Bearing fault signal extraction is a crucial procedure when acquired signals have strong background noise. This paper proposes an adaptive fault signal extraction method based on CEEMDAN and relative entropy difference spectrum. Firstly, the acquired signal is decomposed into several intrinsic mode functions (IMFs) by complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), each IMF contains a part of frequency band information. The relative entropy difference spectrum is obtained based on the calculation of the IMF relative entropy and the difference of relative entropy. According to the designed adaptive fault signal extraction criterion, it can adaptively choose the IMFs which carries on sufficient fault information. Reconstructing the time signal with the chosen IMFs and calculating the reconstructed signal envelope spectrum, it can accurately diagnose the bearing fault. The effectiveness of the proposed methods are validated by the simulation signal and engineering signal. The results show that the reconstructed signal envelope spectrum can accurately find out the fault characteristic frequency of bearing fault, which verified that the proposed method has the ability to adaptive fault signal extraction and diagnose bearing fault.