Remaining Useful Life Prediction Based on Stacked Sparse Autoencoder and Echo State Network

Nowadays, the prediction of industry components' remaining useful life (RUL) has already become a hot topic. In this paper, a RUL prediction method based on stacked sparse autoencoder (SAE) and echo state network (ESN) is proposed. Autoencoder (AE) is an unsupervised learning method that can be used for feature extraction to obtain a health index (HI). To enhance the performance of the HI, a sparse constraint is added and a stacked structure is used to increase depth. The RUL of the industry components is then predicted using a method that maps directly to the RUL value based on the health factor curve. With the capability of encapsulating dynamic time behavior and saving historical information of input data, ESN is selected as the prediction network. The proposed method is verified using the C-MPASS dataset. The experimental results show that the method has better performance than the stacked AE based method.