In　the　health　monitoring　of　rotating　machinery,　there　often　coexists　multiple　fault　sources;　thus　a　multi-source　composite　fault　signal　will　be　collected　by　sensors.　Moreover,　the　vibration　signal　of　rotating　machinery　is　usually　submerged　by　the　environmental　noise　in　practical　engineering　occasions.　These　increase　the　difficulty　of　separation　from　the　composite　fault　signal.　However,　the　effect　of　traditional　wavelet　analysis　method　for　denoising　to　enhance　features　of　fault　signal　is　limited,　because　of　the　energy　dispersion　of　fault　signals.　To　overcome　these　problems,　we　propose　a　convolutive　blind　separation　method　based　on　peak　transform　using　wavelet　analysis.　With　this　strategy,　the　original　signal　energy　in　high　frequency　is　concentrated　into　that　in　low　frequency　band,　through　some　non-linear　transformations.　Next,　the　large　wavelet　coefficients,　which　represent　the　main　features　of　the　signal,　are　retained　to　reconstruct　a　feature-enhanced　signal.　The　convolutive　fixed-point　algorithm　based　on　maximization　of　non-Gaussianity,　is　carried　out　to　separate　source　signals　from　the　feature-enhanced　signal.　Experimental　results　of　rolling　bearing　show　that　the　proposed　strategy　can　effectively　separate　source　fault　signals　from　the　composite　signal　and　detect　the　fault　features.　©　2018　IEEE.