As it’s difficult to get comprehensive fault information with traditional machine model in the interrelated process of fault knowledge in rotating machinery fault diagnosis，an immune genetic algorithm (IGA) is proposed to optimize Elman neural network. Fault vibration signals are decomposed into several stationary intrinsic mode functions (IMF) first，then the instantaneous amplitude energy of the IMF which has the fault characteristics are computed and regarded as the input characteristic vector of the Elman neural network optimized by IGA algorithm for fault classification. EMD decomposition adaptively isolates fault vibration signals from original signals. IGA algorithm has more superior performance on global optimization and convergence speed. So it can improve the fault diagnosis accuracy and the adaptive dynamic memory of the Elman neural network. The result of rolling bearings fault simulation experiments show that，compared with traditional fault diagnosis model，the proposed method significantly improves the diagnostic accuracy and generalization ability of the typical failure of the rolling bearings.