L1-NORM Based Robust Classification Of EEG Signals In BCI System

Electroencephalogram (EEG)-based brain-computer interfaces (BCIs) provide a new interactive manner for handicappeds with neuromuscular disabilities to communicate with the outside world by translating EEGs into control signals for external devices.One core issue in BCIs is to perform accurate and robust classification of EEG signals recorded under different mental states. Common spatial patterns (CSP) is one of the most widely used spatial filtering approaches which can capture the event related synchronization (ERS)/desynchronization (ERD) information and extract discriminative patterns during motor imaginary tasks. However, CSP is prone to be affected by the appearance of noise and outliers in EEG signals because of the L2-norm based divergence expression. Besides, CSP is based on the dynamic behavior of single signals without using the coupling information between two EEG signals.This paper proposed two rubost CSP algorithms namely waveform length regularized L1-norm based common spatial patterns (wlCSPL1) and sparse L1-norm based common spatial patterns (sp-CSPL1) respectively based on L1-norm theory. Besides, the study proposed a statistics based phase synchronization measurement method namely phase lag weighted by signed-rank (PLSR) and obtained coupled features of CSP and PLSR by three coupling strategies. In order to show the superiority of these proposed algorithms, experiments in one simulated data set and five true recorded EEG data sets of motor imaginary tasks were executed. The experimental results show that, wlCSPL1 extracted more anti-noise spatial filters than CSP did, which improved classification accuracies of 17 subjects. As for sp-CSPL1 algorithm, which is expected to obtain more sparse spatial filters by adding a L1-norm regularization term, was proved to be effective for channel reduction. As for PLSR, the average result on nonstationary EEG data set shows better performenc of PLSR than the other algorithms, on the other hand, classification performances of the three feature coupling strategies were confirmed, especially, the dimensionality reduction coupling strategy obtained higher accuracies then CSP for all subjects.