A static-dynamic hybrid graph neural network model incorporating physical information for remaining useful life prediction of aircraft engines

Abstract The remaining useful life (RUL) prediction of aircraft engines is challenging due to strong, time-varying correlations among multivariate sensor signals. This paper presents TVAE-HybridGAT, a physics–data fusion framework combining a Temporal Variational Autoencoder (TVAE) with a static–dynamic hybrid graph neural network to jointly model temporal evolution and sensor interdependencies. The TVAE encodes multivariate time series into a probabilistic latent space to obtain robust temporal features. A static graph derived from physical priors captures stable structural relationships, while a dynamic graph computed from the TVAE latent features captures transient, emergent sensor couplings. A multi-head graph attention module adaptively fuses the two graphs, emphasizing informative connections and reducing the influence of noisy or irrelevant links. Experiments on the C-MAPSS and N-CMAPSS benchmarks show that TVAE-HybridGAT attains strong and consistent RUL prediction performance across datasets, demonstrating the benefit of combining physics-informed structure with data-driven, adaptive graph learning.