The working conditions of mechanic system in real life are generally studied by analysis of signals so that the exact conclusions will be drawn. These signals emanating from mechanic system commonly contain a mixture of different oscillations. For a reliable conclusion, it is necessary to separate a set of physically meaningful modes from the mixture and background noise. Based on that, a new method for bearing fault extraction is proposed in this paper. At first, a novel decomposition algorithm named empirical wavelet transform (EWT) is employed to decompose the fault signal into a set of AMFM components that have a compact support Fourier spectrum. And then, KL divergence method is used to select the sensitive component. Finally, the fault characteristic frequency is extracted by a new demodulation method called energy operator of symmetrical differencing (DEO3S) that can restrain the end effect, and the instantaneous frequency is obtained at the same time. The results of the simulation and bearing fault diagnosis experiments indicate that the method can effectively extract fault characteristic frequency, certifying its feasibility and superiority in comparison with the previous methods.