Abstract: It often occurs to common mechanical equipment,which is easily damaged by high frequency,that fault diagnosis is time-consuming and inaccurate. This paper proposes an improved fault diagnosis model with optimized Elman neural network supported by the bearing data from Case Western Reserve University（ CWRU）. Given that the common fault vibration signal of rolling bearing is nonlinear and nonstationary,we first apply Hilbert Huang Transform（ HHT）to the original signal data process in forms of decomposition,dimensionally reduction and transformation so as to gain the characteristic quantity of the fault by extracting the spectral envelope sensitive to fault signal and able to distinguish the fault frequency. Then,by improving the Elman neural network,the time is delayed at the corresponding position of the network input,and its length is adjustable according to the input feature dimension. Finally,we build a fault diagnosis model. The diagnostic results demonstrate clearly and intuitively that the model has a very good improvement in terms of diagnostic accuracy and learning efficiency.