In order to extract the effective rolling bearing features and eliminate the redundancy between features,a deep neural network (DNN) based on stacked sparse autoencoder (SSAE) and Softmax layer was proposed for bearing fault diagnosis. 12 statistical features and 6 timefrequency domain features were extracted from the vibration signal,then the obtained features were used to construct an 18-dimensional feature set. The highdimensional feature set obtained the highlevel feature without redundancy by greedy layerwise learning by stacked sparse autoencoder. Finally,the advanced features were input into the Softmax classification layer for bearing fault diagnosis. The experimental results show that compared with traditional BP and SVM classifiers,DNN can be used to identify the type of rolling bearing fault more accurately.