Key Technology And Experiment Research On EEG-Based Brain-Computer Interface

Brain-computer interface (BCI) is a communication system between brain and computers directly.It does not depend on the normal output pathways of peripheral nerves and muscles and reflects human’s mental activity by EEG, it is a new communication and control method. BCI research has been a focus in brain-science, human machine automatic control, rehabilitation engineering and biomedical engineering.It is not only an important way to understand and improve human’s brain function, but also can help the person who is deformity or severe motor disabilities to control external device and improve the quality of their life.It can also help normal persons as the auxiliary control approach in special conditions. It has wide application prospect in many areas.EEG is weak electric signal produced by cerebral cortex, it can be recorded from scalp by electrodes.The essential of BCI is to recognize particular pattern of EEG and transform it to external information as the advance rules. BCI includes EEG recording, pre-processing, feature extraction, feature selection,feature classification and external control.For all BCI technologies, EEG recognization is the most important part, it concerns whether the input can be transfer to output correctly.This paper studies EEG pre-processing, feature extraction, feature selection,feature classification deeply, just to improve the BCI classification accuray rate,and prove the availability of the method in this paper by BCI competition data.EEG pre-processing is to remove the artifacts in EEG, it can provide good source signal for later disposal. This paper presents a pre-processing method which combines the independent component analysis (ICA) and time-frequency analysis.The aritfacts can be represented by independent component, and use time-frequency analysis method to find out the artifact component. FastICA and extended Infomax algorithm are used to deal with EEG data, Choi-Williams distribution is used to get the artifact component, at last make the artifact component to zero and resume the EEG data. Analyze the EEG after artifact removing, the result shows the method can remove EOG and 50Hz interference component effectively.The purpose of feature extraction is to transform EEG to feature vector which is the representiton of different mental task, it is the most important part in EEG recognization because it affects the feature selection method and classifier design.In this paper, according to the sub-band of wavelet packet(WP) energy method, combine the EEG data in each channel with different ways, get the sub-band of WP energy as the feature,and find out that the relationship between channels can help to improve classification accuray rate.Finally a method called common spatial pattern(CSP) based on sub-band of WP is presented, this method combines the frequency of EEG and the relationship between channals, and it can get a better result.2005 BCI competition Data Set IVb is used to prove the effectivity of the method.The purpose of feature selection is to select more effective features in feature extraction result. The purpose of feature classification is to yield the corresponding class label according to the input feature. The original featue is extracted by CSP based on sub-band of WP method in this paper.The feature selection method called optimal gene chromosome mutation is presented based on GA,the method uses k-nearest neighbors (KNN) classfier as the standard of good chromosome,and get a better result for classification accuracy rate to prove the effectivity of the method, and later use learning vector quantization(LVQ) and support vector machine(SVM) as the the standard of good chromosome.The experiment result compares the performace of the three kinds of classfiers.An experiment platform and experiment method called motor imagery BCI based on different frequency sound stimulus is presented in this paper.The method presented before is used to deal with the 16 channals EEG data which recorded by the experiment. The results of many subjects show that though the channals are reduced to 16, it can also get a good classification accuracy rate, and the speed is faster than before.It is a good foundation to research on-line BCI.