Traffic-Aware System-Level Modeling for 5G-Enabled Factory of the Future
Dhungana, Nabin (2026)
2026
Master's Programme in Computing Sciences and Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2026-05-28
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202605276440
https://urn.fi/URN:NBN:fi:tuni-202605276440
Tiivistelmä
Fifth Generation Mobile Network (5G)-enabled Factory of the Future (FoF) environments require communication networks to simultaneously support stochastic heterogeneous traffic with fundamentally different latency and reliability requirements, all competing for the same shared radio resources. Existing approaches address scheduling enhancements, resource allocation, or architectural design in isolation, without capturing bidirectional scheduling, resource contention, and heterogeneous stochastic traffic in a unified system-level framework.
To address the gap, a custom discrete-event simulator is developed for a 5G-enabled industrial environment. The simulator models a 5G Access Point (AP) serving heterogeneous Uplink (UL) traffic comprising high-bandwidth Automated Guided Vehicle (AGV) video streams and latency-critical AGV control signals as Class-1 traffic and periodic sensor data as Class-2 traffic, alongside their feedback Downlink (DL) traffic. The resource management is abstracted under a two-class Discriminatory Processor Sharing (DPS) scheduling policy. The system further incorporates a physical-layer model, a co-located edge server with object detection, and bi-directional scheduling to emulate a realistic FoF deployment. Job drop rate and End-to-End (E2E) cycle latency serve as the primary Key Performance Indicators (KPIs) to evaluate system performance.
The system is evaluated across three bandwidth configurations (40, 60, and 100 MHz) with progressively higher device counts. The results show that simultaneous KPI compliance for both traffic classes requires server utilization to remain below approximately 65%. Below this threshold, a clearly identifiable compliant weight range emerges; it widens further as utilization falls, providing tolerance for time varying load. Although average E2E latency targets are met across all configurations, per-payload AGV video latency compliance does not reach 100% even at the optimal weight and highest bandwidth, owing to large payload variability under stochastic mixed-traffic contention, a gap that bandwidth scaling reduces but does not close. Analysis further reveals that the DL direction contributes 8–15% of total E2E latency and that DL queue pressure grows non-linearly with scheduling weight, reinforcing the necessity of bidirectional modeling.
A 50 MHz configuration is used exclusively for the DPS versus network slicing comparison. The assessment demonstrates that shared scheduling provides superior multiplexing gain, maintaining simultaneous KPI compliance for both classes across a well-defined operating window that static slicing cannot replicate at any resource split. Finally, an Upper Confidence Bound (UCB)-based Multi-Armed Bandit (MAB) framework is applied to automate DPS weight selection for the most constrained configuration, converging on the optimal weight within 500 training rounds and achieving performance equivalent to that of the best manually identified weight. However, correlated arm rewards caused slow convergence, suggesting that more sample-efficient methods would be better suited to this problem.
To address the gap, a custom discrete-event simulator is developed for a 5G-enabled industrial environment. The simulator models a 5G Access Point (AP) serving heterogeneous Uplink (UL) traffic comprising high-bandwidth Automated Guided Vehicle (AGV) video streams and latency-critical AGV control signals as Class-1 traffic and periodic sensor data as Class-2 traffic, alongside their feedback Downlink (DL) traffic. The resource management is abstracted under a two-class Discriminatory Processor Sharing (DPS) scheduling policy. The system further incorporates a physical-layer model, a co-located edge server with object detection, and bi-directional scheduling to emulate a realistic FoF deployment. Job drop rate and End-to-End (E2E) cycle latency serve as the primary Key Performance Indicators (KPIs) to evaluate system performance.
The system is evaluated across three bandwidth configurations (40, 60, and 100 MHz) with progressively higher device counts. The results show that simultaneous KPI compliance for both traffic classes requires server utilization to remain below approximately 65%. Below this threshold, a clearly identifiable compliant weight range emerges; it widens further as utilization falls, providing tolerance for time varying load. Although average E2E latency targets are met across all configurations, per-payload AGV video latency compliance does not reach 100% even at the optimal weight and highest bandwidth, owing to large payload variability under stochastic mixed-traffic contention, a gap that bandwidth scaling reduces but does not close. Analysis further reveals that the DL direction contributes 8–15% of total E2E latency and that DL queue pressure grows non-linearly with scheduling weight, reinforcing the necessity of bidirectional modeling.
A 50 MHz configuration is used exclusively for the DPS versus network slicing comparison. The assessment demonstrates that shared scheduling provides superior multiplexing gain, maintaining simultaneous KPI compliance for both classes across a well-defined operating window that static slicing cannot replicate at any resource split. Finally, an Upper Confidence Bound (UCB)-based Multi-Armed Bandit (MAB) framework is applied to automate DPS weight selection for the most constrained configuration, converging on the optimal weight within 500 training rounds and achieving performance equivalent to that of the best manually identified weight. However, correlated arm rewards caused slow convergence, suggesting that more sample-efficient methods would be better suited to this problem.