Abstract

Existing methodologies for predictive maintenance in aviation engines have largely diverged into data-centric and physics-centric models, each constrained by their unchangeable limitations. In this study, a hybrid frame work integrating both perspectives was developed to address fatigue-induced failures in the CFM56-7B en gine. Specifically, a Bayesian Physics-Informed Neural Network (B-PINN) was constructed, embedding Paris' Law within a deep learning structure and modeling key fatigue parameters as probabilistic distributions. Selected sensor data from NASA's CMAPSS FD004 dataset was employed to assume latent stress signals and simulate fatigue crack propagation. The results show that the proposed model has advantages on interpret ability and reliability of fatigue predictions but also quantifies uncertainty through variational inference and Monte Carlo dropout

DOI: doi.org/10.63721/25JPAIR0109

To Read or Download the Article  PDF