Deviation-aware digital twin-enabled joint user association and task partitioning for industrial IoT offloading

Abstract

Latency-critical industrial Internet of Things (IIoT) applications outstrip on-device computing capability, yet digital twin (DT)-assisted multi-access edge computing (MEC) is typically modeled as a perfectly synchronized mirror and treats edge association and task partitioning separately. This paper proposes a deviation-aware three-layer DT-assisted offloading framework spanning IIoT devices, micro base stations (MBSs), and a macro base station (BS). Twins of devices and MBSs are maintained at the BS and explicitly parameterized by CPU-frequency, transmit-power, and bandwidth deviations, so decisions are made from realistically imperfect twin state and the resulting delay-estimation error is quantified. Under delay and energy constraints, discrete user association and continuous partitioning ratios are jointly optimized to minimize average offloading time plus a service-failure penalty, with distinct delay models derived for parallel independent subtasks and serially dependent subtasks. Because the problem is non-convex, NP-hard, and hybrid-variable, a deep multi-agent parameterized Q-network (DMAPQN) is developed in which each device twin acts as an agent and a global mixing network couples local hybrid-action Q-values, preserving the discrete–continuous action structure. Simulations show a 10 MB task completes in 330 ms versus 415, 596, 617, and 705 ms for dichotomy, random, full-MBS, and local execution—a 20.5% gain over the strongest baseline.