An LLM-based Cross-Domain Fault Localization in Carrier Networks
Jiafu Ma, Sai Han, Guangquan Wang, Zelin Wang, Fengxia Fan, Shiwei Ye · 2024
The increasing complexity of carrier networks poses significant challenges for root cause analysis (RCA) and fault localization across diverse domains, including Optical Transport Networks (OTN) and IP Radio Access Networks (IPRAN). Traditional approaches, predominantly reliant on rule-based systems and manual expertise, are often inadequate to manage the scale and intricacies of modern network environments. In this paper, an intelligent method, CrossRCA, is presented to address cross-domain fault localization and RCA in carrier networks. Powered by large language models (LLMs), CrossRCA is designed to dynamically align incoming alerts with relevant diagnostic workflows, aggregate critical context-aware diagnostic information, predict incident root cause categories, and generate explanatory narratives to support network engineers. Pre-trained LLMs, augmented with prompt engineering and Retrieval Augmented Generation (RAG) techniques, are utilized to encapsulate domain-specific knowledge and enable sophisticated contextual reasoning across heterogeneous network domains. As a result, accurate and automated fault diagnosis is facilitated. Experimental evaluations conducted on real-world network fault datasets demonstrate significant improvements in diagnostic accuracy and reductions in troubleshooting time, particularly in complex cross-domain scenarios. The transformative potential of LLM-driven frameworks to advance autonomous network management and enhance operational efficiency in carrier networks is highlighted in this paper.