A GNN-boosted reinforcement learning framework for maintenance optimization in multi-dependency manufacturing system

Maintenance optimization in manufacturing systems faces challenges due to multiple dependencies, including cost-sharing in maintenance operations, interactive degradation across interconnected machines, and competition for limited spare part resources. These interdependencies, together with high-dimensional decision spaces and evolving system dynamics, limit the effectiveness of conventional optimization methods such as heuristic scheduling rules and Markov decision processes (MDPs). Existing studies rarely address these heterogeneous dependencies in a unified framework. To bridge this gap, this paper proposes a Multi-Hierarchical Graph Neural Network–based reinforcement learning framework (MHGNN-PPO) for dependency-aware maintenance decision optimization. Specifically, MHGNN constructs a multi-layered graph to explicitly represent economic, stochastic, and resource dependencies. By combining hierarchical graph structures with cross-graph attention aggregation, MHGNN dynamically captures intricate inter-machine relationships and refines system state representations. Furthermore, a constraint-integrated and stability-enhanced PPO algorithm is developed to ensure feasible decision-making and reliable convergence in high-dimensional maintenance environments. To validate the effectiveness of the proposed approach, an experiment was conducted on a manufacturing system consisting of 24 machines. Comparative experiments demonstrate that the proposed MHGNN-PPO framework achieves higher cumulative rewards than baseline methods, while maintaining convergence stability across different system scales. Ablation studies revealed that excluding any component notably gains lower rewards, highlighting the necessity of explicit dependency modeling. Additionally, robustness tests further validate the adaptability and resilience of MHGNN-PPO under varying dependency intensities.