Deep reinforcement learning-based decision support system for optimal maintenance planning in bridge digital twins

Bridge infrastructures are inevitably exposed to a wide range of deterioration mechanisms and external damages over their service life, which can gradually impair structural integrity, compromise functionality, and increase safety risks. These challenges, coupled with the growing complexity of infrastructure systems and limited financial resources, highlight the need for implementing intelligent and proactive asset management solutions. To address these needs, this paper presents a Reinforcement Learning-based Decision Support System (RL-DSS) framework for bridge maintenance optimization, leveraging a Multi-Objective Deep Centralized Multi-Agent Actor-Critic (MO-DCMAC) algorithm that is designed for application within bridge Digital Twin environments. The proposed RL-DSS implements a three-objective MO-DCMAC framework that models maintenance cost, structural failure risk, and serviceability as separate objectives within a vector-valued reward formulation, while employing a fixed linear utility function for policy optimization. The proposed framework was evaluated through three real-world bridge case studies, each representing different service conditions, to demonstrate its capability in providing long-term maintenance planning. Its performance was benchmarked against multiple established rule-based policies commonly used in practice. The results revealed that the MO-DCMAC agent achieved the highest overall utility score, demonstrating superiority by generating more refined and practical maintenance policies. By actively penalizing high-frequency site visits and strategically balancing intervention costs with risk mitigation, the RL agent prioritizes fewer but more impactful repair actions. The proposed RL-DSS reflects an adaptive, informed, and cost-effective decision-making process that captures the complexity of real-world maintenance scenarios, surpassing the limitations of traditional rule-based approaches.