A Novel Two-Stage Prognostics Framework for Aero-Engine Remaining Useful Life Estimation Under Dynamic Operating Conditions

Deep learning-based remaining useful life (RUL) prediction for aero-engines has been widely used to forecast impending failures and ensure flight safety. However, most existing methods ignore the intricate impacts of dynamic operating conditions on the monitoring data gathered by multiple sensors, thus hindering accurate estimation of RUL. To address this problem, we propose a novel two-stage prognostics framework for aero-engine RUL estimation under dynamic operating conditions. The core idea is to disaggregate the health and operational condition information in the monitoring data as much as possible to ensure that the input data to the estimation model is least likely to be negatively affected by dynamic operating conditions, thus improving the RUL prediction accuracy. Specifically, the first stage starts with training an unsupervised variational autoencoder with data decomposition capabilities. Subsequently, the well-trained encoder and decoder components with uniform operating condition data will generate high-quality monitoring data embedded with health information and without distractions caused by dynamic operating conditions. In the second stage, high-quality monitoring data are combined with current RUL estimation methods to extract degradation characteristics and make accurate RUL predictions. Finally, the effectiveness and superiority of the proposed framework are demonstrated using publicly available aero-engine degradation datasets.