The　adaptivity　of　feature　extraction　is　a　key　problem　in　rehabilitation　with　brain　computer　interface.　A　multi-domain　feature　fusion　method　was　proposed　for　EEG.　The　method　is　mainly　based　on　Hilbert-Huang　transform　(HHT)　and　common　spatial　subspace　decomposition　(CSSD)　algorithm　and　denoted　as　HCSSD.　Firstly,　a　relative　distance　criterion　is　defined　to　select　the　optimal　combination　of　channels　in　consideration　of　the　distinction　of　event-related　desynchronization　(ERD)　extent　induced　by　different　motor　imagery　tasks.　Then　HHT　and　CSSD　are　applied　to　extract　the　time-frequency　feature　and　spatial　feature　for　optimal　EEG　signals　respectively.　Furthermore,　serial　feature　fusion　strategy　is　employed　to　construct　time-frequency-spatial　feature.　Finally,　learning　vector　quantization　(LVQ)　neural　network　is　designed　to　classify　the　motor　imagery　electrocorticography　(ECoG)　data　in　BCI　Competition　III.　The　data　were　recorded　from　the　same　subject　and　with　the　same　mental　tasks,　but　on　two　days　with　about　one　week　in　between.　The　average　recognition　accuracy　is　92%　with　much　less　channels　used.　Experiment　results　show　that　HCSSD　can　enhance　the　adaptability　and　robustness　of　feature　extraction,　and　the　recognition　accuracy　is　also　improved.　This　is　helpful　for　further　research　of　portable　BCI　system　in　rehabilitation　field.