Estimating and Leveraging Uncertainties in Deep Learning for Remaining Useful Life Prediction in Mechanical Systems

Many researchers in the prognostics and health management community have begun exploring the use of deep neural networks for predicting remaining useful life (RUL) of mechanical systems. These models have consistently reestablished the state-of-the-art in RUL prediction performance on common benchmarks, such as the NASA C-MAPSS Aircraft Engine dataset. However, they do not attempt to capture the multiple sources of uncertainty that are inherent in their predictions. This paper presents an approach for estimating both epistemic and heteroscedastic aleatoric uncertainties that emerge in deep neural network models that are trained for RUL prediction and demonstrates that quantifying their overall impact on predictions can be extremely valuable in real-world systems, where decisions are sometimes made during uncertain operating conditions. First, a novel deep neural network architecture is proposed that demonstrates competitive performance on the NASA C-MAPSS FD001 and FD003 datasets. Then, this network is adapted to estimate epistemic and heteroscedastic aleatoric uncertainties in the RUL prediction problem. Finally, a study is carried out to observe the effects that augmenting the RUL truth data, i.e. utilizing piecewise linear truth curves in place of the actual truth data, have on the perceived uncertainties in the system. Case studies on the C-MAPSS FD001 dataset will show that utilizing the actual RUL truth data can yield more meaningful uncertainty estimates and more insight into the relationship between sensor data and an engine's time-to-failure.