PhD Oral Exam - Nguyen Phuc Tran, Computer Science

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

The evolution of mobile networks toward 5G and Beyond 5G (B5G) including network slicing and Open RAN (O-RAN) architectures, requires a shift from reactive operation toward proactive and increasingly autonomous network management. A central challenge is to maintain stringent Service Level Agreements (SLAs) while managing complex root cause analysis (RCA) processes and reducing the energy footprint of both network infrastructure and the AI models used to manage it.

This thesis develops and evaluates complementary components of a reference architecture for cognitive and sustainable communication-network management. The architecture organizes the contributions into proactive service assurance, automated root cause analysis, sustainable AI, and energy-aware infrastructure optimization, while distinguishing experimentally demonstrated interfaces from proposed interfaces and future end-to-end integration.

At the proactive-automation layer, the thesis develops LSTM-FSD for short-term traffic forecasting and LP-KPI for estimating sensitive Key Performance Indicators (KPIs). These models address the dynamic behaviour of network slices and provide data-driven visibility into future network states.

At the orchestration layer, a proactive closed-loop algorithm for network-slice assurance (PCLANSA) uses the predicted traffic and KPI estimates to dynamically manage virtual resources and link capacities. In the evaluated 24-hour 5G simulation, the first parameter setting prevents KPI violations across the four heterogeneous slices. For the eMBB slice, it reduces mean resource consumption by 54.85% relative to a static peak-provisioned configuration without closed-loop adaptation.

At the diagnostic layer, the thesis introduces TelcoInsight and the Fault Diagnostic Analytics Service (FDAS). TelcoInsight combines Large Language Models (LLMs) with Retrieval-Augmented Generation (RAG) to construct an RCA knowledge base from unstructured historical support tickets. FDAS defines a standards-oriented service and versioned-artifact contract through which ticket-derived knowledge can support ranked diagnostic results. This interface is proposed and characterized offline; held-out end-to-end diagnostic evaluation and implementation in the 5G Core remain future work.

To reduce the environmental footprint of the intelligent components, an evaluation pipeline examines the energy-performance trade-off associated with LLM quantization and pruning. On the evaluated hardware and workloads, the 16-bit pruning configuration achieves energy savings of up to 40.42% relative to the 32-bit configuration; the corresponding performance effects are reported for the downstream telecommunications tasks.

Finally, a mixed-integer linear programming (MILP) model, named EJPM, and a deterministic k-means- based heuristic, named H_EJPM, are presented for energy-aware joint placement and migration in an O-RAN edge cloud. Over the evaluated 24-hour workload, the theoretical Multi-CU relaxation reduces modeled MILP energy by 5.7% relative to Single-CU. For the feasible Multi-CU solutions included in the comparison, H_EJPM remains within approximately 9.7% of the MILP while respecting the modeled resource-capacity and one-way F1-U delay constraints.

The demonstrated integration in this thesis connects traffic prediction and KPI estimation to proactive network-slice assurance. The diagnostic service contract, sustainable-LLM pipeline, and O-RAN placement-and-migration models complement that control path within the reference architecture. Together, the evaluated components and proposed interfaces provide a technical basis for future AI-native and energy-conscious network-management systems, with integrated cross-component evaluation remaining future work.