A Novel Framework for Machine Remaining Useful Life Prediction Based on Time Series Analysis

The use of Deep Neural Network (DNN) has made great strides in data-driven engineering maintenance and prognostics. As a powerful neural network structure, recurrent neural network (RNN) has achieved very good results in the remaining useful life (RUL) estimation problem, because of its extraction of long-term dependency information from the historical sensor signal. This work focuses on a two-step method containing two RNN models for RUL estimation. A series prediction model is trained for predicting subsequent signal series from existing sensor signal sequences, whereas another model is trained for estimating RUL from the combination of the existing signal sequences and the predicted signal series. Better results are obtained by this method on FD-001 and FD-003 in NASA C-MAPSS dataset.