With　the　popularization　of　brain-computer　interface　(BCI)　technology,　the　research　on　intention　recognition　of　motor　imagery　(MI)　in　electroencephalogram　(EEG)　has　turned　to　the　latest　research　points.　The　BCI-based　system　can　provide　a　powerful　rehabilitation　guarantee　for　patients　with　stroke　and　spinal　cord　injury.　However,　EEG　signals　have　certain　complexity,　which　is　easily　interfered　by　other　noises,　resulting　that　it　is　still　not　enough　to　provide　some　better　practical　application　scenarios.　In　this　paper,　an　improved　framework　has　been　proposed　through　deep　separation　convolution　neural　network　(DSCNN)　and　extreme　learning　machine　(ELM)　to　address　the　recognition　rate　of　patients＇　motor　intention.　First　of　all,　the　collected　one-dimensional　time　series　EEG　signals　are　preprocessed　into　a　two-dimensional　grid　containing　spatial　and　temporal　features.　Afterwards,　the　DSCNN　is　utilized　to　extract　the　temporal　features　and　spatial　features,　respectively.　Thirdly,　the　ELM　classifier　is　utilized　to　classify　five　kinds　of　MI　actions　according　to　the　extracted　temporal　and　spatial　features.　Experimental　results　indicate　that　the　presented　framework　can　achieve　an　excellent　intention　recognition　rate　of　97.88%　through　the　public　EEGMMIDB　datasets.　Moreover,　the　training　time　was　greatly　shortened　from　13h30min　to　9h10min　with　a　reduction　rate　of　about　32%　under　the　same　hardware　configuration,　which　is　superior　to　some　advanced　models.　Therefore,　the　proposed　idea　not　only　accelerates　the　training　speed　of　the　model,　but　also　can　boost　the　application　research　of　BCI　based　rehabilitation　efficiently.