| Brain computer interface (BCI) builds an independent communication channel between the brain and external devices without using brain’s normal output pathways of peripheral nerves and muscles, and acts as a new interactive way. BCI technology has a widely application prospect in rehabilitation, auxiliary control, neurorobotics, entertainment and so on.In non-invasive BCI research, electroencephalography (EEG)-based BCIs are the most common for studies, because the recording method is safe, convenient and inexpensive. Various types of EEG signals have been used in BCIs. This thesis takes motor imagery EEG data and P300potential as key point, systematically describes the framework of the BCI, offline algorithm and online realization, and achieves some good results.BCI is a complex system integrating hardware and software platforms. Now, several hardware and general-purpose software are available to build an online system rapidly. On the base of platform comparison and existing experimental condition, Neuroscan hardware devices and BCI2000software are combined to build an online BCI system. The thesis provides some online experimental experiences for future research on BCI by the secondary development of BCI2000.Signal processing algorithm converts brain signals into control signals, the results of algorithm directly affect the performance of BCI system. The thesis presents a study on some frequently used signal processing method in detail. Due to the low information transfer rate and low recognition accuracy in BCI, the thesis proposes an algorithm based on common spatial pattern (CSP), Hilbert transformation and support vector machine (SVM) for feature extraction and classification of multi-channel four-class motor imagery EEG signals. With the datasets of BCI competition III, experimental results show that the proposed algorithm produces high classification accuracy and less time consumption, moreover, classification result can be further improved at the expense of algorithmic complexity by the threshold adjustment. With the increase of threshold, average classification accuracy is improved from87.22%to92.22%while the time consumption is increased to2.15times.To choose the best solution, different feature extraction and classification algorithms are usually considered in BCI experiment. However, there is complementary information between different methodologies or different features that can improve the classification performance. In order to use the complementary information efficiently, the confidence coefficient is proposed to determine classification difficulty levels of samples in a classifier, and then combination of multiple classifiers based on confidence coefficient is proposed in this thesis. The technique uses different classification strategies by measuring base classifier outputs values to utilize the complementary information. Meanwhile, the execution speed is taken into account. We perform comparative analysis of five feature extraction methods and their combinations on EEG signal from BCI competition IV. Experimental results show that the proposed combination method could utilize complementary information between multiple classifiers effectively. The average kappa value of our combination method is0.63which has15.38%of rising than that of all base classifiers.The P300event-related potential (ERP), with advantages of highly stability and not need initial training, is one of the most commonly used in BCI applications. The traditional row/column paradigm (RCP) that flashes an entire column or row of a visual matrix has succeeded in helping patients to spell words. However, the RCP remains subject to errors that slow communication, such as adjacency-distraction errors and double-flash errors. After extensive research, a new visual stimulus presentation paradigm called the submatrix-based paradigm (SBP) is proposed in this thesis. The SBP divide the whole keyboard matrix into several submatrices, and each submatrix flash in single display paradigm (SDP) mode and performs ensemble averaging method according to sequences separately. With an increasing number of sequences, SBP can detect the target submatrix and the attended item simultaneously. Compared with RCP, SBP can improve the practical bit rate by10.8%. Besides, the SBP has advantages of eliminating adjacency-distraction errors and double-flash errors, flexible division of matrix, better expansion and user acceptability.To further improve the information transfer rate and realize online BCI system, an adaptive SBP online BCI is proposed. The adaptive algorithms employ a threshold which dynamically limits the number of sequences. Two threshold algorithms named maximum value algorithm and pseudo-kurtosis algorithm are proposed and compared in the thesis. Online experimental results show that both adaptive algorithms are effective, and the highest practical bit rate of maximum value algorithm is34.36bits/min which is9.04%higher than that of checkerboard paradigm (CBP) and11.2%higher than that of Jin’s n-flash adaptive system (NFA). |
PDF Full Text Request