In view of the complexity of degraded data of rolling bearings and the problem that traditional life prediction methods can not make full use of the correlation of data, thus the prediction accuracy is not high, a residual life prediction model combining deep belief neural (DBN) network and long-short term memory (LSTM) neural network was proposed. In this model, band-pass filter was first used to denoise the vibration data of rolling bearings, and then the prediction starting point of the model was determined according to the trend diagrams of root mean square and kurtosis features in the whole life cycle of the bearing; Secondly, the optimized 4-layer DBN network was used to extract depth features and be used in LSTM training and testing. The reliability of the proposed model is proved through the bearing life cycle test, and compared with the prediction result of traditional LSTM、BP neural network and DBN-BP, the validity of the model in this paper is verified.