Synchronization-Constrained Multi-Agent Scheduling for Low-Latency Industrial Digital Twin Networks

Industrial digital twin networks require low-latency delivery and temporal alignment of correlated multi-source traffic under dynamic congestion. This paper proposes a synchronization-constrained multi-agent traffic-scheduling method for industrial digital twin networks. The method combines priority-based traffic shaping, a calibrated communication-level digital twin, and an event-time-aware constrained asynchronous multi-agent proximal policy optimization (Async-MAPPO) scheduler. Communication nodes are modeled as cooperative agents under centralized training and distributed execution. The objective considers end-to-end delay, link-load concentration, packet-delivery failures, and the arrival-time span of correlated packet groups. A non-negative Lagrange multiplier is updated using the batch-average synchronization residual, while an event-time-dependent discount factor accounts for unequal physical intervals between consecutive network events. The method is evaluated using an NS-3-based communication digital twin with a Unity visualization frontend, and the learned policy is deployed as quantized scheduling tables on a Kintex-7 FPGA testbed. Under 90% offered load, the deployed policy reduces mean and P95 end-to-end delays by 10.6% and 12.8%, respectively, and reduces the synchronization-violation rate by 43.5% relative to a strict-priority baseline. Results show improved delay performance and temporal consistency for correlated industrial traffic.