The　working　conditions　of　wind　turbine　rolling　bearings　are　always　complex,　therefore　obtained　fault　vibration　signals　are　non-stationary　and　non-linear.　However,　the　traditional　method　based　on　time-frequency　domain　has　some　problems,　such　as　inaccuracy　and　poor　adaptability　when　extracting　fault　features.　To　solve　the　problem,　a　novel　feature　extraction　algorithm　for　fault　diagnosis　is　proposed　based　on　Local　Mean　Decomposition(LMD)　and　Morphological　Fractal　Dimension(MFD),　and　combined　with　Extreme　Learning　Machine　(ELM)　to　conduct　wind　turbine　bearing　fault　diagnosis.　This　method　considers　the　different　damage　degree　and　different　fault　types　of　rolling　bearing　at　the　same　time.　Firstly,　the　raw　vibration　signal　is　adaptively　decomposed　by　LMD　into　several　Product　Functions　(PFs)　which　in　different　frequencies.　Secondly,　the　correlation　coefficients　between　all　PFs　and　the　raw　signal　are　calculated,　and　the　first　three　PF　components　with　the　maximum　correlation　coefficient　value　are　selected　as　sensitive　variables.　The　fractal　dimension　of　selected　PFs　is　estimated　by　morphology　to　construct　fault　feature　vector.　It　is　taken　as　the　input　of　ELM　to　develop　fault　diagnosis　model.　Finally,　the　experimental　results　show　that　the　proposed　method　improves　performance　for　detecting　the　bearing　faults.　The　method　has　also　high　computational　efficiency　and　accuracy.　©　2020,　Solar　Energy　Periodical　Office　Co.,　Ltd.　All　right　reserved.