In order to construct a trend health index that accurately characterizes the degradation process of rolling bearings and im ? prove the prediction accuracy of remaining useful life （RUL） of rolling bearings， a neural network model （LSTM-SA） combining long-short term memory （LSTM） and self-attention mechanism （Self-Attention） is proposed for RUL prediction of rolling bear? ings. The envelope spectrum of the original signal is obtained by using envelope demodulation， and then the envelope spectrum is segmented and the Pearson correlation coefficients of the corresponding frequency bands are calculated to extract the degradation features with monotonicity and trend. The degradation features are normalized and processed as the input of the LSTM-SA model， and the LSTM is used to adaptively extract the temporal internal correlation of the degradation features and the Self-Attention is used to screen key information. By eliminating the interference of useless information and mining deep-level features， healthiness in? dexes are constructed and degradation curves can be obtained. By determining the failure threshold， the degradation curves are fit? ted by the least squares method， and the life failure point is predicted， which realize the RUL prediction of rolling bearings. The ex? perimental results on the PHM2012 dataset show that the proposed method reduces the average absolute error by 43.99%， 63.11% and 60.00%， respectively， and improves the average score by 10.87%，45.71% and 34.21%， respectively， compared with other literature. The experimental results on the actual engineering data show that the average prediction error of the proposed method is higher than that of standard-RNN and CNN by 39.58% and 74.86%， respectively.