An algorithm based on the similarity of dynamic time warping (DTW) and Bi-directional long short-term memory (Bi-〖JP〗LSTM) neural network was proposed to solve the problems such as the existed degradation indexes were poorly monotonic and insensitive to bearing anomaly, which would make it difficult for data-driven deep learning algorithms to achieve accurate prediction of bearing life. The information entropy was used to extract the degraded information of rolling bearing, and a continuous time series was constructed; reference templates and test templates were constructed for similarity calculation through time series division, the DTW algorithm was used to calculate the similarity between templates, which was used as a health index to describe the degree of bearing degradation; the Bi-LSTM network was trained with health index and predicted the lifetime of the bearings.The research was carried out by using the full life data of the French IEEE PHM 2012. The result shows that the entropy index optimized by DTW algorithm can better reflect the degradation process of rolling bearing; compared with entropy index before optimization, the optimized index can significantly improve the accuracy of bearing life prediction when the prediction algorithm is kept unchanged during life prediction of rolling bearing.