Hierarchical multi-agent reinforcement learning for operations and maintenance disruption recovery in high-end equipment manufacturing under compound risks

Compound disruptions in high-end equipment manufacturing can simultaneously fragment operations and maintenance (O&M) resource networks and constrain critical resource flows. Using natural disasters and trade sanctions as representative compound risks of topology-based network fragmentation and capacity-constrained flow bottlenecks, this study addresses a recovery problem in which network reconfiguration changes the feasible space of downstream resource scheduling, while scheduling outcomes feed back through system-level evaluation to influence upstream reconfiguration. The problem is further complicated when multiple projects compete for heterogeneous, category-specific resources under contractually binding delivery requirements. An integrated recovery model based on a dual-layer business–resource network is developed to coordinate network reconfiguration and resource scheduling while balancing recovery cost, duration, and pending repair rate. To solve the coupled problem, a dual-agent algorithm based on hierarchical multi-agent proximal policy optimization (HMA-PPO) is proposed. Its main architectural feature is a hierarchical conditional coordination mechanism: the reconfigured network is transmitted forward to construct the state and feasible decision conditions of downstream resource scheduling, while the network-reconfiguration and resource-scheduling decisions are jointly evaluated as an integrated recovery solution, with the resulting system-level evaluation shared by both agents for policy updating. Numerical experiments show that HMA-PPO achieves the lowest mean objective value among the tested baselines, reaches the exact optimum on all three tested small instances, and requires approximately 0.24 s for post-training online decision making. Its relative solution-quality advantage becomes more evident as the tested decision space expands; on the large network, the mean objective value is 12.84% lower than that of Multi-Agent Proximal Policy Optimization (MAPPO) and 7.21% lower than that of Adaptive Large Neighborhood Search (ALNS). Scenario analyses further show that compound risks increase total O&M cost by 25.17% on average relative to the single natural-disaster scenario, while backup-supplier configuration and operating margins materially affect recovery resilience. These findings provide a unified perspective on coupled structural recovery and constrained resource allocation for resilience management in advanced manufacturing systems and operations research. Enterprises should diversify backup-supplier portfolios while jointly assessing supplier quality, usable capacity, transportation conditions, and delivery reliability rather than relying on supplier count alone. They should also use stress testing to identify scenario-specific operating buffers and prioritize scarce resources according to project urgency, resource criticality, and contractual delivery requirements.