Remaining useful life (RUL) prediction is one of the difficulties in prognostics and system health management (PHM) of bearings. Moreover, due to the limitations of test cost and equipment, it is not easy to obtain abundant full life cycle experimental samples of bearings, which cannot provide enough training samples to further improve the accuracy of RUL prediction. To address these problems, this article proposes a rolling bearing RUL prediction method based on coupled diffusion process and temporal attention (CDTA) mechanism. This method first augments the original sequence with the coupled diffusion process, which preserves its randomness and reduces the difficulty of training the subsequent inference network. Then, it introduces a temporal attention unit (TAU) to enhance the network’s ability to extract and fuse temporal features and dependencies within and between samples, which improves prediction accuracy. This article conducts experimental verification on PHM 2012 datasets and shows that the proposed method achieves significant improvement in prediction accuracy compared with existing RUL prediction methods.