies in the temporal domain.Therefore,an algorithm termed spatial filtering and temporally down-sampling(SFTDS)was proposed for ERP feature extraction in this thesis.SFTDS was solved efficiently within an SNR maximization framework,with the spatial filters and the weighted down-sampling vectors could be optimized simultaneously.In addition,in order to improve the generalization ability,a7)2-norm regularization term was added in the objective function of SFTDS.Compared with the traditional ERP feature extraction method and the spatio-temporal discriminant analysis algorithm(STDA),the classification performance of SFTDS was significantly improved,therefore SFTDS had a good prospect in online applications.Finally,to ameliorate the performance of the model under small sample setting,a Riemann distance-based transfer learning(RDTL)method was presented in this thesis.The designation of spatial filters for the classical feature extraction method CSP,was based on the joint diagonalization of the estimated covariance matrices under the two states.When the number of training samples was relatively small,the estimation of the covariance matrices was usually not accurate enough,which resulted in the variation of the estimated model.Therefore,it was of great practical significance to improve the performance of the model by transferring discriminant information from other subjects.The selection of auxiliary samples was very challenging because of the large variations between subjects and the nonstationary nature of EEG.Riemann distance was employed by RDTL as the criterion for selecting auxiliary samples,and the model of the target subject was designated by the training samples and auxiliary samples.Experimental results showed that,RDTL could select auxiliary samples effectively from other subjects when there were few training samples for the target subject,and the performances of the spatial filters and classifiers optimized by RDTL were significantly improved.The algorithms proposed in this thesis were all applied to real EEG data to verify the effectiveness.Compared with the existing algorithms,the algorithms of this thesis have the advantages of higher performance and lower computational complexity,and could be applied to the online BCI systems.