Constrained time-dependent loss LSTM for bearing remaining useful life prediction

Remaining useful life (RUL) prediction is an essential prognostic health management and maintenance method. Data-driven RUL prediction methods with various neural network (NN) architectures have been widely adopted in previous literature. The inconsistency of prediction by less discriminative feature combinations is a significant problem in the NN-based method, leading to insufficient convergence and irrational predictions. In this work, we study the optimization of a long short-term memory (LSTM) network architecture to improve the accuracy of RUL prediction by constraining time-dependent loss of LSTM (CTDM-LSTM) according to the degradation. First, we focus on the reasonable range of a good prediction from zero to actual RUL. The error beyond this range is more weighted to control the loss gradient. Additionally, we consider the relevance of prediction uncertainty over time by reducing the loss at an early stage of degradation. The experiment results demonstrate that the proposed loss function optimization method can effectively predict the RUL of rolling bearings and show higher prediction accuracy.