Search Results

Predictive maintenance of multi-component systems requires decisions that account for both local component degradation and system-level redundancy. This paper proposes a formal framework for synthesizing interpretable maintenance policies from a system functional architecture. The system is modeled as a network of Dynamic Priced Timed Game Automata (DPTGA): each component automaton encodes degradation dynamics, stochastic failure windows, and repair and replacement costs, while a System Health Monitor automaton is automatically derived from the minimal cut sets of the functional architecture and detects system failure as soon as any cut set becomes unavailable. We execute the DPTGA semantics in an explicit-state simulator and optimize a compact threshold policy with the Cross-Entropy Method (CEM). The policy contains one age-fraction threshold per component and an embedded subsystem-safe predicate that forbids any preventive action that would immediately realize a cut set, making the learned strategy provably safe and directly interpretable as a rule-based maintenance plan. On a seven-component benchmark comprising a parallel subsystem and a 2-out-of-5 redundant block, the CEM policy reduces mean cumulative cost by 15% and median cost by 16% relative to the best rule-based heuristic (corrective, systematic, and condition-based maintenance), while maintaining a 98.7% system survival rate. The approach therefore provides a compact and explainable alternative to deep reinforcement learning for architecture-aware predictive maintenance.