کنفرانس مهندسی برق ایران

صفحه اصلی / سی امین کنفرانس بین المللی مهندسی برق

Application of Transfer Learning in Optimized Filter- Bank Regularized CSP to Classification of EEG Signals with Small Dataset

نویسندگان :

M. Moein Esfahani¹ Hossein Sadati²

1- Faculty of Electrical Engineering K. N. Toosi University of Technology Tehran, Iran 2- Faculty of Electrical Engineering K. N. Toosi University of Technology Tehran, Iran

کلمات کلیدی :

Brain-Computer-Interface،BCI،EEG،Motor Imagery،FBRCSP،Common Spatial Pattern

چکیده :

Application of Brain-Computer Interface (BCI) systems to develop a path between brain and external devices, such as Electroencephalography (EEG) signal acquisition, is extensively under study in regard to brain electrical activities. EEG is an inexpensive brain cognition and imaging method with high temporal resolutions for feature extraction in Motor Imagery tasks. The common spatial pattern (CSP) and its optimized algorithms are effective methods for discriminating and classifying EEG Signals. To classify motor imagery tasks in EEG signals, we need to implement the CSP algorithm to extract features and discriminate spatial patterns based on movement tasks in two-class motor imagery signals. Furthermore, owing to the amount of noise in EEG signals and the limited number of trials per subject, we need to optimize the conventional CSP algorithm by adding a penalty term in the denominator of the CSP cost functions. In this study, due to differences in each subject's neural activities, we employed transfer learning which used the information for other subjects to regulate features of the subject. Additionally, BCI Competition III dataset IVa was analyzed. Furthermore, this study presents the optimized Filter Bank Regularized CSP algorithm with Transfer Learning to perform the classification of the electroencephalography (EEG) motor imagery signals. Moreover, to compare the efficiency of the proposed algorithm, the conventional CSP and the proposed optimized CSP have been weighed, and results for both methods are presented. The results at the end explain that the classification with 10-fold cross-validation in comparison with that of the proposed method achieves approximately 15% and 21% higher accuracy against the R-CSP and conventional CSP, respectively.