Session: 06-17-01 AI Technology for Ocean Engineering - I

Submission Number: 180694

Hazard Identification Support Tool Utilizing a Large Language Model for Improving Efficiency and Comprehensiveness in Risk Analysis 

In recent years, technical development for the practical use of autonomous ships has progressed rapidly. For large-scale and complex systems, such as autonomous ships, it is important to determine the operational procedures and conceptual design of the entire system at the initial design stage. In addition, it is essential to identify and address serious hazards and risks at the initial design stage to avoid rework at later stages, such as detailed design, construction, or demonstration tests.

The Structure Model-Based Hazard Identification (SMB-HAZID) method is proposed as a suitable method for hazard identification in large-scale and complex systems such as autonomous ships, which enables comprehensive hazard identification by brainstorming utilizing structure & task (ST) diagrams, keywords and checklists. ST diagrams were developed based on a class diagram of Unified Modeling Language (UML) and could help analysts understand the structure of the target system and its components’ tasks and the interactions among them. However, it was revealed thorough its application to a hypothetical autonomous ship that repetitive works for a huge number of hazards would increase cognitive load for analysts, which could result in reduced work efficiency and potential human errors such as overlooking critical hazards.

To address this issue, this study develops a hazard identification support tool that leverages a Large Language Model (LLM) to automate parts of the SMB-HAZID process. The tool automatically generates prompts using structured system data derived from an ST diagram, inputs them into the LLM and records the output from the LLM in a worksheet format.

A prototype of the tool has been developed in Python3, integrating GPT-4.1 nano via Application Programming Interface (API) and verified through the application to the hypothetical autonomous ship. The results demonstrated that the tool could produce appropriate hazard descriptions when prompts were refined to include relevant context and structured information on the target system with clear and explicit instructions; however, there is room for further improvement of the tool by some means such as prompt optimization.

Furthermore, the findings gained through the verification of the prototype suggest that combining human expertise with Artificial Intelligence (AI) capabilities can improve both efficiency and comprehensiveness of hazard identification. By automating mechanical tasks, human analysts can focus on complex, knowledge-intensive aspects of safety analysis, with reduced cognitive load.

Presenting Author: Megumi Shiokari National Maritime Research Institute, MPAT

Presenting Author Biography: Dr. Shiokari has some experience in risk analysis for ships and onboard systems as well as offshore wind turbines. She has recently contributed to the development of risk assessment frameworks for autonomous ships including the development of hazard identification methods for autonomous ships. These insights have also been applied to support risk management in autonomous ship demonstration projects in Japan.
She has also involved in discussions on development of regulations for maritime autonomous surface ships (MASS), in cooperation with Japanese Government, at the Maritime Safety Committee (MSC) of the International Maritime Organization (IMO).

Authors:

Megumi Shiokari National Maritime Research Institute, MPAT
Yuki Tomita the University of Tokyo
Kazuhiro Aoyama the University of Tokyo