Cascading failure analysis of critical infrastructure in conflict zones: a large language model-driven knowledge graph framework

Modern geopolitical conflicts increasingly target interdependent critical infrastructures, where localized damage can trigger cascading failures, leading to widespread disruption of essential public services. Traditional intelligence safety assessment frameworks struggle to process the dynamic, unstructured data streams from conflict environments. To bridge this gap, this paper introduces a novel probabilistic framework for dynamic cascading failure analysis of Critical Infrastructure (CI) under attacks by integrating Large Language Model (LLM) and Knowledge Graph (KG). Our novel dual-graph architecture first employs an LLM to automatically construct a Dynamic Battlefield Event Knowledge Graph (BEKG) from real-time Open-Source Intelligence (OSINT), capturing tactical events as they unfold. This dynamic event layer then serves as evidence for a probabilistic inference process on a static Infrastructure Service Dependency Knowledge Graph (ISDKG), which is modeled as a Bayesian Network (BN). This integration enables the simulation of the propagation of cascading failures, translating initial physical damage into a quantifiable risk assessment of widespread service disruptions. Validated through a case study on Ukraine's infrastructure, our system demonstrates high fidelity in automated knowledge construction, provides novel insights into cascading impact analysis, and offers a powerful decision-support tool for assessing the systemic vulnerability and resilience of complex infrastructure networks in conflict zones.