Aiming at fault diagnosis problems of rolling bearings, a hybrid intelligent diagnosis model is proposed based on stacked sparse auto encoders (SSAE), improved gray wolf optimization algorithm (IGWO) and support vector machine (SVM). Firstly, by making use of the excellent ability of SSAE in feature self-extraction, adaptive learning of deep frequency-domain features of fault signals can be realized. In addition, sparse penalty term is introduced to enhance the generalization. Secondly, the high-level feature vectors are taken as input to the SVM for classification and recognition, whose parameters are optimized by the IGWO algorithm. The proposed model fully combines the powerful feature self-learning ability of deep neural network and the excellent performance of SVM in classifications on small samples. Identification on vibration signals of different fault types can be achieved in a more reliable and accurate way, avoiding the drawbacks of manual feature extraction. Moreover, contrast experiments are conducted for validation. The results show that the model proposed in this paper has better performance in fault diagnosis accuracy compared with traditional methods, and the diagnosis accuracy can be over 98%.