Variational mode extraction (VME), inspired by variational mode decomposition (VMD), is a novel fault diagnosis technique that can efficiently extract narrowband modes from multi-component signals. Compared with VMD, VME is more accurate and faster when extracting the narrowband component. However, the preset center frequency ω
c
and balance factor α seriously affect the performance of VME. Therefore, a spectral-coherence guided variational mode extraction (SCVME), capable of determining the hyper-parameters automatically, is proposed for fault diagnosis of rolling bearings. First, considering the advantages of spectral coherence (SCoh) for characterizing the cyclostationarity of bearing faults, its energy spectrum is further constructed. The energy spectrum of SCoh can intuitively reveal the fault information energy hidden in each frequency, which provides sufficient support for the determination of the center frequency ω
c
. Subsequently, a novel signal evaluation index named cyclic pulse intensity (CPI) is proposed to adaptively optimize the balance factor α. It has been verified that the proposed CPI index is superior to these common metrics, including kurtosis, spectral kurtosis (SK) and l2/l1 norm, in identifying periodic pulses. Finally, the modes containing fault information are accurately extracted by VME according to the optimal parameters (ω
c
, α). The effectiveness of the proposed method has been demonstrated by the simulations and experiments analysis. In addition, comparisons with the VMD and Autogram methods are carried out to highlight the superiority of the SCVME method.