Chronologically Guided Deep Network for Remaining Useful Life Estimation

Abstract

In this paper, we introduce a new chronological loss function for training models to predict remaining useful life (RUL) of industrial assets based on multivariate time-series observations. The chronological loss, an alternative to the more traditional mean-squared error (MSE) loss, incorporates a monotonicity constraint, an upper bound, and a lower bound on the RUL estimates at each time step. We also present a fully-convolutional network (FCN) as a superior competitor to the current state-of-the-art approaches that are based on LSTM. Our experiments on public benchmark datasets demonstrate that deep models trained using chronological loss outperform those trained using the traditional MSE loss. We also observe that the proposed FCN architecture outperforms LSTM-based predictive models for RUL estimation on most datasets in this study. Our experiments demonstrate the potential of the proposed models to assist in observing degradation trends. Finally, we derive a sensor-importance score from the trained FCN model to enable cost savings by minimizing the number of sensors that need to be placed for asset monitoring without sacrificing RUL estimation accuracy.