Probabilistic Methods Applied Research In Multi-task Eeg Brain-computer Interface

The goal of a Brian Computer Interface (BCI) consists of the development of a novel communication channel that translates the user's intention into control signals for peripheral equipments via brain signals. The common BCI systems contain the processes of data acquisition, feature extraction and pattern recognition. In practical BCI Systems, precision and information transfer rates (ITRs) are two important factors to measure their performances. Especially, ITRs is directly relevant to the real-time communication between human and the output device such as a computer application. At present, however, BCI applicability is seriously limited by its low ITRs and low accuracy. A feasible way to solve this problem is to change the usual binary decision to a more diverse decision. But, Great difficulties in both signal processing and machine learning stand out when the number of brain patterns increases. Thus BCI research is required interdisciplinary cooperation and only by combining machine learning and signal processing techniques based on neurophysiological knowledge will the largest progress be made.In this work, I mainly deal with the combination of signal processing and machine learning which lies in the area of pattern recognition. So, firstly, we have adopted support vector machine with posteriori probability (PSVM) method and Bayesian linear discriminant analysis (BLDA) probabilistic output model to obtain probabilistic voting results. Then, with the posterior probability or probability, two methods named EPSVM and EBLDA were adopted for increasing the performance of BCI. The main procedure of EPSVM was increase the size of training sets by adding test samples with big posterior probabilistic of SVM ,and the EBLDA is increasing the size of training set with the probabilistic output of BLDA .We test those methods on two four class motor imagery datasets. The feature vectors were extracted by multi-class common spatial patterns (CSP) algorithm and probabilistic classifiers were used for classification and augmented the size of train sets respectively. We also compared the two traditional multi-class classification methods (one-versus-rest SVM and Mahalanobis distance classifier) with these probabilistic classifiers. Meanwhile, we investigated the combination of time-frequency-spatial feature and the classifiers combination. the results showed that: (1) probabilistic information can improve the performance of BCI; (2) probabilistic information can be utility to enlarge the training dataset by adding test samples with big probability, thus to get an even more robust classifier; and these new expanding methods can reduce the training process and increase machine learning adaptation. (3) The combination of features and the multi-level classification can improve the performance of BCI, and this is considered to be a hot topic in the future research.