Attention-based Gate Recurrent Unit for remaining useful life prediction in prognostics

Abstract

An essential process in prognostics and health management (PHM) is remaining useful life (RUL) prediction. The traditional Recurrent Neural Networks (RNNs) and their variants are not very efficient at solving the regression problems of RUL prediction. Given this problem, an attention-based Gate Recurrent Unit (ABGRU) for RUL prediction is proposed in this paper. Firstly, the dataset is preprocessed, and the RUL labels are modeled using the piecewise linear degradation method. Then, a GRU network based on an encoder–decoder framework with an attention mechanism is proposed. The network can assign weights according to the importance of feature information and effectively use the feature information to predict RUL. The validity of the proposed framework is verified in the NASA C-MAPSS benchmark dataset. The results show that the presented method outperforms the existing state-of-the-art approaches and provides a new solution for RUL Prediction.