A causal resilience-based and topology-aware framework for root cause analysis and system hardening in graph-based critical infrastructures

Evaluating system safety in critical infrastructures requires identifying the root causes of failures and defining hardening strategies. Existing data-driven approaches rely on statistical correlation rather than causation and fail to isolate true causal drivers due to the scarcity of historical failure data. This study proposes a causal framework that integrates physical topology with causal inference to execute root cause analysis and guide infrastructure hardening. A hybrid structural learning approach constructs a causal graph by combining data-driven causal discovery with physical topological constraints. The methodology explicitly models latent variables to account for unobserved confounding, eliminating the restrictive assumption of causal sufficiency. The methodology employs counterfactual reasoning to derive causal-based metrics that evaluate system dynamics under hypothetical scenarios. The methodology is evaluated on cascading failures from the IEEE 123-bus power system, diverse urban infrastructure topologies, and a real large-scale power distribution grid subjected to low-frequency, high-impact extreme weather events. Results demonstrate that the proposed hybrid structural learning improves the root cause identification accuracy of several state-of-the-art causal discovery algorithms. In failure location tasks, the framework achieves 92.29% component-level and 98.32% measurement-level accuracy, significantly surpassing state-of-the-art baselines. Furthermore, the application of counterfactual resilience metrics for infrastructure hardening yields up to a 68.2% reduction in performance loss, surpassing standard feature importance and correlation-based investment strategies across all tested network topologies.