From Barrier Assurance to Cumulative Risk: A Knowledge Graph-Based AI Framework for Major Accident Hazard Risk Management

Abstract

Abstract Major Accident Hazard (MAH) risk in oil and gas assets rarely arises from the failure of a single component, equipment item or barrier. Instead, risk emerges cumulatively when maintenance deferrals, operational constraints, inspection findings, integrity anomalies, temporary repairs and project activities interact with degraded asset conditions and weakened safety barriers across complex systems (Reason, 1997, Kletz, 2009). Conventional work prioritization methods typically evaluate these activities using static, discipline-specific or equipment-level criteria. Although effective for local decision-making, such methods can obscure the aggregate effect of multiple individually tolerable impairments on a shared MAH scenario. Prior work at a GCC onshore COMAH site demonstrated that artificial intelligence (AI), retrieval-augmented generation (RAG) and large language models (LLMs) can automate identification and verification of Safety Critical Elements (SCEs), generate performance standards, verify assurance tasks in the computerized maintenance management system (CMMS), and reduce engineering effort by more than 90% (Stewart et al., 2025). This paper extends that framework from static barrier assurance to dynamic cumulative risk prioritization. The proposed method combines a Semantic Asset Knowledge Graph with a GraphRAG retrieval layer (Edge et al., 2024) and an agentic AI risk reasoning engine. Structured asset, maintenance, integrity and operational datasets are integrated with unstructured safety cases, bow tie studies, performance standards, inspection reports and maintenance data. These datasets are normalized and linked to represent physical, functional, safety-critical and causal dependencies between equipment, barriers, protection layers and MAH escalation pathways. The architecture enables specialised AI agents to identify affected barriers, assess degradation states, detect common dependencies, evaluate escalation pathways and rank competing work. In the current implementation, barrier mapping and cumulative risk clustering were demonstrated, with additional agent capabilities defined for future deployment. Application of the method demonstrated improved visibility of cumulative risk clusters and demonstrated the use of graph-based reasoning to support MAH-focused prioritisation. The system identified clusters of cumulative risk in which individually low-priority work items combined to create credible MAH precursors, consistent with the Swiss Cheese model of accident causation. The approach reduced manual engineering screening effort by approximately 40% and identified more than USD 100 million of production value at risk associated with suboptimal prioritization. The work demonstrates a practical and scalable method for moving from barrier assurance to dynamic, asset-level MAH risk management.