The vibration signal of fault bearing is always non-stationary and multi-component. The strong background noise and low signal–noise-ratio make it more difficult to extract the fault features from the raw signal accurately. In this paper, based on the enhanced synchrosqueezing wavelet transform (SWT) which couples with the adaptive optimal cumulative frequency range, a new feature extracting method is proposed. The adaptive optimal cumulative frequency range is deduced according to the relationship between wavelet coefficient of vibration signal and the supporting interval of wavelet basis. The optimal parameters of wavelet and frequency range are calculated by the minimum-energy error criterion that is based on the integrity and orthogonality of the intrinsic mode types function. In the proposed method, first, the raw signal is denoised via a data-driven wavelet soft-threshold denoising method. Then, the preprocessed signal is decomposed into a series of intrinsic mode type (IMT) functions by the advanced SWT method, and the spectral kurtosis is used to select the sensitive IMT. Finally, the signal is reconstructed by the selected IMT and the ridge feature. The instantaneous frequency feature of sensitive IMT functions is analyzed by Hilbert transform. The proposed method is verified by two different experimental verifications, the results show that the proposed method could classify the fault type of bearing accurately.