Abstract: The vibration signals of bearings are usually typical fault samples in reality. Therefore we propose a tion (LMD) approximate entropy and the support vector non-stationary and it is difficult to obtain a large number of fault diagnosis method that uses the local mean decomposi-machine (SVM). We decompose the original non-stationa-ry acceleration vibration signals into several stationary production functions (PFs). Because the approximate entro-py values of vibration signals in different frequency bands change when faults occur in bearings, we can judge whether a fault occurs and decide its type by calculating the approximate entropy and can establish the SVM by using as its input the approximate entropy features extracted from the PFs that contain the major information on faults. Experimental results show that our diagnosis method can effectively diagnose faults of bearings.