As　a　nonlinear　time-varying　and　non-stationary　signal,　Motor　Imagery　Electroencephalography　(MI-EEG)　has　attracted　many　researchers　to　use　its　time-frequency　feature　extraction　by　using　Discrete　Wavelet　Transform　(DWT)　in　brain　computer　interfaces　(BCIs).　Though　a　few　people　have　devoted　their　efforts　to　exploring　its　nonlinear　nature　from　the　perspective　of　manifold　learning,　they　hardly　take　into　full　account　both　time-frequency　feature　and　nonlinear　nature.　To　obtain　features　that　can　fully　describe　the　information　from　a　nonlinear　nature　and　time-frequency　perspective　of　MI-EEG,　a　novel　feature　extraction　method　is　proposed　based　on　the　Locally　Linear　Embedding　algorithm　(LLE)　and　DWT.　The　multi-scale　multi-resolution　analysis　is　implemented　for　MI-EEG　with　DWT,　and　the　valid　time　and　frequency　windows　are　determined　in　advance　by　a　Wigner-Ville　distribution.　In　view　of　the　nonlinear　structure　in　MI-EEG,　LLE　is　applied　to　the　approximation　components　to　obtain　the　nonlinear　features,　and　the　statistics　of　the　detail　components　are　calculated　to　obtain　the　time-frequency　features.　After　an　organic　combination　of　the　two　features,　a　Back-Propagation　neural　network　optimized　by　a　Genetic　Algorithm　was　employed　as　a　classifier　to　evaluate　the　effectiveness　of　the　proposed　feature　extraction　method.　Compared　with　conventional　DWT-based　methods,　the　proposed　method　has　a　better　effect　on　feature　visualization　with　an　obvious　clustering　distribution　and　improves　the　classification　results　and　their　stability.　This　paper　successfully　achieves　manifold　learning　in　signal　processing　of　EEG.