A　denoising　method　of　the　improved　neighboring　coefficients　and　the　undecimated　multiwavelet　transform　is　proposed,　Hilbert-Huang　time-frequency　analysis　is　applied　as　the　post-processing　method.　The　proposed　method　is　applied　to　the　incipient　fault　diagnosis　of　planetary　gearboxes.　In　the　planetary　gearbox,　the　fault　response　is　quite　weak;　the　vibration　is　obviously　non-stationary　and　evidently　nonlinear;　low-frequency　characteristics　are　easily　immersed　in　heavy　noise.　Therefore,　the　existing　fault　diagnosis　theory　and　technology　for　traditional　fixed-shaft　gearboxes　fail　to　solve　the　difficulty　in　the　planetary　gearbox　fault　diagnosis.　The　undecimated　multiwavelet　transform　has　the　shift-invariant　property　in　time　domain,　which　can　effectively　weaken　the　Gibbs　phenomena　in　the　neighborhood　of　the　discontinuities.　The　improved　neighboring　coefficients　can　select　variant　sizes　of　neighboring　window　and　flexible　thresholds　at　different　decomposition　levels,　which　can　correctly　extract　the　incipient　fault　features　in　the　non-stationary　signals.　Hilbert-Huang　time-frequency　analysis　can　intuitively　represent　the　non-stationary　and　nonlinear　features　of　the　collected　signals.　Experiments　indicate　that　the　proposed　method　can　correctly　extract　the　weak　fault　features　caused　by　the　incipient　pitting　defects　in　the　planetary　gearbox.　©　2013　Journal　of　Mechanical　Engineering.