A single feature cannot show the operational state of a bearing during its entire life cycle. Therefore, a rolling bearing performance deterioration prediction method based on an SAE and the TCN-attention model is proposed. The SAE method is used to fuse the timedomain indicator and the frequency-domain indicator to construct the performance degradation characteristic indicator. The evaluation indices are used to comprehensively evaluate multiple performance degradation indices, and the fused feature indices together, to filter out the features that have a good overall performance. Attention is added to the TCN model, and the output state weight of the TCN model is calculated through a scoring function to increase the important information weight and the prediction accuracy. The appropriate network structure and parameter configuration are determined, and the rolling bearing performance degradation prediction model is established. A validation is performed using publicly available datasets from the University of Cincinnati and XJTU-SY. The results show that the method is more sensitive to the critical information part of the long time series than the other models. At the same time, the average absolute error and the root mean square error are minimized, the accuracy of the rolling bearing performance degradation prediction is high, and the model has a strong robustness and generalization abilities. Additionally, the model has practical engineering value for predicting the health status of equipment.