Early fault warning of power plant auxiliary engine based on improved stacked autoencoder network

In order to improve the predictive ability of auxiliary engine faults, combined with the advantages of unsupervised learning methods in deep learning, an early fault warning method for power plant auxiliary engine based on improved stacked autoencoder network is proposed. The method takes the historical normal data of the auxiliary engine as the training set, utilizes the nonlinear expression ability of the stacked autoencoder (SAE) network to express the relationship between the variables of the auxiliary engine, and introduces the batch normalization (BN) algorithm to optimize network performance. For the input observation vectors, the SAE network gives the corresponding reconstruction vectors, constructs the similarity based on the fusion distance to represent the deviation between the observation vector and reconstruction vector. When the auxiliary engine starts to deviate from the normal state, the deviation between the observed value and reconstructed value increases, and the similarity drops to the warning threshold, which indicates that the engine fault appears. The normal data and fault data of the medium speed coal mill of a certain thermoelectric unit are used to conduct test and verification respectively. The results show that the SAE network with BN algorithm introduced has lower reconstruction error. The proposed fault warning method can make early warning before the coal mill is tripped, which indicates that the method can effectively make fault warning of auxiliary engine fault and has certain engineering application value. © 2019, Science Press. All right reserved.