KG-VLM: visual analytics for network fault diagnosis via knowledge graph-enhanced vision-language model

Network fault diagnosis in modern telecommunications systems requires rapid and accurate root cause identification while maintaining interpretability for operational deployment. Traditional rule-based approaches lack adaptability, while pure deep learning methods suffer from high computational costs and limited explainability. We propose KG-VLM, a visual analytics framework that synergistically integrates knowledge graphs with lightweight vision-language models for interpretable network fault diagnosis. Our approach leverages a knowledge graph to encode domain-specific causal relationships, enabling rapid candidate filtering through structured reasoning. A graph visualization module transforms fault propagation patterns into spatial visual representations that capture topological structures, thereby facilitating crossmodal understanding. A lightweight vision-language model then performs semantic reasoning over visual graph patterns and textual alarm contexts, achieving deep understanding without computational overhead. Experiments on the Telstra Network Disruptions dataset demonstrate that KG-VLM achieves 0.7234 F1-score, outperforming the best baseline by 4.0 percentage points while maintaining 8.4ms inference latency suitable for edge deployment. Ablation studies validate that the visualization module contributes significantly by encoding spatial relationships that enhance visual reasoning. Our framework demonstrates that principled integration of symbolic knowledge, visual analytics, and neural representation learning provides an effective solution for operational network management.