Adapting Network Information Into Semantics for Generalizable and Plug-and-Play Multi-Scenario Network Diagnosis

Abstract

Leveraging large language models (LLMs) for fault diagnosis in communication networks is considered a promising direction. However, communication-network faults often manifest as cross-layer degradations and are reflected in heterogeneous observations, making them difficult to directly represent for LLM reasoning. This makes it challenging to express network information in a form that can be effectively understood by LLMs. To address this issue, we propose an LLM-based Network Semantic Generation (LNSG) algorithm, which integrates semanticization and symbolization methods to uniformly represent multimodal network information. Based on LNSG and LLMs, we further present NetSemantic, a plug-and-play and data-light semantic fault diagnosis framework. It enables rapid adaptation to diverse network environments and provides efficient fault diagnosis capabilities. Experimental results show that, without task-specific supervised training or model fine-tuning, NetSemantic achieves 96.10% anomaly detection accuracy and 89.49% fault diagnosis accuracy on the NS-3-based network fault dataset, and further achieves 90.36% overall diagnosis accuracy on a real-world PCIe fault dataset.