Remaining Useful Life Estimation for Systems with Abrupt Failures

Abstract

Data-driven Remaining Useful Life (RUL) estimation for systems with abrupt failures is a very challenging problem. In these systems, the degradation starts close to the failure time and accelerates rapidly. Normal data with no sign of degradation can act as noise in the training step, and prevent RUL estimator model from learning the degradation patterns. This can degrade RUL estimation performance significantly. Therefore, it is critical to identify degradation mode during the training step. Moreover, in the application step, predicting RUL when the system is in normal mode and is not showing any sign of degradation can generate inaccurate estimations, and reduce faith in the model. In this paper, we propose a new RUL estimation method that incorporates an early degradation mode detection step to automatically identify the earliest point of time at which the degradation starts to happen. When the degradation mode is detected, a Long Short Term Memory (LSTM) neural network is applied to predict system RUL. As a case study, we apply the proposed method for RUL estimation in 2018 PHM Data Challenge. The case study demonstrates that our solution achieves more accurate RUL estimation compared to several baseline methods.