Noise robustness analysis of sparse representation based classification method for non-stationary EEG signal classification

Brain-computer interface (BCI) systems provide a new communication and control channel between people and external devices [1]. In these systems, users can control an external device using their intention or imagination without making any real muscle movement. Therefore, these systems are very helpful for people who are suffering from severe motor diseases. The electroencephalogram (EEG) is widely used for measuring brain signals in BCI systems because of its low cost, no space restriction, and high temporal resolution compared with other equipment such as functional magnetic resonance imaging (fMRI) and magneto encephalogram (MEG) [2,3]. However, scalprecorded EEG signals are very sensitive to noise. In particular, in the case of motor imagery based BCI, which uses induced EEG signals while the subject imagines limb movements [2,3], the instability of imagery task, non-stationarity of signals, and lack of concentration are among main obstacles to effectively process the EEG signals. In addition, it is difficult to collect a large set of training samples because of the subject’s fatigue. The raw EEG signals are associated with high dimension owing to the large number of EEG channels; hence, it is difficult to collect volume of data samples that are large enough for good training. Therefore, EEG signal processing is very important and many research efforts have been focused on this issue [5–7].