The　local　mean　decomposition　(LMD)　method　can　adaptively　decompose　a　vibration　acceleration　signal　into　a　series　of　product　functions　with　different　frequency　characteristics.　However,　in　the　process　of　calculating　local　mean　and　interpolation,　this　method　not　only　increases　the　computational　complexity,　but　also　produces　serious　modal　aliasing.　In　order　to　improve　the　computational　speed　and　reduce　modal　aliasing,　a　new　method　is　proposed　to　optimize　local　mean　decomposition.　The　proposed　method　uses　an　order　statistics　filter　to　estimate　the　upper　and　lower　envelopes　of　the　signal.　The　moving　average　method　is　replaced　by　the　envelope　mean　method　to　calculate　the　local　mean　function　and　envelope　function.　Subsequently,　spectral　negentropy　is　used　as　the　stopping　criterion　of　iteration　and　to　filter　the　component　which　may　contain　fault　information.　Strict　termination　conditions　in　LMD　are　replaced.　The　invalid　components　avoid　decomposition　and　the　computational　efficiency　is　improved.　The　dependence　on　extreme　points　is　reduced,　so　the　endpoint　effect　is　suppressed.　The　experimental　results　show　that　the　proposed　method　is　capable　of　identifying　bearing　faults　and　is　suitable　for　the　diagnosis　of　composite　faults　of　rolling　bearings.