Research On ISVM Classification And Design Of Feedback And Application System In BCI

A Brain Computer Interface (BCI) is a communication channel between humanbrain and the outside world, which does not depend on peripheral nerves andmuscles and other conventional output channels. It makes use ofElectroencephalogram (EEG) generated in the brain neural activity to achieve directcommunication and control between human brain and a computer or other electronicdevices, the essence of which is infer idea or purpose through brain signals.A BCI system mainly consists of four parts: signal acquisiton, signal processing,feedback unit and application unit. The signal processing includes feature extractionand feature classification. Based on spontaneous EEG, three parts are studied in thispaper: feature classification, the design of feedback unit and application unit. Thesespecifically include:1) Batch incremental support vector machine (BISVM) classification isdiscussed. BISVM is the intergration of incremental learning, SVM and batchtechnology. The problem of lack of continuity in training model can be effectivelysolved in incremental SVM (ISVM) classification. It takes full advantage of theresults of historical data, and update classifier by the new vector in real time.BISVM, which absorb batch idea, trains the data by ISVM group by group. It takesless training time, and improves system performance effectively.Based on dataset of2008BCI Competition IV and our experimental EEG data,features are extracted by WPD and CSP. Then SVM, ISVM and BISVMclassification are adopted. The classification results show that both ISVM andBISVM classification accuracy are significantly higher than SVM, and BISVMclassification accuracy is slightly higher than ISVM. In the case of similar accuracy,BISVM classification is much faster than ISVM.2)3D virtual human BCI feedback system is designed based on VRP software.The feedback system is expected to improve the adaptability of human brain to the computer, to train subjects produce the best EEG signals. Virtual human model isestablished in3Dsmax. Virtual human model actions and virtual scene functions aredefined in VRP. Left hand motor imagination and right hand motor imagination ofthe subjects are corresponded to left arm or right arm stretching exercise of virtualhuman. The feedback unit communicates with the BCI signal processing unit bydatabase. The system calls the database commands translated by EEG classificationresults in real time, and left or right arm stretching exercise is performed by virtualhuman relevantly. According to the action of virtual human, subjects adjust theirstatus actively to better motor imagination. With intuitive of virtual human feedback,subjects are more prone to EEG signals which are easier to distinguished in realisticenviroment, and the training time is expected to reduce.3)3D virtual car application is accomplished on Java3D. Firstly static scenemodels and car model are set up and scene layout is designed in Java3D. Then thebehavior of virtual car is defined. Communication and evaluation functions of thesystem are designed at last. Collision detection and statistical functions, whichcalculate the control accuracy of subjects, and the timer function is also added.Processed EEG signals are translated into identifiable control commands, which aresent by TCP/IP in real time, to control the virtual car movement. Strong left handmotor imagination and right hand motor imagination of the subjects control the carturn left or right, weak left hand or right hand motor imagination of the subjectscontrol the speed of virtual car. The real time system, with realistic3D virtualenvironment and ideal operating environment of the actual hardware system, lay thefoundation for the practical application of BCI.