Optimizing Long-Term Maintenance Pathways for Sustainable Infrastructure: A POMDP and Reinforcement Learning Approach with Explainable Knowledge Layers

Introduction: Traditional infrastructure risk management heavily relies on static prioritization frameworks that fail to capture the multi-stage, time-variant nature of asset deterioration, sequential budgetary constraints, and partial observability stem from imperfect inspection data. This study addresses these gaps by implementing an artificial intelligence (AI)-assisted dynamic decision-support framework designed to map optimal multi-year strategic intervention pathways under deep epistemic uncertainty—formally defined as the persistent divergence between the true physical state of an asset and the noisy, imperfect sensor observations available to decision-makers.

Methods: The framework integrates a dual-layer architecture: (1) a Core Dynamic Engine utilizing a Partially Observable Markov Decision Process (POMDP) combined with a Proximal Policy Optimization (PPO) reinforcement learning agent to handle continuous multi-asset resource allocations subject to strict annual budgetary constraints; and (2) an Explainable AI (XAI) surrogate layer leveraging a tree-based Random Forest classifier and SHAP values to extract transparent governance rules from the black-box policy network.

Results: Validated across a 20-year operational horizon using a semi-empirical campus facility benchmark base aligned with real-world deterioration statistical bounds, the proposed POMDP-PPO framework achieved an absolute cumulative reward shift of +1.80 units over traditional dynamic Analytic Hierarchy Process (AHP) greedy baselines and significantly lowered asset failure rates to 2.1%. Sensitivity analyses confirmed structural policy robustness against severe sensor telemetry measurement noise (up to 150% standard deviation inflation). The global surrogate model demonstrated high discrete action classification accuracy (macro F1-score = 0.95) and robust continuous value function approximation ($R^2 = 0.96$), effectively formalizing actionable engineering logic rules.

Conclusions: The results indicate that coupling robust stochastic optimization with post-hoc explainability layers successfully mitigates institutional trust resistance. The architecture demonstrates significant scalability potential for large-scale municipal portfolios, satisfying public accountability standards while sustaining infrastructure resilience