1、EEG SIGNAL PROCESSING EEG signal modelling 1 Available features 2 Classification algorithms 3 Independent Component Analysis 4 CONTENT Sparse Representation 5 1 EEG signal modelling Bioelectricity 1 Signal generation system 2 BIOELECTRICITY SIGNAL GENERATION SYSTEM Excitation model SIGNAL GENERATION
2、 SYSTEM BIOELECTRICITY Linear Model SIGNAL GENERATION SYSTEM BIOELECTRICITY Nonlinear Model 2 Available features Basic features 1 Modern methods 2 Temporal Analysis Signal Segmentation:label the EEG signals by segments of similar characteristics.BASIC FEATURES MODERN METHODS Temporal Criteria BASIC
3、FEATURES MODERN METHODS Frequency Analysis Suboptimal DFT,DCT,DWT;Optimal KLT(Karhunen-Lo ve)Demerits:complete statistical information,no fast calculation.BASIC FEATURES MODERN METHODS Signal Parameter Estimation AR model:Merits:Outperform DFT in frequency accuracy.Demerits:suffer from poor estimati
4、on of parameters.Improvements:accurate order&coefficients.MODERN METHODS BASIC FEATURES AR coefficients estimation methods Yule-Walker aryule(x,p)Merits:Toeplitz matrix Levinson-Durbin,fastest!Demerits:with window bad resolution of PSD MODERN METHODS BASIC FEATURES AR coefficients estimation methods
5、 Covariance method arcov(x,p),armcov(x,p)Merits:without window good resolution of PSD Demerits:slow Burg arburg(x,p)Merits:accurate approximation of PSD Demerits:line skewing&splitting MODERN METHODS BASIC FEATURES MODERN METHODS BASIC FEATURES Comparison Principal Component Analysis Use same concep
6、t as SVD Decompose data into uncorrelated orthogonal components Autocorrelation matrix is diagonalized Each eigenvector represents a principal component Application decomposition,classification,filtering,denoising,whitening.MODERN METHODS BASIC FEATURES 3 Sparse Representation Sparse Approximation 1
7、 Sparse Decomposition 2 Over-complete dictionary atoms Hilbert space :Signal:Error:“Sparse”:lN,satisfy limited error.SPARSE APPROXIMATION SPARSE DECOMPOSITION,1,2,.kDd kKKNyHllrrr Iyd(,)inflllyy DyyNHRMajor algorithms:Basic Pursuit,Matching Pursuits,OMP Matching Pursuits(MP):1st:kth:SPARSE DECOMPOSI
8、TION SPARSE APPROXIMATION 0(1,.),rikiy dsupy d 001,rryy ddR y1(1,2,.),kkrikkiR y dsupR y d 10,nnknrrknyR y ddRy 与与 正交正交 nrd1kRyK-SVD:training dictionary Potential applications for EEG:Coefficients features ERP detection Abnormal EEG detection Classification of different status of EEG SPARSE DECOMPOS
9、ITION SPARSE APPROXIMATION 4 Classification algorithms Common methods 1 Na ve Bayes LDA:Linear Discriminant Analysis HMM:Hidden Markov Modelling SVM:Support Vector Machine K-means ANNs:Artificial Neural Networks Fuzzy Logic COMMON METHODS 5 Independent Component Analysis ICA approaches 1 Application
10、 2 Independent Component Analysis ICA APPROACHES APPLICATIONS ICA APPROACHES APPLICATIONS ICA approaches:Factorizing the joint PDF into its marginal PDFs Decorrelating signals through time Eliminating temporal cross-correlation function BSS:Blind Source Separation Normal brain rhythms,event-related sources Artefacts eye movement&blinking,swallow APPLICATIONS ICA APPROACHES THANKS!
