Aiming at the strong nonlinearity and high uncertainty in the flight process of launch vehicle, and the significant impact of thrust descent faults on the reliability and safety, an interpretable machine learning model based on attention mechanism is proposed to improve the accuracy and robustness of thrust descent fault detection, fault engine location, fault degree estimation, and trajectory prediction after faults. The attention layer is used to extract the features of the high-dimensional time series flight monitoring data, and the feature matrix is used to express the high-dimensional time series data succinctly. Then the self-attention and fully connected network are used to predict the position and degree of thrust descent, the feature vector is decoded by the long-term and short-term memory unit to realize the accurate prediction of flight trajectory. The proposed integrated model is tested on the thrust descent data set to verify the effectiveness. The results show that the accuracy of the proposed model is 96. 0% for the fault location, the accuracy is 94. 7% for the fault severity estimation, and the average trajectory prediction error is 0. 94%. The proposed model has good application effect in thrust descent fault modes.