Learning for predictions: real-time reliability assessment of aerospace systems

View Video Presentation: https://doi.org/10.2514/6.2021-1478.vid Prognostics and Health Management (PHM) aim to predict the Remaining Useful Life (RUL) of a system and to allow a timely planning of replacement of components, limiting the need for corrective maintenance and the down time of equipment. A major challenge in system prognostics is the availability of accurate physics based representations of the grow rate of faults. Additionally, the analysis of data acquired during flight operations is traditionally time consuming and expensive. This work proposes a computational method to overcome these limitations through the dynamic adaptation of the state-space model of fault propagation to on-board observations of system's health. Our approach aims at enabling real-time assessment of systems health and reliability through fast predictions of the Remaining Useful Life that account for uncertainty. The strategy combines physics-based knowledge of the system damage propagation rate, machine learning and real-time measurements of the health status to obtain an accurate estimate of the RUL of aerospace systems. The RUL prediction algorithm relies on a dynamical estimator filter, which allows to deal with nonlinear systems affected by uncertainties with unknown distribution. The proposed method integrates a dynamical model of the fault propagation, accounting for the current and past measured health conditions, the past time history of the operating conditions (such as input command, load, temperature, etc.), and the expected future operating conditions. The model leverages the knowledge collected through the record of past fault measurements, and dynamically adapts the prediction of the damage propagation by learning from the observed time history. The original method is demonstrated for the RUL prediction of an electromechanical actuator for aircraft flight controls. We observe that the strategy allows to refine rapid predictions of the RUL in fractions of seconds by progressively learning from on-board acquisitions.