Robust feature construction against non-stationarity for EEG brain-machine interface

Electroencephalographic (EEG) signals are known to be non-stationary and easily affected by artifacts. therefore, it is particularly important to alleviate non-stationarity in order to construct useful brain-machine interface (BMI) systems in real environments. in this manuscript, we show analysis of nonstationarity in the motor imagery data recorded with a portable EEG device over 15 days and 83 runs. we found non-stationary changes possibly caused by drowsiness, bad hardware conditions, and switching what to imagine for movements of specified limbs. matrix PCA was used for exploratory analysis and data visualization. then, we explain some of robust feature construction techniques such as stationary common spatial patterns (SCSP) and maxmin CSP which will be applied to our data.