A skill-enhanced retrieval-augmented generation expert system framework for offshore accident analysis: Combining fuzzy inference, Dempster-Shafer evidential reasoning, and expert-gated skill adaptation

Traditional rule-based expert systems for offshore safety assessment suffer from labour-intensive knowledge acquisition and an inability to evolve with operational experience. This paper presents a Skill-Enhanced Retrieval-Augmented Generation (RAG) Expert System Framework in which the “Skill” is a versioned, self-contained module encoding domain-specific fuzzy inference templates with expert-gated weight updates. Applied to FPSO–shuttle tanker collision risk during tandem offloading, the framework compresses 245 IF–THEN rules into 12 weighted templates across five Skills, achieving Pearson r = 0.987 and F1 = 0.944 for high-risk classification. One feedback cycle confirmed that the version-control and weight-update mechanism operates as intended, with five of six proposed adjustments, ranging up to 4.7%, approved and committed by the expert panel. This result is not a claim of superior risk-assessment accuracy relative to the established baseline. It is evidence that the knowledge-engineering architecture, grounded in maintainability, modularity, and traceability, can be put into practice and audited at every step. The contribution is architectural, not computational. The framework is positioned as a feasibility-oriented knowledge-engineering contribution: it demonstrates that Skill-structured RAG can support maintainable, modular, and traceable offshore safety knowledge bases, not that it outperforms established expert-assessment methods on general offshore risk problems.