As an important component in smart grid, the Feedback Terminal Unit(FTU) occasionally faces the unexpected shutdown due to the extreme operation environment. The failure rate prediction of massive Feedback Terminal Units(FTUs) is investigated by using the Stacked Denoising Autoencoder (SDAE) failure rate estimation method improved by the Dropout Regularization operation to prevent overfitting. Adadelta algorithm is employed to optimize the learning rate. An accurate failure prediction is realized with satisfied learning rate. A series of experiments are conducted to verify the advantages of the proposed method in solving the FTU failure estimation problem