Evaluating LLM Prompting Strategies for Industrial Functional Safety Risk Assessment

Industrial Cyber-Physical Systems (ICPS) require accurate and transparent risk assessments to comply with safety standards such as ISO 12100 and ISO 13849. While Large Language Models (LLMs) offer automation potential for these assessments, their inconsistencies and lack of structured reasoning necessitate human oversight. This study evaluates four prompting strategies-Zero-Shot, Role-Based, Retrieval-Augmented Generation (RAG), and RAG+Role-Based-for hazard classification and required Performance Level (PLr) estimation. Results show that RAG significantly improves accuracy (75.8 %) over ZeroShot (24.2%) and Role-Based (25.8%), while RAG+Role-Based offers no additional benefit. Future work should enhance retrieval mechanisms with expanded safety rule embeddings and multistage reasoning to further refine high-risk classifications.