Rec-PdM: Decision-calibrated recursive multi-agent systems for uncertainty-aware prescriptive maintenance in smart manufacturing

Abstract

Predictive maintenance creates industrial value only when uncertain health evidence leads to an appropriate intervention under asymmetric failure costs and operational constraints. We present Rec-PdM, a decision-calibrated recursive multi-agent system for uncertainty-aware prescriptive maintenance. Its six decision agents maintain specialized beliefs about asset condition, fault state, remaining useful life, economic risk, action selection, and verification; production-impact and human-prescription interfaces connect these beliefs to factory context and operator-facing decisions. RecursiveLink reconciles local beliefs with downstream decision consequences while preserving the provenance of each update. The method treats maintenance as a partially observable degradation process and calibrates prediction and uncertainty against decision regret, tail risk, feasibility, and communication overhead. Evaluation on six public datasets spans degradation prognosis, process failure, vibration-based diagnosis, and acoustic anomaly detection, with controlled decision studies of cost asymmetry, coordination, resource constraints, dependent faults, and operating drift. RecursiveLink achieves the best mean-cost rank among the evaluated coordination mechanisms, while calibrated Rec-PdM ranks in the top two in 98.3% of the evaluated cost regimes without an unexpected failure. Under shared fleet constraints, the coordinator remains in the top two by mean cost in 88.8% of 240 scenarios; prequential recursive guarding reduces unserved failures from 70 to 5 under drift. These results show that decision cost, tail risk, and traceable coordination complement predictive accuracy when maintenance evidence must be converted into action.