A confidence-aware multi-fidelity digital twin framework for time-critical incident-driven AMR fleet management

Abstract Effective disruption response in Autonomous Mobile Robot (AMR)-enabled shop floors requires decisions that are not only accurate but also temporally aligned with the evolving physical system. In practice, this requirement is difficult to satisfy using single-fidelity Digital Twin (DT) approaches, whose computational latency can exceed the operational decision window. This paper presents a confidence-aware multi-fidelity DT framework for incident-driven AMR fleet management, designed to support deployable scheduling decisions under real-time constraints. The framework integrates three digital abstraction levels: a real-time Digital Shadow (DS) for state acquisition via OPC-UA, a bounded-time surrogate optimiser (NSGA-II) for rapid rescheduling, and a high-fidelity Simulink DT for post-decision execution audit. Fidelity allocation is governed through a confidence-aware mechanism based on three measurable reliability dimensions: execution uncertainty, communication integrity, and temporal alignment. This enables conditional authorisation of surrogate decisions within empirically validated operating envelopes. The framework is implemented through a structured data and decision pipeline and evaluated on a stochastic battery-module assembly flow shop under machine breakdown (MBD), AMR breakdown (AMRBD), and demand-change (DC) scenarios. Results show that the surrogate produces actionable schedules within 3-10 s while maintaining close agreement with high-fidelity execution (MAPE $$\le $$ 1.03% for makespan and $$\le $$ 0.94% for energy). A comparative delay analysis further shows that single-fidelity response latencies of 230-410 s lead to unrecoverable congestion equivalent to 4-7 continuously waiting jobs. The results demonstrate that temporal alignment, rather than model fidelity alone, is a governing design constraint for DT-enabled disruption response in manufacturing systems.