Hierarchical domain-physical topology enhanced Fully-Connected Spatial-Temporal Graph Neural Network for remaining useful life prediction

Accurate and reliable Remaining Useful Life (RUL) prediction is paramount for ensuring the operational reliability and safety of complex industrial systems. However, existing deep learning paradigms often fail to incorporate explicit structural biases derived from physical dependencies, neglecting inherent physical mechanisms and failing to provide rigorous uncertainty quantification. To surmount these challenges, this paper proposes a Hierarchical Domain-Physical Topology Enhanced Fully-Connected Spatial-Temporal Graph Neural Network (HDPT-FC-STGNN). First, to bridge the heterogeneity between symbolic domain knowledge and continuous sensor streams, a Hierarchical Domain-Physical Topology (HDPT) is constructed leveraging high-dimensional semantic embeddings. A hierarchical message-passing mechanism is subsequently devised to explicitly encode topological dependencies, thereby fusing static physical priors with dynamic sensor streams. Second, to mitigate spurious correlations in noise-intensive environments, a Structure-Aware Probabilistic Graph Sampling strategy is introduced. By selectively injecting stochasticity into unreliable data-driven edges, this mechanism treats the graph topology as a latent variable, bolstering the model robustness against epistemic ambiguity. Finally, a Bayesian uncertainty quantification module, governed by a risk-sensitive hybrid loss function, is implemented to yield high-fidelity RUL estimates alongside well-calibrated confidence intervals. Extensive experiments on the C-MAPSS benchmark demonstrate that the proposed method significantly outperforms state-of-the-art baselines in both deterministic precision and probabilistic reliability, notably reducing the RMSE by approximately 6.9% on the FD001 subset while ensuring high probabilistic reliability.