Toward accurate RUL and SoH estimation using reinforced graph-based physics-informed neural networks enhanced with dynamic weights
Title: Enhancing RUL and SoH Precision via Reinforced Graph-Based Physics-Informed Neural Networks with Dynamic Weighting
Abstract:
Precise forecasting of Remaining Useful Life (RUL) and State of Health (SoH) is a cornerstone of effective Prognostics and Health Management (PHM), enabling proactive maintenance strategies and ensuring robust industrial operations. Despite this necessity, hybrid approaches that merge data-driven learning with physics-based constraints typically suffer from reduced accuracy when applied to assets exhibiting distinct degradation patterns, largely due their reliance on static loss weights. To address this limitation, we propose Reinforced Graph-based Physics-informed Networks with Dynamic Weighting (RGPD), a comprehensive framework designed for spatio-temporal degradation modeling and adaptive, physics-guided regularization.
The RGPD architecture employs graph-based representation learning to map the structural relationships of degradation across sensors. A Soft Actor-Critic (SAC) module is integrated to refine latent features, ensuring resilience against noisy data environments. Furthermore, a lightweight Q-learning policy dynamically adjusts the balance among monotonicity, smoothness, and latent-dynamics residual losses throughout the training process.
We validated the proposed framework using three distinct datasets representing engine, bearing, and battery degradation: C-MAPSS, PHM2012, and XJTU. Comparative analysis against the top-performing baselines in existing benchmark tables reveals significant performance gains. Specifically, RGPD achieves an average RMSE reduction of up to 12% on both the PHM2012 and C-MAPSS datasets. Additionally, it lowers the average MAPE by 20% on the XJTU dataset when compared to the second-best reported model. These results across heterogeneous benchmarks highlight the model’s strong generalizability across various degradation systems.
The physics-informed aspect of RGPD is realized through degradation-consistent priors and a residual structure inspired by the Deep Hidden Physics Model. This design enhances physical plausibility in the predictions without necessitating the development of full first-principles models for every specific asset type.
Source: arXiv Generated at: 2026-06-02 00:00:00 UTC
