Spatio-temporal Feature Analysis Of EEG And Application In BCI

The BCI system aims at creating new direct information interaction and communicationchannels between brain and computer without depending on brain's normal output channelsof peripheral nerves and muscles. The BCI research has drawn attention of scientists in brainscience, rehabilitation engineering, biomedical engineering and intelligent information pro-cessing. In this paper, we mainly focus on non-invasive BCI systems based on EEG signalsfrom which several motor imagery (MI) tasks can be recognized. Due to the non-stabilityand weakness of EEG signals, it is very difficult to extract reliable features from EEG signalswith high noise, especially for the spontaneous MI related EEG. According to the charac-teristic of EEG signals, we mainly research on several aspects including experiment design,neurophysiology mechanism, feature extraction algorithms and online BCI systems, and em-phasize particularly on EEG spatio-temporal feature extraction.The main contributions and innovations of this paper have been listed as below:(1) Neurophysiology mechanism. We investigated the relationship between MI andduration of ERD/ERS. Experimental results have demonstrated that repetitive fixed MI canproduce sustained ERD/ERS, which can be viewed as Task Related Sustained Desynchro-nization/Synchronization (TRSD/TRSS). The key advantages of TRSD/TRSS are not onlynon-phase-locked but also non-time-locked to the cue stimulus. Hence, it is more suitablefor asynchronous BCI.(2) Temporal spatial pattern (TSP). We apply independent residual analysis to extractindependent components with temporal structure, from which we can further extract optimalspatial patterns. Finally, feature selection is performed using mutual information betweenlabels and features. Therefore, the optimal temporal and spatial features of EEG signals canbe obtained simultaneously and classification accuracy has been improved.(3) Common spatial frequency pattern (CSFP). In order to consider the effects of fre-quency patterns for classification of MI during spatial pattern calculation. The proposed CSFP algorithm allows the simultaneous optimization of spatial and frequency patterns en-hancing discriminability of EEG signals by time-frequency analysis based on continuouswavelet transform. Experimental results has demonstrated that better performance can beachieved by CSFP when compared to common spatial pattern (CSP).(4) Incremental common spatial pattern (ICSP). To deal with the non-stationary EEG,feature extraction algorithm must has self-adaptive ability. This paper proposed a novel ICSPalgorithm which can update spatial patterns in real time by incremental learning manner.Hence, it is very suitable for online BCI system.(5) Non-negative tensor sparse factorization (NTSF) and Common tensor discrimina-tive analysis (CTDA). Tensor factorization has been widely focused recently. Based on thenon-negative tensor factorization algorithm, we proposed a new sparseness constraint con-dition and developed the non-negative tensor sparse factorization (NTSF) algorithm whichhas been used in feature extraction of EEG. By the sparseness on the condition mode, themaximal discriminative tensor bases on multi-modes can be obtained and optimal feature co-efficients have been achieved by projecting EEG to tensor bases. Furthermore, we extendedthe CSP based on matrix operation to tensor sense and proposed CTDA algorithm which candiagonalize high dimension covariance tensor of multi-class EEG and obtain common tensorpatterns and maximal discriminant features for classification.(6) BCI game and"Mind-driven Car". We developed two novel BCI applications.The first one is hit-rat game by different MI tasks (synchronous BCI). We have applied slid-ing window techniques to achieve fast response for BCI and analyzed classification perfor-mance with different window length. Furthermore, this game will open a new entertainmentapplication for BCI. The second one is driving a car in 3D Virtual Reality Environmentsby thought. The duration of ERD/ERS caused by MI can be modulated and used as an ad-ditional control parameter beyond simple binary decisions. By this strategy, the complexcontrol functions can be achieved such as control of steering wheel angle and car speed.Furthermore, by cumulative incremental control strategy, the steering wheel rotates moresmoothly, which make the BCI system has error tolerant ability. This system is an asyn-chronous, self-paced BCI which provides a more natural interaction manner.In summary, this paper have investigated EEG patterns and dynamic features duringspecific mental tasks, and proposed several novel feature extraction algorithms which canextract temporal, spatial, and frequency patterns related to MI from complicated EEG sig-nals. We have improved classification accuracy and response speed of MI based BCI bythe novel feature extraction algorithms. The best information transfer rate of 0.55bps can be achieved by our methods. Furthermore, based on these algorithms, we developed anasynchronous, self-paced, real-time BCI system which provided more complicated controlfunctions. These results further reveal neurophysiology mechanism corresponding to MItasks, and provide technology prototype and theory basis for new BCI applications.