|
Enhancing traffic steering in 5G open radio access network using machine learning |
|---|---|
| รหัสดีโอไอ | |
| Title | Enhancing traffic steering in 5G open radio access network using machine learning |
| Creator | Natthapol Sangkool |
| Contributor | Prapun Suksompong, Advisor |
| Publisher | Thammasat University |
| Publication Year | 2568 |
| Keyword | Open radio access network, Traffic steering, Machine learning, eMBB, URLLC, xApp |
| Abstract | The coexistence of enhanced Mobile Broadband (eMBB) and Ultra-Reliable Low-Latency Communication (URLLC) traffic poses a major challenge in 5G Open RAN (O-RAN) networks. Conventional resource allocation approaches, such as network slicing and puncturing-based scheduling, either introduce significant management overhead and inter-slice interference or inevitably degrade eMBB throughput to satisfy URLLC latency requirements. Both approaches also depend on advanced 5G features that remain immature or unavailable in current open-source implementations, leaving service-aware traffic separation for eMBB and URLLC on real hardware an unsolved practical problem.This thesis proposes a practical, slicing-free and puncturing-free traffic steering framework that exploits the programmability of the O-RAN architecture. A machine learning-based UE classification system is developed and deployed as a near-real-time RIC xApp, identifying eMBB and URLLC UEs in real time from standard E2SM-KPM metrics alone, without requiring explicit UE signaling or deep packet inspection. Upon classification, intelligent traffic steering via conditional handover directs each UE to a pre-configured cell optimized for its service profile: a downlink-heavy TDD cell for eMBB and a low-latency, uplink-balanced cell for URLLC.The framework is implemented and evaluated on a physical O-RAN testbed built on srsRAN, Open5GS, and low-cost SDR hardware. To ensure measurement integrity and avoid cross-link interference between co-channel cells operating under different TDD slot configurations, eMBB and URLLC traffic types were evaluated sequentially. Four classifiers — SVM, KNN, XGBoost, and LSTM — all achieved offline classification accuracy exceeding 98, with all four models exceeding 99.94 accuracy and near-perfect URLLC recall under live xApp deployment. System-level evaluation against a random attachment baseline demonstrated median downlink throughput improvements of 27–38 for eMBB UEs, while URLLC P99 delay was reduced to below 1 ms across all models, closely approaching the ideal correct-attachment upper bound. These results confirm that the proposed framework delivers a deployable, low-complexity solution for service-aware eMBB/URLLC traffic separation in practical 5G O-RAN deployments. |
