Risk-Aware Aircraft Fleet Maintenance Under Partial Observability and Airport Resource Constraints

Aircraft fleet maintenance must translate uncertain health information into executable decisions under shared resource constraints. We present belief-guided risk-aware multi-agent scheduling (BRiMS), a framework linking probabilistic health beliefs and diagnostic inspection to hierarchical maintenance decisions and explicit execution checks. Action-level records identify how the policy, controller and environment contribute to constraint satisfaction. Multi-seed simulations evaluate 10-aircraft and 200-aircraft fleets. In the main small-fleet stress case, availability is 0.9217, high-risk exposure is 0.0075 and task completion is 38.9%. The corresponding large-fleet values are 0.916, 0.0077 and 44.0%. Both cases satisfy the prescribed safety and availability constraints in every evaluated episode. Under normal small-fleet demand, completion reaches 82.5% and availability reaches 0.963. Paired comparisons with recurrent proximal policy optimisation (RPPO) support lower high-risk exposure and constraint violation. Diagnostic-quality experiments quantify inspection benefits. Risk-weight, controller-off and frozen-policy tests further identify the conditions in which execution controls sustain feasibility. These results support auditable risk-aware maintenance coordination within the evaluated settings. Joint reporting of risk, availability and service completion provides a basis for adding service requirements and calibrating the framework to operational fleets.