| Brain-computer interfaces (BCI) technology collects and processes the men’ or animals’EEG signals during their thinking activities. It translates their intentions as control signals tomanipulate external devices, such as computers, manipulator or wheelchair, etc., withoutusing the information transmission channel of nerves and muscles. Brian-computer interfacestechnology is hot area of research for control and nerve repair, signal processing, artificialintelligence and pattern recognition and more other relevant technique. BCI technology has ahigh application value in improving the subjects’ live quality, who with severe neuromuscularfunction but thinking normally. Since the characteristic of EEG signal, such as randomness,low signal-to-noise ratio, high dimensions, relative-un-abound training samples, easilyinterfered, and so on. It leads to low classification results of the EEG signal and blocks thepractical application of BCI technology. Based on the characteristics of EEG signal, threeparts of BCI signal processing method were studied. They pre-processing, feature exactionand classification algorithm. Main procedures and contributions are included as follows.Firstly, in order to enhance signal-to-noise ratio, the common average reference (CAR)spatial filter is used to reduce the noise that are common to all electrodes. Then, we proposeda method for extracting optimum frequency band from the filtered EEG signals. Theexperiment results on Data set IVa of BCI competition III show the method can get optimalpass-filter to extract more effective features, and compared with the CSP features through8-12Hz pass-filter, the classification results of features for all subjects are higher5.8%inaverage.Secondly, we propose an optimal channel set selection algorithm based on improvedgenetic algorithm (GA) and maximization of Rayleigh Coefficient (RC). Dataset IVa of BCIcompetition III and dataset of our laboratory are applied on the algorithm, which reduces agreat number of irrelevant channels, and improves the classification results. In this paper, thealgorithm will get optimal channel set by35iterations in average for dataset IVa of BCIcompetition III and25iterations in average for dataset of our laboratory.Lastly, according to the requirement of reducing train samples and shortening thetraining time at the BCI technical developing process, therefore, we propose a semi-supervised model based on RC and graph based semi-supervised learning for theanalysis of EEG signals. In this model, both training data set with known labels and the testdata set without labels are involved. And we use a method based on RC to set the parametersof the semi-supervised model. Then, we compare the classification results of the methodwith CSP and Support Vector Machine SVM method with different number of trainingsamples. The experiments results show that better results are obtained with our proposedmethod at different training sample size.In this thesis, all signal processing methods are proposed based on motor-imagery BCItechnology, and they are realized and simulated by MATLAB. Two different datasets areapplied on those methods, and the results validate the effectiveness of our proposed algorithm. |
