AI-based Analysis Tools for System Safety : Exploring the potential of Large Language Models for system safety analyses in the early stages of product development

Sammanfattning

This thesis explores the potential of artificial intelligence (AI) to enhance system safety analyzes in the early phases of product development. Modern systems, particularly those used for defense applications, are becoming increasingly complex, which requires efficient and proactive hazard identification. Traditional hazard analysis methods are often time consuming and resource intensive, especially when applied early in development when system information is limited. A proactive and preventive approach to safety during system design and development has been shown to be more cost effective than adding safety features to a system after an accident or mishap has occurred. This study utilizes Design Research Methodology (DRM) to structure and implement the project’s phases. The research includes a literature review which deepened the understanding of existing knowledge about AI and system safety. The study investigates the ability of large language models (LLMs), specifically ChatGPT-4o and Gemini 2.0 Flash to perform system safety analyses. It assesses their performance in comparison to system safety engineers. More specifically, it examine show LLMs can be utilized in this context, how their ability to identify hazards varies with different levels of system information, and the quantity and relevance of the hazards they identify. The research applies a case study that analyzes an SAI unit (Safety, Arming, and Initiation) developed at Saab. The case study focused on Functional Hazard Analysis (FHA), an analysis method chosen because it is mandatory at Saab and applicable in early development phases and for a wide range of systems. The investigation was conducted through three progressively detailed tests, where the LLMs were provided with varying levels of information: functions only; functions and subfunctions; and functions, subfunctions, and function descriptions. The LLMs’ results were compared quantitatively and qualitatively with a reference FHA performed by system safety engineers at Saab. The results from the case study show that LLMs can generate a substantial number of potential hazards. However, their results require expert evaluation due to the inclusion of irrelevant or incorrect information.