To improve the prediction ability of auxiliary equipment fault, a fault warning method for power plant auxiliary equipment is proposed. It is based on improved stacked autoencoder network, which fuses the advantages of unsupervised learning methods in deep learning. The method takes the historical normal data as the training set and utilizes the nonlinear expression ability of the stacked autoencoder (SAE) network to indicate the relationship among the variables of the auxiliary equipment. The batch normalization (BN) algorithm is utilized to optimize network performance. For the input observation vector, SAE network offers the corresponding reconstruction vector. The similarity based on the fusion distance is constructed to represent the deviation between the observation vector and the reconstruction vector. When the auxiliary equipment starts to deviate from the normal state, the difference between the observed value and the reconstructed value increase. The similarity drops to the warning threshold, which indicates that the machine is fault. Normal data and fault data of a medium speed mill of a thermoelectric unit are used to test and verify the proposed method. Experimental results show that SAE network with BN algorithm has lower reconstruction error. The model can provide fault warning before the coal mill trips. Therefore, the method can effectively make fault warning of auxiliary equipment, which has certain engineering application value.