Reconstruction of missing data for heavy-haul trains is an important factor in ensuring safe train operation. However, the existing methods of generative model require a complete data set for training, and it is very difficult for them to solve the issue of missing data completely at random. For this, this paper proposes a new attention-generative adversarial network to reconstruct missing data. First, a mask matrix is designed to locate the missing data, and the gradient descent algorithm is applied in combination with the output probability matrix of the discriminator, so that the mask matrix can still filling up the data well in the case of incomplete data set. Then, the prompt matrix is derived based on the mask matrix to solve the problem of model overfitting and accelerate the convergence. Finally, an attention mechanism is introduced into the whole GAN to improve the expression of data features by the feature extraction network. The experimental results show that the mean square error and mean absolute error percentage indexes of reconstruction accuracy can be kept below 1.5 for measurement data at different missing rates, and the reconstructed data can also well conform to the distribution law of measurement data.