Grounded expert systems for offshore safety: enhancing rule-based risk assessment with retrieval-augmented generation (RAG) for auditable explanations

Ren, J ORCID: 0000-0003-4640-824X, Jenkinson, I and Tobora, O Grounded expert systems for offshore safety: enhancing rule-based risk assessment with retrieval-augmented generation (RAG) for auditable explanations. In: ICET 2026 . (International Conference on Electronics Technology (ICET), Chengdu).

Text ICET 2026 May final.pdf - Accepted Version Access Restricted until 1 June 2026. Download (554kB)

Abstract

This paper proposes a novel framework for enhancing traditional rule-based expert systems in offshore safety with a Retrieval-Augmented Generation (RAG) layer to provide dynamic, auditable, and human-readable explanations. Operations in the offshore industry, such as tandem loading between FPSOs and shuttle tankers, are complex and high-risk, often managed by expert systems built on extensive rule bases. While effective, these systems can lack transparency and become difficult to audit as they scale. Our approach maintains the deterministic logic of the original expert system as the single source of truth for decision-making while leveraging a Large Language Model (LLM) to generate grounded, narrative justifications for each decision. We present a two-phase methodology: an offline process to convert a technical rule set into an indexed, natural language knowledge base, and a real-time process that uses metadata filtering to retrieve the exact triggered rule and generate a formal explanation. This "Grounded Expert System" transforms an opaque decision-making tool into a transparent, auditable, and maintainable knowledge asset, addressing a critical need for explainability in safety-critical domains.

Item Type:	Conference or Workshop Item (Paper)
Subjects:	T Technology > TK Electrical engineering. Electronics. Nuclear engineering
Divisions:	Engineering
Publisher:	IEEE
Date of acceptance:	28 April 2026
Date Deposited:	07 Apr 2026 15:24
Last Modified:	07 Apr 2026 15:24
URI:	https://researchonline.ljmu.ac.uk/id/eprint/28341

View Item