Recently,　it　is　the　key　of　brain-computer　interface　(BCI)　technology　to　extract　electroencephalogram　(EEG)　data　features　effectively　and　classify　them　accurately.　As　a　challenging　research　topic　in　the　field　of　BCI,　the　decoding　technology　of　MI　classification　based　on　EEG　can　provide　a　reliable　and　convenient　way　of　information　interaction　for　patients　with　spinal　cord　injury,　disability　and　stroke.　However,　the　recorded　EEG　signal　is　easily　interfered　by　other　signals,　which　leads　to　its　low　signal-to-noise　ratio.　This　paper　proposes　a　model　based　on　deep　learning　network　to　decode　EEG-based　MI　actions,　combining　deep　separation　convolution　network　(DSCNN)　and　bidirectional　long　short-term　memory　(BLSTM)　neural　network.　Firstly,　DSCNN　with　two　layers　of　logical　convolution　is　used　to　extract　the　spatio-temporal　feature　information　of　EEG-based　MI　tasks,　and　then　the　spatial　feature　information　is　further　extracted　through　the　ordinary　convolution　network,　which　is　connected　into　a　one-dimensional　vector　by　one　layer,　and　then　transmitted　to　BLSTM　to　further　extract　the　temporal　feature　information.　Finally,　the　EEG-based　MI　task　is　decoded　by　the　softmax　function.　A　prominent　decoding　rate　is　obtained　to　evaluate　the　performance　of　the　model　with　the　open　physiological　EEGMMIDB　datasets,　which　is　superior　to　other　advanced　models.　The　research　of　EEG-based　MI　algorithm　model　proposed　can　effectively　promote　the　application　of　brain-computer　interface　technology　in　medical　field.　©　2022,　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.