To achieve accurate prediction of the remaining useful life (RUL) of brake discs, ensure train braking safety, and optimize economical maintenance, this paper proposes a prediction model based on the fusion of selfattention mechanism and long short-term memory network (BiLSTM- SA), which takes the crack propagation life as the division basis. Firstly, the test data of brake discs are collected and the working conditions are calibrated, and a thermal-mechanical coupling finite element model is established to obtain the simulation dataset. Secondly, a Time-GAN neural network is constructed, which enhances data through a double-layer LSTM generator and a physical constraint discriminator. Its distribution similarity, root mean square error and coefficient of determination are significantly better than traditional models. Finally, the BiLSTM-SA fusion prediction model is proposed, which uses bidirectional LSTM and self-attention mechanism to capture temporal dependencies and key features. In the prediction of single expanding cracks, the RMSE is reduced by 49.8% and 46.5% compared with the traditional LSTM and TCN-LSTM, respectively. In complex working conditions, the RMSE and Score are optimized by 25.5% and 51.1% , respectively, significantly improving the prediction accuracy and robustness.This study can provide a reliable technical solution for condition monitoring and preventive maintenance of highspeed train brake discs.