The　impulse　feature　from　an　early　diagnosis　of　bearing　fault　is　often　drowned　by　the　noise　background,　and　is　usually　very　difficult　to　extract.　To　solve　the　problem,　a　new　method　was　presented　here,　which　was　based　on　the　Local　Mean　Decomposition　(LMD)　and　wavelet　de-noising.　The　LMD　was　used　to　decompose　the　original　signal　of　the　bearing　into　serval　PF　components　which　were　retained　using　the　principle　of　maximum　kurtosis　and　cross-correlation　coefficients　to　keep　only　the　reasonable　ones.　Compared　to　the　traditional　PF　component　selection　process,　our　new　method　captured　more　fault　impulse　features　in　the　selected　PF　components.　For　these　retained　PF　components　were　first　de-noised　by　a　db10　wavelet　of　5　layers,　and　then　were　used　to　reconstruct　the　high　frequency　signal　of　each　component　layer　by　the　method　of　superposition.　Finally,　the　envelope　spectrum　analysis　was　applied　to　the　derived　the　spectral　kurtosis　to　give　the　result　of　rolling　bearing　fault　diagnosis.　A　test　experiment　was　conducted　with　our　bearing　fault　simulation　platform.　The　collected　data,　including　signal　from　bearing　outer　ring,　inner　ring　and　ball,　was　analyzed　using　the　method　proposed　in　this　paper.　The　result　shown　that　the　new　method　can　effectively　enhance　the　impulse　features　in　the　signal,　also　improve　the　fault　diagnosis　efficiency.　©　2017　IEEE.