Abstract: In the prediction of the remaining life of bearings, the traditional method has the problem of information redundancy in the multi-domain features of bearing vibration signals and dimensionality reduction of high-dimensional features. In order to solve this problem, a prediction method based on Sparse Autoencoders （SAE） feature fusion and Bi-directional Long Short-term Memory （BiLSTM） combined with Self-Attention （SA） is proposed. Prediction method. First, multi-domain feature extraction is performed on the original signal to construct an initial feature set containing time, frequency and time-frequency domain metrics; then, a comprehensive evaluation system is established based on the three dimensions of feature relevance, robustness and monotonicity, and sensitive degraded features with scores higher than the preset thresholds are screened as the input to the model, and the features extracted from time, frequency and time-frequency domains are fused and downgraded by SAE to obtain the low-dimensional fusion features. Finally, the fused features are input into the BiLSTM-SA module to further extract the timing information and assign feature weights, and validated on the PHM2012 dataset. The results show that the proposed method exhibits better performance in several evaluation indexes compared with the three models with BiLSTM-SA, SAE-BiLSTM and SAE-BiGRU, proving its effectiveness in predicting the remaining life of bearings.