Study On EEG Feature Exaction And Pattern Recognition Based On Time-Frequency Methods

Electroencephalography (EEG) has obvious advantages of non-invasion, easy to acquisitionand good time resolution. So it plays an irreplaceable role in the research of neural informationengineering, especially in a brain-computer interface (BCI) application. Because EEG signal isnon-stationary, nonlinear and with low spatial resolution, more and more attention has been paid tothe study of characterization and classification methods for EEG signals.For the EEG-based estimating of mental fatigue, an experimental scheme was firstly proposedto speed up fatigue and a self-rating scale was developed to measure the severity of mental fatigue,and then the EEG of three fatigue status from non-fatigue cases to severe fatigue cases weresampled. A new method for non-stationary and nonlinear data processing, the Hilbert-Huangtransform (HHT), was applied to investigate the characters of the electroencephalogram for fatigued.The fatigue index was developed based on hilbert marginal spectrum. Compared with Short-timeFourier Transform (STFT) and Wavelet Transform (WT), HHT has better resolution bothin time domain and in frequency domain. Furthermore, empirical mode decomposition (EMD) hasmore adaptive ability for non-stationary EEG signals.Subsequently, this paper presented a new methodology for automated sleep stage identificationbased on the time frequency distribution of electroencephalogram using the Sleep-EDF databasefrom MIT-BIH. Considering intrinsic mode functions (IMF) might cover too wide a frequencyrange such that the property of mono-component could not be achieved with traditional HHT, a newtechnique was presented which was combined with Wavelet Package Transform (WPT), and thenthe improved HHT method was applied to the identification of sleep stage. WPT would improve thenarrow band behavior of signal frequency. The bandwidth refinement and instantaneous energycalculation were implemented to realize automatic sleep-stage determination. The average stagingaccuracy obtained by the new method was up to 87.37%, higher than that from original HHT.Finally, EEG-based BCIs have become a hot spot in the study of neural engineering, andidentifying the brain activity through EEG has become a major challenge in the design ofefficient BCIs, therefore the time-frequency method was be applied to the classification of motorimagery tasks for BCI competition. On the basis of Event-related Desynchronization (ERD) andEvent-related Synchronization (ERS) features extraction from EEG by WPT and HHTtime-frequency analysis method, this paper proposed the weight optimal and screening algorithmsfor IMFs to reduce high dimensions during feature extraction and negative influence of IMFs with low correlation. The results showed that this new method improved in accuracy and rapidity whenidentifying ERD/ERS phenomenon.