One of the primary cause of malfunction generated by machinery is bearing damage, and its vibration signal has the characteristics of weak, non-stationary and nonlinear. It proposes multi-scale sub-band sample entropy concept aiming at the problem of eigenvalues and eigenvectors that can’t be accurately extracted from the weak signals, firstly, obtaining the multi-scale signal by wavelet packet decomposed, and then, sub-band decomposing each signal of the scales, finally, solving the sample entropy of each sub-band, this method can dig deep into the essential characteristics of the weak signal. It presents that adopting smooth pseudo Wigner-Ville to solve the problem of uneven distribution of energy density of non-stationary signal, it can be used to deal with the time and frequency aggregation of the instantaneous symmetric correlation function of non-stationary signal, make the signal energy is evenly distributed. It proposes that using LPP (Locality Preserving Projection) decomposition to settle the problem of can’t accurate image of the mainstream of the nonlinear data, LPP in the process of projection to keep the relationship between the optimal data of the local neighborhood, the main manifold can be accurately excavated from nonlinear data. The paper adopting a group of normal, inner ring fault, balls fault and outer ring fault signal as the original data to verify this method’s validity, he experimental results prove that this method can effectively to separation and identification of signal failure.