Failure Threshold Setting for Wiener-process-based Remaining Useful Life Estimation

Abstract

Wiener-process-based methods are frequently used in remaining useful life (RUL) estimation for prognostic health management of industrial systems. RUL can be estimated by the time difference between the current time and the expected first hitting time at which a Wiener-process-based degradation process exceeds a pre-set failure threshold (FT) for the health indicator values. Although FT setting significantly impacts the prediction accuracy, there is no known standard way to determine the FT for a given degradation process of health indicator values. In this paper, we address the issue of FT setting for a Wiener-process-based RUL estimation method and experimentally investigate the impacts of FT setting on the prediction accuracy. We apply a simple Wiener-process model to predict the lifetime of turbofan engines from the sensor data in the benchmark dataset. Our experimental results show that the FT setting by the mean value of health indicators before the failure times in the training data sets tends to achieve higher prediction accuracy than the FT setting using the health indicator values at the failure times. The results imply that a conservative threshold setting is preferable in terms of prediction accuracy for Wiener-process-based RUL estimation.