Source-linked AI summary
Digital Twin for Pre-Deployment Validation of AI-Driven Safety-Critical Industrial Edge Control Loops
Sara Cavallero, Federico Tonini, Tony Chahoud, Giampaolo Cuozzo, Davide Borsatti, Walter Cerroni, Maurizio Fodrini, Riccardo Marini
TL;DR
Existing digital twins often separate process and network models, although industrial closed-loop performance also depends on application dynamics. This paper proposes an integrated modular twin and validates it against a real 5G-connected AMR deployment; predictions closely match measurements, and placement analysis quantifies feasible timing margins.
Problem
Existing digital twins rarely jointly model application dynamics and network behavior, although industrial closed-loop performance depends on processes, latency, data quality, and application decisions.
Source: §I · p. 1Method
The paper combines Process and Network DTs with an AI-oriented Decision Module, validates the framework in a real industrial deployment, and studies application placement.
Source: §I · pp. 1–2, §VIII · p. 9Results
RTT prediction error averaged 5%, end-to-end communication latency error 2%, and RSRP maximum absolute deviation 2 dB; feasible uncertainty margins were 1.2 on-premises, 1.08 edge, and 0.55 remote cloud.
Source: §VIII · p. 9Takeaways & Limitations
The results indicate that integrated industrial digital twins can support what-if analysis, proactive system design, and latency-focused deployment decisions.
Source: §VIII · p. 9The deployment analysis focuses on latency and control-loop feasibility; scalability, operational costs, resource availability, and resiliency are outside the study’s scope.
Source: §VIII · p. 10
Abstract
from arXiv · showhide
Industrial environments are increasingly characterized by the tight interaction among physical processes, communication infrastructures, and intelligent applications. In this context, Digital Twins (DTs) have emerged as a key technology for system analysis and optimization. However, existing DT solutions typically focus either on industrial processes or communication networks, while lacking an integrated and application-aware perspective. To fill this gap, this paper proposes a modular DT framework for industrial environments that jointly models physical processes, wireless communications, and application logic within a unified architecture. The feasibility of the proposed framework is experimentally validated through a real-world Proof-of-Concept (PoC) implemented in the BI-REX pilot line, involving a 5G-connected Autonomous Mobile Robot (AMR) transporting hazardous liquids and remotely controlled by an AI-driven application. The proposed DT is used to reproduce the behaviour of the real deployment and to investigate the impact of different placements of the AI application, including on-premise, edge, and remote cloud execution scenarios. Experimental results demonstrate a close agreement between DT predictions and PoC measurements in terms of both network-level metrics, such as Reference Signal Received Power (RSRP) and latency, and end- to-end application metrics, including application-level Round- Trip-Time (RTT). Moreover, the analysis shows how inaccuracies of network modeling can critically affect the feasibility of latency-sensitive industrial control loops, highlighting the potential of integrated DTs as tools for the pre-deployment design and validation of next-generation industrial systems.
I. INTRODUCTION
Industrial digital twins often model processes or networks separately, although closed-loop performance also depends on application dynamics. This paper proposes and experimentally validates a modular framework that jointly represents those elements and studies AI-application placement.
I. INTRODUCTION
- Industrial systems combine heterogeneous devices and data streams, while subsystem disturbances and dynamic wireless conditions can affect safety, production, and communications.These systems include sensors, actuators, controllers, robots, mobile vehicles, edge nodes, and cloud services.Source: §I · p. 1
- Existing digital twins rarely integrate application dynamics with network behavior, despite closed-loop performance depending on processes, communication latency, data quality, and application decisions.Industrial testing is costly, time-consuming, potentially disruptive, and cannot feasibly cover every operating condition.Source: §I · p. 1
- The proposed modular framework combines a Process DT, Network DT, and Decision Module to model processes, wireless communications, and application logic.The Decision Module reproduces application logic and decision-making processes.Source: §I · p. 2
- The BI-REX proof of concept uses a 5G-connected AMR carrying hazardous liquids and remote AI predictions to trigger clamp closure when a spill risk is detected.The use case represents a closed-loop system whose performance depends on physical processes, communications, and application intelligence.Source: §I · p. 2
- The framework is experimentally validated by comparing digital-twin predictions with deployment measurements for application-level RTT and network-level RSRP and communication latency.The contribution description identifies both end-to-end application and network metrics for the comparison.Source: §I · p. 2
- The deployment study provides quantitative placement guidelines for latency-sensitive applications based on the accuracy of network and application models.The paper describes this experimentally validated, jointly modeled framework as one of the first of its kind.Source: §I · p. 2
II. RELATED WORK
The related-work discussion identifies process-focused and network-focused twins as largely separate, while the proposed architecture integrates process, network, and application models. Its shared infrastructure connects the industrial scenario with the twin modules through two-way data and service flows.
II. RELATED WORK
- Prior industrial digital twins often focus on process or network operations, while this framework combines process, network, and application models in a modular closed-loop architecture.The framework is experimentally validated in a real-world industrial proof of concept.Source: §II · p. 2
III. PROPOSED DIGITAL TWIN FRAMEWORK FOR INDUSTRIAL ENVIRONMENTS
- A shared data infrastructure connects the industrial scenario to Process, Network, and Decision Modules, importing real measurements and distributing twin services and synthetic data.The framework section introduces these components as jointly modeling the physical process, communication infrastructure, and application logic.Source: §III · pp. 2–3, Figure 1 · p. 3
A. Architecture Overview
The framework links the industrial scenario and three digital-twin modules through shared data and iterative information exchange. Process and network models can be simulated and combined, with real-world measurements supporting synthetic-data generation and calibration.
A. Architecture Overview
- The architecture pairs physical process, assets, and network with Process, Network, and Decision Modules, using shared infrastructure for real measurements, twin services, and iterative design.Information from one module can inform the others when evaluating system performance under alternative deployment configurations.Source: A. Architecture Overview · p. 3
B. Process Digital Twin Module
- The Process DT models the industrial environment, assets, and process dynamics, simulates their evolution, and generates synthetic data calibrated with real-world measurements.Its simulation supports realistic and reproducible operating conditions for analysis, validation, and testing.Source: B. Process Digital Twin Module · p. 3
C. Network Digital Twin Module
- The Network DT models radio propagation, communication technologies, and protocols, then simulates network configurations and combines network information with the Process DT.The co-simulation captures mutual influence between process and communication dynamics under different operating conditions.Source: C. Network Digital Twin Module · p. 3
D. Decision Module
The Decision Module supplies predictive and prescriptive application logic, while the industrial case study tests a 5G-connected mobile robot transporting hazardous liquids. Its sensing-to-inference-to-command path supports spill-prevention actions.
D. Decision Module
- The Decision Module uses real and synthetic data to train and run machine-learning models that forecast system conditions and support proactive decisions.It can host machine-learning, statistical, rule-based, or other application-level logic.Source: D. Decision Module · p. 3
IV. INDUSTRIAL CASE-STUDY
- The BI-REX case study addresses spill risk when an AMR transports hazardous liquids and abrupt braking can spill containers, threatening nearby operators.The proof of concept combines deep learning and 5G to address this safety issue.Source: §IV · pp. 3–4, A. Block Scheme · p. 4
A. Block Scheme
- A motion sensor sends acceleration and angular-velocity data over an OAI-based 5G standalone network to a local server, which runs spill prediction and can command gripper closure.The sensor reports periodically, and the downlink command is sent when the algorithm predicts a potential spill.Source: A. Block Scheme · p. 4
B. Deep Learning Design
The paper frames spill prediction as a binary Fault/Non-Fault classification over MIR motion time series. A sliding-window CNN predicts Fault before a potential spill, with an example showing 400 ms of anticipation.
B. Deep Learning Design
- A Fault is an obstructed path with potential for a spill; the spill sample and its A preceding samples are labeled Fault, and earlier samples Non-Fault.Source: B. Deep Learning Design · p. 4
- Each time-series row records timestamped acceleration, z-axis angular velocity, position, and orientation; consecutive rows are separated by sampling period TS.Source: B. Deep Learning Design · p. 4
- The CNN processes preprocessed, normalized windows of W rows that shift over time using a sliding-window mechanism.Source: B. Deep Learning Design · p. 4
- The CNN predicts a spill-related Fault 400 ms before the event in the real-world inference example.Source: B. Deep Learning Design · p. 4
C. Working Condition
The working condition is whether the robot can complete gripper closure within the available time before a spill. The PoC meets this timing constraint, while the delay model provides a bound for evaluating application-server placement under uncertainty.
C. Working Condition
- The 5G communication flow includes uplink sensor data and downlink actuation commands, with TCP endpoint overhead included in the delay model.Source: Figure 4 · p. 5, C. Working Condition · p. 5
- The model separates the application timing requirement TI, 5G communication delay TC, server preprocessing TO, CNN inference TE, and gripper actuation delay TA.Source: C. Working Condition · p. 5
- In the PoC, empirically estimated TI=200 ms exceeds measured TC+TO+TE=66 ms plus TA=50 ms (116 ms total), meeting the timing constraint; predicting the spill and activating the gripper prevents bottle fall.Source: C. Working Condition · p. 5, Figure 5 · p. 5
- The PoC uses a dedicated server, and the study considers on-premises, edge-cloud, and remote-cloud CNN placement because TT depends on application location.Source: C. Working Condition · p. 6
- The placement bound limits TT, the CN-to-application-server delay; a positive uncertainty margin reduces the allowable delay to account for network and application variability.Source: C. Working Condition · p. 6
V. DIGITAL TWIN IMPLEMENTATION
The implementation combines a process replica, a simulated 5G network, and an AI decision module in a modular message-broker architecture. Their interaction supports controlled application validation and comparison of digital predictions with PoC measurements.
V. DIGITAL TWIN IMPLEMENTATION
- Independent services publish and consume data through RabbitMQ, decoupling producers such as sensors and network interfaces from aggregation, inference, and DT consumers.Source: §V · p. 6
- The Process DT reconstructs the BI-REX line and MIR motion, then generates labeled synthetic time series across controlled trajectories, speeds, and environmental configurations.Source: §V · p. 6
- The Network DT combines ray-traced, time-varying RSRP predictions with a 3GPP-compliant network simulator for scheduling and message exchanges.Source: §V · pp. 6–7
- The Decision Module trains and validates the 1D-CNN on Process-DT synthetic data, enabling repeatable measurement of preprocessing and inference delays and what-if deployment analyses.Source: §V · p. 7
A. Round Trip Time
RTT is measured per fault event from the first uplink sample that triggers a Fault prediction to receipt of the corresponding gripper command. The paper averages this measure across blockage events and obtains delay terms differently in the PoC and DT.
A. Round Trip Time
- Per-event RTT runs from acquisition of the first uplink motion sample triggering a Fault prediction to onboard receipt of the corresponding gripper-closing command.Source: A. Round Trip Time · p. 7
- The overall RTT metric is the average across the blockage events considered in the analysis.Source: A. Round Trip Time · p. 7
- The PoC measures TC, TO, and TE experimentally, whereas the DT estimates them through simulation campaigns.Source: A. Round Trip Time · p. 7
B. Reference Signal Received Power
The section defines blockage-event RSRP and explains how the PoC and DT obtain it. Their KPI comparison covers two blockage positions.
B. Reference Signal Received Power
- RSRP for blockage event i is the received power measured by the UE during that event interval.The DT synthesizes it by summing the powers of K propagation paths, whose channel gains reflect the industrial environment’s geometry and materials.Source: B. Reference Signal Received Power · p. 7
- The PoC polls the UE in real time using standardized AT commands, while the DT computes event RSRP with Sionna using the deployment’s UE–gNB coordinates, frequency, and bandwidth.Source: B. Reference Signal Received Power · p. 7
- Table III compares PoC and DT KPIs at two blockage positions; it reports DT results as mean ± standard deviation over 50 runs and PoC results over 10 runs.Blockage #1 is near the gNB and Blockage #2 is farther away.Source: Table III · p. 8, Figure 7 · p. 8
VII. RESULTS
The DT predictions closely match PoC measurements for end-to-end delay and capture the RSRP change between blockage positions, with an offset in absolute RSRP. The placement analysis shows that tighter timing requirements and larger uncertainty margins reduce the available cloud options.
VII. RESULTS
- At both blockage locations, DT-predicted average RTT differs from PoC measurements by less than 4 ms, and every experimentally measured RTT falls within DT-predicted variability (about 6 ms standard deviation).Source: §VII · p. 8
- Approximately 2 dBm: the DT’s systematic RSRP offset from the PoC; both capture the signal decrease from the nearer to the farther blockage position.The relative change is preserved despite the absolute-value discrepancy.Source: §VII · p. 8
- The results validate the DT’s capability to reproduce the industrial process, 5G network behaviour, and application logic in the case study.Source: §VII · p. 9
- M > 0.55 makes remote-cloud deployment infeasible; M > 1.08 leaves only on-premises deployment, which allows a maximum margin of M = 1.2.At M = 0 and TI = 200 ms, the available TT budget is about 82 ms, allowing any of the considered servers.Source: §VII · p. 9
- At TI = 160 ms, remote servers cannot host the application and on-premises allows M = 0.6; at TI = 120 ms, only on-premises is suitable, with almost no uncertainty margin.Source: §VII · p. 9
VIII. CONCLUSIONS
The integrated Digital Twin closely reproduces real-deployment metrics and supports application-placement analysis under delay uncertainty, within a study focused on latency-related feasibility.
VIII. CONCLUSIONS
- The framework integrates Process DT, Network DT, and AI-oriented Decision Module components and was validated in a BI-REX PoC with a 5G-connected robot transporting hazardous liquids.An AI-driven application predicts dangerous manoeuvres and can proactively trigger the onboard gripper’s safety closure.Source: §VIII · p. 9
- RTT prediction error averaged 5%, end-to-end communication latency error 2%, and RSRP maximum absolute deviation 2 dB, indicating close field-trial agreement.The framework reproduces both network-level and application-level behavior.Source: §VIII · p. 9
- Before application timing requirements are violated, uncertainty margins reach 1.2 on-premises, 1.08 at the edge, and 0.55 in the remote cloud.The deployment analysis compares margins and application requirements against reference delays for cloud solutions.Source: §VIII · p. 9, Figure 8 · p. 9
- The results position integrated industrial DTs as tools for accurate system replication, what-if analysis, proactive design, and deployment optimization.The deployment study examines latency-related control-loop feasibility and resilience to uncertainty.Source: §VIII · p. 9
- The deployment analysis covers latency-related control-loop feasibility, leaving scalability, operational costs, resource availability, and resiliency outside this study’s scope.Source: §VIII · p. 10