Due to the inconsistent distribution of bearing degradation data under different working conditions, the prediction performance of remaining useful life by using deep learning and other techniques is limited. Moreover, most of existing transfer learning-based prediction methods fail to fully exploit the inherent degradation trend under different working conditions. To solve these problems, a prediction method of bearing remaining useful life is proposed based on deep temporal feature transfer. Firstly, a health indicator(HI) construction model based on deep temporal features is proposed. The model uses the temporal convolutional network(TCN) to exploit inherent temporal features from the degradation trend of multiple bearings, and builds the HI sequence of the bearings in source domain. Then, a domain adaptation algorithm for minimizing sequence similarity is proposed. The HI sequence of source domain is used as meta-information of degradation trend for selecting common representative features between target domain and source domain. Finally, the support vector machine is used to construct the prediction model. Experiments are performed on the IEEE PHM Challenge 2012 bearing whole-life dataset, and the results show the proposed HI can effectively reflect the degradation trend, and significantly improve the prediction accuracy of remaining life.