A method for payload capacity estimation and program angle reconstruction is proposed based on neural networks to deal with the problem that it is difficult to reconstruct the trajectory after thrust drop. Optimal trajectories with thrust drop faults are generated offline based on engineering practice to train the neural networks. Features of the program angle are quite different in stage 1 and other stages. Different modes are taken. Impact of hyperparameters on the training process is also analyzed and a random search method is taken to optimize the hyperparameters. The proposed method is data-based and can avoid the complicated dynamics of the launch vehicle， and the trajectory reconstruction can even be used in endo-atmosphere. Numerical simulation shows that the proposed method can estimate the payload capacity accurately and the reconstructed program angle has a small error compared with the optimal trajectory. The arithmetic speed of the proposed method is also faster than those of other methods.