Due　to　strong　noise　in　vibration　signals　of　rolling　mills,　it　is　very　difficult　to　identify　their　fault　location　and　to　extract　their　feature.　For　incipient　faults,　the　spectrum　analysis　is　not　effective.　Using　the　band　pass　filtering　feature　of　wavelet　transformation,　the　signal　can　be　decomposed　in　the　specific　frequency　band.　Furthermore,　the　single　level　reconstruction　of　the　decomposed　signals　can　separate　the　valuable　components　from　the　noise　successfully.　At　the　same　time,　the　global　reconstruction　after　processing　the　high　frequency　decomposing　coefficients　can　realize　the　same　goal.　The　signal　processing　results　for　the　vibration　signal　of　bearings　of　a　rolling　mill　after　using　several　kinds　of　denoising　technology　show　that　the　low　frequency　information　containing　fault　features　is　extracted,　and　the　characteristic　frequency　of　incipient　faults　can　be　captured　from　the　frequency　spectrum　of　the　denoised　signal.　The　proposed　method　can　improve　accuracy　of　fault　diagnosis　decomposing.