From Causal Graphs to Problem Diagrams: Enhancing Safety Requirement Modeling with LLM

Safety requirements play a crucial role in ensuring the reliability of Cyber-Physical Systems. Causal graphs provide structured support for the generation and validation of safety requirements by revealing the logical links between system behaviours and potential threats. Traditional methods rely on manual experience to analyse failure causality and safety requirement causality, resulting in inefficiency and limited traceability. Although Problem Frames (PF) can effectively describe system requirements and boundaries, they lack explicit causal logic, hindering causality traceability. In this paper, we propose a causal enhancement modelling approach that fuses LLMs with fault analysis. Firstly, the causal graphs are generated through a combination of prompt engineering and manual inspection patterns, and problem diagrams are constructed accordingly. Secondly, the causal logic is embedded in the PF as a semantic layer, and the safety requirement propagation paths between the physical and information domains are explicitly proposed. Finally, a comparative experimental evaluation was conducted using a multi-rotor UAV GPS system as a case study to verify the usability of the proposed method. The results indicate that our method increased the average number of safety requirement domains by 4.47 and interactions between domains by 7.33. Our approach enables fault-driven safety requirements modeling, enhances PF’ semantic expressiveness and interpretability, and offers a semi-automated, reliable solution that bridges AI causal reasoning with formal modeling for complex systems.