Process-Aware Cross-Layer Adaptation for O-RAN-Enabled Industrial Systems

Abstract: Wireless networks increasingly support closed-loop industrial applications in which sensed data must be transmitted, processed, and converted into an action before physical process makes the result obsolete. Throughput, latency, and inference accuracy measured separately cannot determine whether such an application completed a useful task. We propose a Process-Aware Co-adaptation Engine framework that combines application outcomes, process state, radio telemetry, edge-compute state, and sensing configuration to select coordinated operating points across the complete loop. We evaluate the proposed approach in a factory-inspection case study that integrates a physics-based digital twin, a programmable 5G O-RAN network, and edge-based visual inference. The experiments show that the preferred resource allocation changes with production speed and that adapting individual system components independently can be inefficient. We further show that efficient configurations can be identified with relatively few full-system evaluations.

• The paper introduces PACE, a process-aware co-adaptation engine that jointly selects radio resources, edge CPU, sensing resolution, and process settings using correct-on-time rate, production goodput, and resource efficiency.
• PACE achieved efficiency of 0.96 at 0.5 m/s versus approximately 0.027 for a fixed high-resource policy, while joint adaptation consistently outperformed radio-, CPU-, or resolution-only strategies, which never exceeded 0.24.
• The evaluation shows that tree-ensemble surrogate search reached about 90% of exhaustive-search performance with roughly one-tenth as many evaluations, but safe deployment still requires drift detection, uncertainty calibration, cross-timescale coordination, and richer testbeds.

The paper proposes PACE, a Process-Aware Co-Adaptation Engine that coordinates sensing, radio, edge-compute, and process controls in wireless closed-loop industrial applications, evaluated end-to-end on a factory-inspection testbed (2608.13372).

Motivation and problem statement

The central observation is that in closed-loop applications—inspection, robotic control, remote operation—a correct decision has no value unless it arrives before the physical process makes it obsolete. Component metrics (throughput, average latency, inference accuracy) each diagnose local degradation but none establishes that a correct decision reached the actuator in time. The physical process itself sets the deadline: doubling conveyor speed halves the capture-to-sort window without any change to the radio channel, edge server, or inference model. Consequently, the preferred cross-layer operating point shifts with process state, and the binding bottleneck migrates among uplink transfer, inference, and sensing quality.

The authors distinguish co-design—jointly selecting cross-layer parameters for one operating condition—from co-adaptation, which updates the operating point at runtime as process and infrastructure states evolve. They argue co-adaptation is necessary because the active bottleneck moves across sensing, radio access, edge computing, and the process itself, and because O-RAN's near-real-time RIC and E2 interface provide a programmable actuation path while the application objective must sit above the RAN (the RAN cannot observe inference correctness, edge saturation, sensing configuration, or the process deadline).

Application-level metrics

The paper defines three metrics computed from per-cycle application logs (ground truth, returned class, capture/return timestamps, process-derived deadline):

• Correct-on-time rate: the fraction of cycles in which the classification is both correct and returned before the part reaches the sorting point. A timely incorrect decision and a correct late decision are both failures.
• Production goodput: correct-on-time parts per minute, which prevents the correct-on-time rate from favoring slow conveyors with generous deadlines.
• Production efficiency: $$\eta = \text{goodput}/(\text{uplink PRB share} \times \text{CPU cores})$$, a multiplicative denominator that penalizes configurations reserving large radio and compute allocations simultaneously. The authors note this is a relative measure within the measured platform and that other deployments could meter energy, cost, or accelerator time instead.

Efficiency is explicitly subordinate to hard constraints: the controller enforces a minimum correct-on-time rate before comparing feasible configurations, with fallback allocation or process-rate reduction when no candidate qualifies.

Architecture

PACE observes four signal groups—radio state, edge state, sensing/process state, and application outcomes—and constructs a cross-layer operating state. Per-stage timing separates communication delay from inference delay; process state determines remaining slack; outcomes distinguish radio congestion, compute saturation, sensing limitations, and deadline contraction. Each candidate operating point specifies uplink PRB share, edge CPU quota, image resolution, and an allowed process setting; the selector rejects constraint-violating candidates before ranking by efficiency, preventing locally reasonable but outcome-irrelevant actions (e.g., adding PRBs when inference already consumes the deadline).

Selection can use a measured operating map, a surrogate model with uncertainty estimates, or both, enabling informative trials rather than full sweeps. Actions are dispatched hierarchically: radio changes within the near-RT RIC window, edge quotas over seconds, sensing settings after sustained conditions, conveyor speed only at batch boundaries with production/safety approval. A digital twin (Isaac Sim) supports offline exploration of disruptive configurations and online rehearsal, with the authors emphasizing that live feedback remains necessary because twin predictions can err.

Case study and numerical results

The testbed integrates Isaac Sim for the physics-based process, an OpenAirInterface 5G NR stack with 106 PRBs and FlexRIC for telemetry and PRB actuation, and a fixed YOLO-based detector trained on GC10-DET for edge inference. Every configuration is measured under a persistent 30 Mbit/s competing uplink flow. The design space—10–90% PRB share, 1–8 CPU cores, three resolutions, four conveyor speeds (0.5–2.0 m/s)—yields 960 full-system operating points.

Three results stand out:

• The efficient allocation shifts with process speed. At 1280×720 resolution, the most efficient measured configuration at 1.0 m/s uses 35% uplink and 2 cores; at 1.5 m/s it moves to 75% uplink and 3 cores, with radio, edge platform, detector, and resolution unchanged. Static provisioning for one speed either wastes resources or misses deadlines at the other.
• Joint selection dominates single-axis adaptation. The fixed high-resource policy (90% PRB, 8 cores, 720p) achieves efficiency ≈0.027 at 0.5 m/s, versus 0.96 for joint selection—a roughly 35× gap. Policies adapting only the radio slice, CPU quota, or resolution never exceed 0.24 in the sweep. Joint selection reaches 0.96, 0.72, 0.54, and 0.40 at 0.5–2.0 m/s; the decline reflects the shrinking set of configurations that meet tighter deadlines. The minimum-resource policy matches the joint optimum at 1.0 m/s (0.72) but falls to zero at 1.5 m/s once its deadline margin is exhausted.
• Few full-system evaluations suffice. In a replay over the measured map, TPE reaches only ≈0.76 after 200 evaluations, whereas bootstrapped tree-ensemble UCB and Thompson-style sampling reach ≈0.90 within ~100 evaluations and 0.92–0.94 by 200, against an exhaustive best of 0.96—about 90% of exhaustive performance after roughly one tenth of the 960-configuration space.

Limitations and open challenges

The authors are explicit that the case study isolates targeted couplings rather than deployment disturbances: the measured deadline ends at decision return because robotic actuation time was not repeatable enough to include; background traffic, channel conditions, edge workload, detector weights, geometry, and downlink allocation are all fixed; and the numerical optima characterize this specific platform. The search replay additionally assumes stationary outcomes per configuration, so deployment must add context-shift detection, uncertainty recalibration, and surrogate updates.

The paper identifies three open challenges. Cross-time-scale coordination: radio, edge, sensing, and process controllers must share a bottleneck estimate and account for actions in progress, with hysteresis, dwell times, and bounded steps to prevent oscillation; stability analysis under delayed telemetry remains open. Safe adaptation: the operating map does not cover fading, interference, traffic bursts, or model drift; exploration must respect unequal reversibility of controls (PRB/CPU changes are reversible; resolution changes affect quality; conveyor changes disrupt production planning), and twin-guided actions must be gated by calibrated prediction error. Monitoring and evaluation: trustworthy per-cycle traces across administrative boundaries, telemetry age awareness, richer resource denominators, and benchmarks that report deadline misses and production disruption incurred while adapting—not just the final operating point—are all lacking.

Conclusion

PACE demonstrates, on a full-stack O-RAN testbed, that process state must enter cross-layer resource selection: the efficient operating point moves with conveyor speed, single-domain adaptation is measurably inadequate (≤0.24 versus 0.96 efficiency), and surrogate-based search identifies near-optimal configurations with roughly an order of magnitude fewer evaluations than exhaustive search. The framework's practical value depends on solving the coordination, safety, and monitoring problems the authors leave open—particularly maintaining correct-on-time performance while the operating environment, and the model of it, drift.