To effectively extract useful information from nonlinear and non-stationary vibration signals of rolling bearings, a method is presented by combining an optimized morphological filter algorithm with Elman Neural Network. First, some product function (PF) components from vibration signals of the rolling bearing are obtained by local mean decomposition (LMD). Then, the PF component with the biggest kurtosis value is selected according to the kurtosis value criterion, and it is demodulated by adaptive morphology filter with multi-size and structure element. Thus, the energy feature vector is extracted as the input parameters of the Elman Neural Network. At last, the fault status and type of the rolling bearing are confirmed. The simulation analysis and experimental study show that this method could effectively extract fault features of the rolling bearing， more pronounced compared with the traditional method of high frequency resonance demodulation. In addition, compared with Wavelet Packet Analysis BP Neural Network Fault Diagnosis, this method is more feasible and effective with a higher recognition rate.