A signal de-noising method combining maximum between-cluster variance-empirical wavelet transform (MBCV-EWT) with singular value difference spectrum is proposed. It helps rolling bearing to overcome pattern aliasing and ensure the integrity and independence of each frequency component during vibration signal de-noising. First, in light of the uncertainty of traditional interval partitioning, an MBCV-EWT signal decomposition method is proposed. The signal spectrum is adaptively divided by the maximum inter-class variance, and a band-pass filter is constructed on each partition interval. Then, aiming at the redundancy of AM-FM component, impulse index is proposed to be the screening criteria of AM-FM and the best component is selected as a follow-up target. Finally, singular value decomposition is used for AM-FM. Signal de-noising is achieved according to the singular value difference spectrum. The simulation and experimental results show that the proposed method can achieve adaptive spectrum division. The problem of pattern aliasing can be effectively overcome, and the main components of the components obtained by decomposition are independent and complete. The amplitude modulation frequency components are accurately screened, and the effect of de-noising is obvious, so as to lay the foundation for fault recognition and prediction.