چكيده به لاتين
A brain-computer interface system (BCI) based on electroencephalogram signals creates communication pathways between the human brain and external devices. P300 signals are one of the brain signals types used for BCI systems. It is a visual evoked potential that appears in a person's brain signal, 300 milliseconds after stimulation. One of the most critical factors which have always been important in P300-based BCI systems is communication speed and accuracy. In this study, the keyboard control was performed online with a P300 signal. An innovative graphical user interface with nine keys has been designed as the main menu. One key to turn the system on and off, and eight other keys to support 56 keyboard characters, including letters, numbers, punctuation, and features such as delete, space, enter, uppercase, lowercase, and open the word page. First, in the feature extraction section, the super-covariance matrix of each trial calculated to enter a symmetric positive definite (SPD) space. This space helps distinguish well between the properties of the target and non-target classes and allows the use of Riemannian geometry relationships. Then we used the Riemann graph dimension reduction method of classes to reduce the super-covariance matrix dimensions, which simultaneously brings the same classes closer together and separates the dissident classes. The geometric filter -minimum distance to Riemannian mean (fgMDRM) method was used in the classification section.
The proposed method applied to dataset IIb from BCI competition II and dataset II from BCI competition III. We achieved better results compared to the proposed methods of recent studies. Online and offline experiments were performed in two sessions on six human volunteers aged 30-24. Compared with four methods of convolutional neural network (CNN), Linear Discriminant Analysis (LDA), Support Vector Machines (SVM), and the minimum distance to the Riemannian mean (MDRM). In all cases, the proposed method appeared more robust. We were able to show the discriminative of the two classes in the form of a scatter plot for all six candidates. In online experiments, asked people to write the phrase "Brain-Computer Interface (BCI) {Enter} 2020" with two or three trials. The average classification accuracy 95/83 and 96/33 for two and three trials, respectively, and the average data transfer rate (ITR) were 71/76 and 48/47 for two and three trials, respectively. Comparing the online results with other recent studies show a 1.83% increase in accuracy and 14.76 bits per minute in ITR.
