Background:　Electroencephalogram　(EEG)　provides　a　non-invasive　approach　to　measure　the　electrical　activities　of　brain　neurons　and　has　long　been　employed　for　the　development　of　brain-computer　interface　(BCI).　For　this　purpose,　various　patterns/features　of　EEG　data　need　to　be　extracted　and　associated　with　specific　events　like　cue-paced　motor　imagery.　However,　this　is　a　challenging　task　since　EEG　data　are　usually　non-stationary　time　series　with　a　low　signal-to-noise　ratio.　New　method:　In　this　study,　we　propose　a　novel　method,　called　structure　constrained　semi-nonnegative　matrix　factorization　(SCS-NMF),　to　extract　the　key　patterns　of　EEG　data　in　time　domain　by　imposing　the　mean　envelopes　of　event-related　potentials　(ERPs)　as　constraints　on　the　semi-NMF　procedure.　The　proposed　method　is　applicable　to　general　EEG　time　series,　and　the　extracted　temporal　features　by　SCS-NMF　can　also　be　combined　with　other　features　in　frequency　domain　to　improve　the　performance　of　motor　imagery　classification.　Results:　Real　data　experiments　have　been　performed　using　the　SCS-NMF　approach　for　motor　imagery　classification,　and　the　results　clearly　suggest　the　superiority　of　the　proposed　method.　Comparison　with　existing　methods:　Comparison　experiments　have　also　been　conducted.　The　compared　methods　include　ICA,　PCA,　Semi-NMF,　Wavelets,　EMD　and　CSP,　which　further　verified　the　effectivity　of　SCS-NMF.　Conclusions:　The　SCS-NMF　method　could　obtain　better　or　competitive　performance　over　the　state　of　the　art　methods,　which　provides　a　novel　solution　for　brain　pattern　analysis　from　the　perspective　of　structure　constraint.　(C)　2015　Elsevier　Ltd.　All　rights　reserved.