Research On Subject-Independent BCI Based On Event-Related Potential

Compared with conventional subject-depend brain-computer interface(SDBCI), a subject-independent BCI(SIBCI) can be ready-to-use for a new user with less or even no calibration procedure. The shortened calibration process means time and energy saving and improvement of subsequent use effect for patients with motor function damage. Meanwhile, it is more in line with the requirements of BCI for the large population of healthy users. However, the existing research on event-related potential(ERP) based SIBCI were only implemented in a small amount of subjects and the real time performance has not been explored yet. Therefore, the feasibility of SIBCI and its influence factors are needed to be further explored.Focusing on the above problems, this thesis designed two experiments. In offline experiment, Farwell paradigm was used to visually evoke ERP signal. 55 subjects were recruited for the experiment. Coherent average algorithm was used for extracting ERP time domain features and an ensemble of Linear Discriminant Analysis(LDA) classifiers was used for character identification. The cross-subject classification accuracy was 86.64% and it showed the feasibility of classification across individuals for BCI. 65.45% of the subjects got accuracy of above 90% which indicated that for most subjects the calibration process could be even omitted. The influence factors of the cross-subject performance were further explored. The performance could be improved by adding information from the testing subject into SI models. It fastened the calibration procedure with saving 93.93% time compared to the SD models. Besides, the results showed that the number of people for training and gender also had influence on the cross-subject performance. In order to further verify the offline analysis conclusion, an online real time experiment with the same paradigm was implemented among 20 subjects. SD, SI and Semi SI models established by data one week ago were used to test the real time data and got accuracy with 97.5%, 87.7% and 91.1% respectively. It demonstrated the individual stability of ERP signal and the feasibility of SIBCI in real time using. The generalizability of SIBCI was explored on two databases from other experiments and the scope of application and general performance of SI model was tested.In conclusion, performance of SI models were explored from various perspectives such as individual, time and experimental paradigm. It provides the theory and technology support for the further development of universal BCI application.