Empirical　wavelet　transform　(EWT)　is　a　new　adaptive　signal　decomposition　method　based　on　wavelet　theory,　the　main　idea　is　to　establish　an　appropriate　set　of　empirical　wavelet　filter　banks　for　adaptive　signal　decomposition.　EWT　has　been　demonstrated　its　effectiveness　in　some　applications.　However,　the　unreasonable　spectrum　segmentation　will　lead　to　the　emergence　of　many　invalid　components.　In　this　paper,　a　novel　spectral　segmentation　method　is　proposed　to　improve　the　drawback　of　EWT　in　boundary　division.　The　proposed　method　takes　into　account　the　waveform　of　the　spectrum　itself.　First,　different　spectral　trends　are　obtained　by　iteratively　calculating　the　mean　of　the　upper　envelope　function　and　the　lower　envelope　function　of　the　spectrum.　Then,　the　most　appropriate　one　is　got　according　to　the　criterion　and　the　spectrum　segmentation　is　achieved　by　detecting　the　local　minimum　of　the　trend.　Finally,　empirical　modes　are　obtained　by　a　set　of　bandpass　filters.　The　effectiveness　and　efficiency　of　the　method　are　verified　by　two　simulation　signals.　Finally,　the　proposed　method　is　applied　to　fault　diagnosis　of　the　inner　and　outer　race　of　rolling　bearings,　respectively.　The　results　indicate　that　the　method　can　accurately　and　effectively　identify　fault　information.