Bearings　play　an　important　role　in　the　industrial　system.　Among　all　kinds　of　diagnosis　methods,　empirical　wavelet　transform　has　been　widely　used　for　its　characteristic　of　separating　empirical　modes　from　the　spectrum.　Although　the　empirical　wavelet　transform　restrains　the　mode　aliasing　of　extracting　modal　functions　from　the　time　domain,　it　runs　slowly　and　separates　a　large　number　of　invalid　components.　In　this　paper,　an　adaptive　and　fast　empirical　wavelet　transform　method　is　proposed.　The　method　extracts　the　first　feature　cluster　in　the　Fourier　transform　of　the　spectrum　to　reconstruct　the　trend　spectrum.　The　minimum　points　are　regarded　as　the　initial　boundaries.　The　key　spectral　negentropy　is　proposed　to　extract　the　frequency　band　which　may　contain　main　information.　This　method　reduces　the　number　of　invalid　components　and　computation　time　by　filtering　components　before　reconstruction.　The　simulation　signal　proves　that　the　proposed　method　is　effective　and　the　proposed　key　spectral　negentropy　has　stronger　anti-noise　ability　than　kurtosis.　The　experimental　signals　show　that　the　method　can　successfully　extract　the　fault　features　of　inner　or　outer　rings　of　bearings,　and　is　suitable　for　the　diagnosis　of　composite　faults.