Fault Root Cause Analysis via Adaptive Chain of Graph Reasoning in Wireless Networks

The increasing topological complexity of modern wireless networks has exposed significant limitations in conventional rule-based root cause analysis (RCA) methodologies. These approaches demonstrate inadequate adaptability to the dynamic interdependencies inherent in 5G and beyond network architectures, frequently resulting in extended service disruptions and degraded performance metrics. While large language models (LLMs) exhibit promising reasoning capabilities, their direct application to RCA is constrained by the unstructured representation of network data and the requirement for multi-hop graph-based inference. To address these challenges, we introduce Adaptive Chain of Graph (ACoG) Reasoning, an inference-time framework that structures RCA as a dynamic, LLM-guided graph traversal process without necessitating LLM fine-tuning. ACoG transforms network topology and fault propagation patterns into structured knowledge graphs, then iteratively refines root cause identification through a closed-loop process of graph-to-text conversion, LLM-based reasoning, and adaptive graph modification. Complementing this framework, we implement a lightweight reinforcement learning module that optimizes path selection strategy, enhancing search efficiency while maintaining minimal computational over-head. Comprehensive evaluation on real-world wireless network datasets demonstrates that our LLM-training-free approach achieves superior accuracy and robustness, particularly in complex multi-fault scenarios, while significantly improving interpretability of the reasoning process.