The strong coupling between the bearing and other internal components of the equipment leads to nonlinear relationship between vibration signal and equipment state. Moreover, the single signal feature is difficult to describe the state of the equipment comprehensively, while multifeatures contain more status information, the signal redundancy caused by highdimensional features easily declines the classification accuracy of the model. Therefore, a rolling bearings fault diagnosis method based on generalized morphological difference filter (GDIF) and autoencoder network (AN) is proposed. This method uses the GDIF to reduce the noise of vibration signals, and obtains the lowdimensional intrinsic manifold from the highdimensional features of the signal by the max likelihood estimate (MLE) and AN algorithm, which alleviates the dimension disasters of highdimensional features. Finally, the extreme learning machine (ELM) fault diagnosis model is established to identify the bearing fault types. The experiments show that the method can effectively suppress the noise; and the classification accuracy of the ELM Model can reach 98.04% after dimension reduction of features by AN.