Extracting　features　from　EEG　signals　through　time,　frequency　and　spatial-domain　gives　rise　to　the　problem　of　neglecting　the　property　of　nonlinear　manifold　structure　of　the　data,　and　traditional　feature　extractions　based　on　Euclidean　space　are　unfavorable　for　the　full　utilization　of　the　manifold　feature.　In　this　paper,　a　new　EEG　classification　method　is　proposed　by　representing　EEG　signals　on　the　Grassmann　manifold.　Firstly,　a　low　rank　representation　method　based　on　matrix　recovery　is　introduced　into　the　Grassmann　manifold　to　explore　the　subspace　information　and　re-express　the　EEG　signals.　Then　a　deep　neural　network　is　proposed　to　reduce　the　dimension　of　the　recovered　EEG　data,　and　the　low　rank　representation　makes　the　EEG　data　separable　by　using　T-SNE　visualization;　Finally,　the　dimension　reduced　features　are　input　to　the　traditional　classifier　for　classification.　The　proposed　method　is　performed　on　four　datasets　for　seizure　detection,　disease　recognition　and　mental　arithmetic　recognition　respectively.　The　corresponding　accuracy　reaches　99.99%,　99.61%,　100%,　and　99.72%　which　shows　that　the　proposed　method　outperforms　other　state　of　art　methods　on　EEG　classification.　©　2023　Elsevier　Ltd