Transformer-based pretrained models (e.g., BERT) have recently emerged as effective tools for automated incident analysis but typically require extensive labeled datasets, leading to risks of overfitting and limited interpretability. To address these challenges, this study proposes a novel framework employing generative large language models (LLMs), enhanced with structured, domain-specific Chain-of-Thought (CoT) prompting and retrieval-augmented generation (RAG) strategies. Specifically, the framework integrates hierarchical root cause classification guidelines from the American Bureau of Shipping (ABS) with offshore platform incident reports published by the Bureau of Safety and Environmental Enforcement (BSEE). Preliminary results using GPT-4o-mini indicate that domain-specific CoT prompting significantly improves LLM classification accuracy, achieving notably higher precision compared to basic zero-shot and few-shot methods. Furthermore, incorporating RAG strategies leveraging local domain knowledge bases enhances model reliability and supports continuous refinement through accumulating incident data.
This research highlights the substantial potential of generative LLMs in automating and advancing incident analysis practices, providing a viable path toward digitalizing incident investigation processes and strengthening PSM practices. Ultimately, the proposed approach enables systematic learning from incidents, offering valuable opportunities to reduce both the frequency and severity of future events in the process industry.