An interpretable variational attention-weighted feature decoupling network for remaining useful life prediction of aero-engines under time-varying operating conditions

Abstract Operating conditions of aero-engines often change over time, complicating the prediction of their remaining useful life (RUL). Most RUL prediction methods overlook the impact of time-varying operating conditions on feature extraction and fail to account for the potential interference of operating condition information in sensor data features. Additionally, deep-learning-based methods lack sufficient interpretability. To address these issues, we model sensor data as a coupling of operating condition and equipment degradation information, and propose an interpretable model named variational attention-weighted feature decoupling network (VAFD-Net) for RUL prediction of engines under both discrete and continuous time-varying operating conditions. VAFD-Net separately extracts operating condition features and sensor data features, then uses variational attention weights representing operating condition information to weight sensor data features, improving the model’s robustness. VAFD-Net also introduces three constraints to decouple operating condition and degradation information in latent space, mitigating the impact of signal non-stationarity. Experimental results on the Commercial Modular Aero-Propulsion System Simulation (CMAPSS) and N-CMAPSS datasets indicate that VAFD-Net not only achieves high prediction accuracy under time-varying operating conditions but also reveals the contribution of physical quantities to the engine’s health states through weighted feature maps. Furthermore, latent variable distribution plots enable users to directly infer the RUL based on the position of the latent variables.