Health Management (PHM) is a comprehensive method and strategy that aims to monitor, diagnose and optimize the health status and performance of systems, equipment or components. It includes health monitoring, fault diagnosis, and remaining service life (URL) forecast. Among them, URL is a more important field in PHM. Because the physical knowledge used by precise modeling machines is complicated, data -based driving methods based on learning data have become a promising alternative method for model prediction methods. This article proposes a new time convolutional neural network (TCN) with a soft threshold and attention mechanism for mechanical prediction. The method used in this paper avoids the process of complicated artificial extraction, and uses bearings directly as the input of the network model. A soft threshold mechanism is added to the TCN network, so that more useful information is retained in the activation function ReLu. The threshold is self-adaptive learning by the sub-network trained by the channel attention mechanism, not the certainty value of the threshold distribution. Therefore, under the influence of the attention mechanism, each feature diagram has its own threshold, which retains some key features. On behalf of verify the generalization of the method proposed, the three benchmark data sets related to rolling bearings were tested, and the performance of the development method was compared with several advanced prediction methods. The results show that the development of all three cases has produced accurate RUL predictions, and has good robustness and generalization capabilities.