P300-based Brain-computer Interface And Its Online Semi-supervised Learning

BCI technology has been a hot topic in the fields of artificial intelligence, patternrecognition and signal processing, and has great effect on the disabled who lost the physicalability. At present, although the technology has been applied to medical, entertainment, etc.,there are still some problems, e.g., low communication rate of the system and long trainingtime. To solve these problems, the major objectives of this thesis include:First, an improved P300speller BCI system is designed in this thesis, where the possibletargets and non-targets are recognized according to the subject’s real-time EEG data, and thenthe flashing of recognized non-targets are blocked to shorten the flashing sequence andweaken the interference to subject. Through the analysis of EEG signals collected fromsubjects, the experimental results show that by decreasing the number of the flashingcharacter to shorten flashing sequence, the input speed of characters is improved while theaccuracy is basically not declined. Therefore information transfer rate (ITR) is improved,which is helpful to solve the practicality of the system.Second, to solve problem that the training is time-consuming, semi-supervised learningcan be considered to reduce the time of collecting labeled data, and here we study an onlinesemi-supervised learning of P300BCI system. Generally, semi-supervised learning uses smallamount of labeled data to train an initial model and relatively large amount of unlabeled datato update the model. In this thesis, the model is updated by the unlabeled data obtained onlineafter the described initial modeling. In order to get the reliable model, we choose normalunlabeled samples and abandon the others so as to reduce the destruction on the model by theabnormal data. For the sample selection problem, we consider the reliability of each unlabeleddata by calculating its classifier response. Thus we can obtain reliable unlabeled data toupdate the classifier gradually. In the experiment, we respectively compare these two methodsof sample selection and non-selection by conducting some offline analysis of online testingdata. The experimental results show that, in the online semi-supervised learning BCI speller system, the proposed sample selection method can improve the accuracy though with smallamount of labeled data.