Study Of Online Semi-supervised Learning For P300-Based Brain-Computer Interface: The Algorithm And The System

A brain-computer interface can establish a direct communication pathway between a brain and the environment, which does not depend on individual muscle and nerve tissue. Therefore, it’s mainly used for the people who have movement deficits to do functional auxiliary and serve the patients with neurological disorders to do rehabilitation training. After decades of development, the system framework and fundamental methods of BCI have been established. But, to meet the request of practical applications, there are still a large number of problems to be solved. In order to reduce time-consuming training and satisfy the online requirements of high classification accuracy and real-time output, in this thesis, we present our work from two aspects:1) A semi-supervised linear discriminant analysis algorithm, namely SUST-ILDA, was proposed based on linear discriminant analysis(LDA). In order to reduce the training effort before a BCI can be put into use, the semi-supervised learning approach was used. Firstly, a small amount of labeled samples were used to train an initial classifier, and after that, the online unlabeled samples were collected to update the classifier sequentially through a method called self-training to improve the performance of the classifier. In order to reduce the online computational complexity, the incremental updating form of LDA was derived. Theoretically, compared with the existing online semi-supervised learning algorithm- SUST-LSSVM,the proposed algorithm reduces the computational complexity significantly and keeps stable. The experimental analysis was carried out using the third BCI competition data and proved that the proposed algorithm can achieve the similar classification accuracy like SUST-LSSVM, as the number of the online samples were increased. At the same time,it keeps convergent and stable.2) An online semi-supervised speller BCI system was designed and implemented based on P300. Compared with the traditional supervised P300-based BCI speller system, the characteristics of this system were: 1) After short-time supervised training, it can automatically switch to input mode, which greatly reduces the boring and tedious training effort. Meanwhile, the classification accuracy was continuously improved to be stable with the continuous input process. 2) During the normal used phase, the online unlabeled samples were collected to update the classifiers sequentially. Consequently, to the extent, it can adapt EEG’s non-stationary changes(through experimental observation, however, the theory has not been proven). However, the traditional speller system no longer updates the classifier during its work. Thus, the performance of the classifier is reduced. Moreover, it influences the performance of the system. The system takes the advantages of the higher classification accuracy in the cases of small training set, as well as the very low computational complexity and high classification accuracy of SUST-ILDA in the cases of adequate samples. It was implemented through multi-process and multi-thread, along with updating the dual classifier.