Toward accurate RUL and SoH estimation using reinforced graph-based physics-informed neural networks enhanced with dynamic weights

Researchers present RGPD, a physics-informed neural network framework that dynamically balances multiple loss functions to improve Remaining Useful Life (RUL) and State of Health (SoH) predictions across industrial assets. The model achieves up to 20% improvement in accuracy over existing methods by combining graph-based representation learning with reinforcement learning-driven adaptive weighting, demonstrating strong generalization across engine, bearing, and battery degradation datasets.

This research addresses a persistent challenge in industrial prognostics: creating models that maintain accuracy when deployed across assets with fundamentally different degradation patterns. Traditional hybrid physics-informed approaches use fixed loss weights that optimize for specific conditions, limiting their transferability—a critical constraint when maintaining diverse equipment fleets. RGPD overcomes this by introducing dynamic weighting via Q-learning, allowing the model to automatically adjust how strongly it enforces physical constraints (monotonicity, smoothness, latent dynamics) based on real-time training conditions.

The innovation sits at the intersection of three methodological advances: graph neural networks capture inter-sensor relationships that reveal degradation structure; soft actor-critic reinforcement learning refines feature representations in noisy industrial environments; and adaptive loss balancing eliminates manual hyperparameter tuning across asset types. This architectural design acknowledges that physics priors remain valuable—they prevent physically implausible predictions—but their implementation must be flexible rather than rigid.

For industrial operators and maintenance teams, these improvements translate directly to operational benefits. A 12% reduction in RMSE on standard benchmarks like C-MAPSS correlates to more reliable predictive maintenance scheduling, reducing unplanned downtime and extending asset lifecycles. The 20% MAPE improvement on battery degradation data holds particular significance as energy storage systems gain criticality across grid modernization and electric vehicle infrastructure.

The framework's demonstrated generalizability across heterogeneous degradation domains suggests scalability to real-world production environments where equipment diversity is the norm rather than exception. Future adoption depends on integrating this approach into existing Condition-Based Maintenance (CBM) systems and validating performance on proprietary industrial datasets where ground-truth degradation curves remain unavailable.

• →RGPD uses reinforcement learning to dynamically balance physics-informed loss weights, eliminating fixed hyperparameter constraints that limit cross-asset generalization.
• →The framework achieves 12-20% performance improvements over baseline models on three diverse industrial degradation benchmarks (engines, bearings, batteries).
• →Graph-based representation learning captures inter-sensor degradation relationships, improving feature quality in noisy industrial data environments.
• →Physics priors remain embedded through degradation-consistent constraints and residual modeling, improving plausibility without requiring asset-specific first-principles models.
• →Results suggest practical applicability to real-world condition-based maintenance systems facing diverse equipment types with varying degradation behaviors.