The Extraction Of Spontaneous EEG Based On Time, Frequency And Spatial Information

For people, Brain Computer Interface (BCD offers a new communication channel between brain and the external world. More and more research groups have focued on this area in recent years. However, the development of BCI technology is hampered by such factors as the following: the EEG signals are very weak; they are easily affected by the environment and they change sharply for different people. So the aim of this dissertation is to find the right and effective experimental model to analyze the EEG signals through signal recognizing. Based on brain neurophysiology and the relationship between spontaneous EEG and nerve cell activity, the dissertation does a research on the problems in present BCI application, and some algorithms are improved. The followings are the main contents of the research.Firstly, different special filter algorithms are studied and compared. These algorithms are analyzed in order to find effective ways reduce the effect of skull on brain signals and increase nerve cell activity characteristics. The two special filter algorithms—principal component analysis algorithm and blind signal separating algorithm are presented and compared with each other in term of capability of removing the artifact from EEG. Their advantages and disadvantages are also presented.Secondly, The Laplacian filter and blind signal separator are associated with HMM-AR algorithm, which can automatically segment a time series in BCI into discrete dynamic regimes. The HMM-AR model is implemented to find the variation state of brain signals frequency.Thirdly, according as the neurophysiology theory, the data sets of two tasks in BCI are analyzed. The algorithm of common special pattern based on support vector machine is presented to select special filter. The result proves the availability of the common special pattern based on support vector machine.Fourthly, the common special pattern algorithm is extended to multi-class. An extension of the CSP algorithm to multi-class paradigms is introduced with an elaborate description. The experimental result illustrates the performance and validity of the algorithm.Lastly, in order to improve the accuracy of the brain state classification, a time- frequency-spatial filter algorithm is put forward. By combining the frequent character and the time character of brain signals, the common spatial pattern is extended. A new algorithm, named the Time-Frequency-Spatial Filter, is presented, and its feasibility is proved by applying it to a data set. The experimental result illustrates the performance and validity of the algorithm.The originalities of this thesis are the followings: (1) According to the characteristics ofμandβbrain rhythms, a method jointly employing Laplacian filter, ICA transform and HMM-AR is presented for EEG pattern classification. The hybrid method can analyze the BCI data sets in time field and find the variation state of brain signals frequency. So it can find out whether the brain state is "movement" or "rest" through Electroencephalogram (EEG) .(2) A new algorithm is put forward as an extension of the Common Spatial Pattern (CSP) to multi-class cases. In order to implement a suitable Brain-Computer Interface system, it is applied to BCI data sets. The experimental result proves that it can not only improve the information transfer rate, but also ensure a certain accuracy.(3) A Time-Variant Spatial Filter algorithm is put forward. In this algorithm, the parameters of spatial pattern filter are not fixed, but variable with time. This algorithm can improve the accuracy of the brain state classification.(4) A Time- Frequency-Spatial Filter algorithm is put forward to classify EEG signals in small training sets. This algorithm is based on the time-variant spatial filter, but it emphasizes specific frequency bands by simply concatenating input signals and their delayed counterparts.Through the works mentioned above, the thesis puts forward a technology system suitable for the extraction of spontaneous EEG signals. Its function includes: removal of the artifact from EEG, identification of the beginning movement of brain states, feature extraction and classification of brain-computer interface system. The research achievements of the thesis have been applied to internationally published data stets, and made very good performances.