Conflict-Resistant Graph-Based Modeling of Technical System Risks from Weakly Structured Information for Decision Support

This paper develops and validates a conflict-resistant method for scenario-based risk assessment of technical systems using weakly structured event reports. The proposed approach integrates an event-oriented directed graph, the Dempster-Shafer theory of evidence, and monitoring of temporal non-stationarity of graph edges. The method jointly estimates the risk interval, conflict between evidence sources, and the amount of ignorance, thereby improving the transparency of expert interpretation and reducing the likelihood of erroneous scenario prioritization. The scientific contribution lies in a coherent use of these dimensions within an inference procedure in which uncertainty is kept explicit and treated as a quantitative indicator of data quality. Empirical validation is performed on data from the U.S. Nuclear Regulatory Commission. A total of 1,474 event records for 2002-2025 are analyzed. During preprocessing, terminology is normalized, severity classes are harmonized, and entities forming cause-and-effect transitions are extracted. The resulting conflict graph includes 624 vertices and 804 directed arcs. For priority scenarios, interval risk estimates are derived using belief and plausibility measures. For arcs, indicators of accumulated conflict and structural drift are computed in 180-day sliding windows. The results show that frequent evidence inconsistency concentrates in transitions to severe consequences, whereas drift dynamics exhibit a clustered pattern with alternating stable and transitional phases. Practical relevance is associated with a reproducible decision-support procedure. Across all windows, 151 drifting windows out of 1,404 (10.75%) are identified, and 44 non-stationary links out of 208 (21.15%) are detected. Risk intervals support operational prioritization of actions, conflict measures support source verification, and drift indicators support routine updating of the safety model. The method can be adapted to other high-risk engineering domains characterized by incomplete and heterogeneous data.