Research On Brain-computer Interface Based On Analysis Of Nonlinear Characteristics Of Motor Imagery Electroencephalogram

Brain-Computer Interface (BCI) is a communication system that helps individualsto drive and control external devices using only their brain activity, withoutparticipation of peripheral nerves and muscles. BCI is an integrated disciplinesincluding neuroscience, signal processing and computer science etc. Over the past20years, BCI has become a research hotspot in the field of international intelligent science.The core of BCI research is how to translate the user’s EEG signals into the controllingcommands for the external devices. So the most important work of the BCI research isto find the proper signal processing and translating method, which can help todistinguish the mentality tasks by the computer quickly and accurately. Generally, theBCI system can be seen as a pattern recognition system. So a successful EEG-basedBCI system very much depends on whether the following two requirements can besatisfied:①The extracted EEG features are able to differentiate the task-oriented brainstates; and②The methods for classifying such features in real-time are efficient. Howto improve the performance of the features and design an efficient classificationalgorithm are two key points.Nowadays, during the research of BCI based on motor imagery, many featureextraction methods and classification methods applied to EEG generally assume that theEEG signal is linear. However, many researches have shown that the EEG signals arenonlinear. To analyze the EEG signals by linear method will loss much nonlinearfeatures, then the capability of distinguishing different tasks will be decreased. Thispaper, we proposed a new feature extraction method based on the nonlinearcharacteristics of EEG. Then we verified the effectiveness of the new algorithm bysimulations. This paper includes the following aspects of content.①We analyzed the nonlinear characters of EEG dynamical model. Using phasespace reconstruction technology to reconstruct the EEG signals obtained in the EEGmodel. Then we learned the changing principle of the attractors ofhe with the changesof parameters ofp eeandpe i. We also analyzed the nonlinear characters of the realEEG. The maximum Lyapunov indexes of the samples in the two BCI competition’sdatasets were calculated. The results show that almost all the maximum Lyapunovindexes of the samples are bigger than zero, which confirmed the argument of chaos inthe EEG. So the nonlinear analysis method can be used to analyze the EEG signals. ②Some normal chaos features, maximum Lyapunov index, correlation dimensionand Approximate Entropy (ApEn), are calculated. Those three characters were used asthe features of the BCI. The experimental results showed that the maximum Lyapunovindex and correlation dimension do not suit for differentiating the motor imagery tasks.However, the Approximate Entropy is a measure of the probability for producing newmodels in time series; it is more suitable for distinguishing different tasks. Based on theanalysis of the ApEn features, this paper proposed a feature extraction method andclassification method based on ApEn and time window. The proposed method cansimulate the online situation, and classify the tasks in each window. The simulationshows that the classifiers based on theApEn features can distinguish the tasks well.③We proposed a feature extraction schemes based on phase space reconstruction.We proved the construction functions have the filtering capacities which can adjust theamplitude and phase of the signals. Then the phase space features can be distinguishedbetween different tasks more easily. Moreover, the features extracted in the proposedscheme can retain the merit of traditional features, at the same time they also contain theinformation of phase space, so it can improve the classification capabilities of theclassifiers. This paper used the benchmark datasets coming from the BCI competition2003and2005. The mutual information (MI) and the maximum steepness of MI areused as the evaluation standards which are also used in the BCI competitions. Thesimulations show that the proposed method is a competitive method. The classifier gotthe maximum MI0.67on the Graz2003dataset, which is the best result ever known.The simulations on Graz2005dataset also obtained some good results. The phase spacefeatures have the good results based on accuracy criteria and mean maximum steepnessof MI.④In order to solve multi-classification problems using common spatial pattern(CSP), we proposed a combination method based on binary tree termed BCSP, whichput the CSP filters and Fisher classifiers on the nodes of the binary tree. Theclassification process bases on binary search. In BCSP, the number of the filters andclassifiers is less than "one versus rest" method. Moreover, The most calculating stepsto solve the N class problems islog2N, which improves the efficiency and results ofclassification largely.⑤Developing and realized a BCI online game platform. We developed an onlinegame platform based on the research of this paper. The platform is a game system basedon Neuroscan, and user can play Hangman game by imaging left/right hand moving. The platform has training module, testing module and game module. The user can use itto complete all the training, testing and play actions. C3, C4and O1channels are usedto gather the EEG signal, and the EEG from C3and C4are used to extract the featuresof right/left hand motor imagery tasks. We also used rhythm in O1to act as theconfirming signals. Phase space features and Fisher classifiers are used in this system.Six people took part in the experiments, the experiments results show that the phasespace features improved the classification accuracy, and then it proved the effectivenessof the phase space features.At the end of the dissertation, we summarized the research contents of this paper,and show the main achievements of the research. Finally, the author points out the nextwork of the future researches.