Probabilistic ML based RUL prediction of a turbofan engine

In the last several years, Prognostics and health management (PHM) have benefited greatly from advancements in machine learning (ML) and deep learning (DL). Despite these advancements, the uncertainty in ML predictions is often not adequately addressed. As black boxes, ML and DL technologies are often employed to provide simply point estimates of remaining usable life (RUL) without taking uncertainties into account. For safer and more informed decision-making, it is crucial to precisely quantify the level of certainty in these estimates and present them as confidence intervals. This paper focuses on the issue of quantifying the uncertainty in ML predictions, especially RUL estimates.Two probabilistic ML approaches are applied in the estimation of the RUL of a Twin-Flow Turbojet. When estimating the turbojet RUL prediction uncertainty, the parametric approach seeks to maximize probability while lowering model variance to provide a reduced prediction uncertainty and, thus, a more dependable RUL prediction. The quantile regression method relies on minimizing the quantile loss function. The two methods produced similar results, in the form of a curve of estimated RUL confidence intervals with respect to the operating cycle number. For the parametric approach, this interval is determined by the compromise parameter; for the non-parametric approach, it is determined by the quantile levels. We conducted our investigation using the Train_FD003 learning base from NASA’s CMAPPS (Commercial Modular Aero-Propulsion System Simulation) and Google Colab environment.