Plug and produce manufacturing systems have the potential to significantly increase the efficiency and adaptability of manufacturing pipelines. However, to fully leverage this potential, it is essential to have efficient solutions for hazard safety detection to avoid bottlenecking the efficiency gains. A crucial step in hazard safety detection is effectively relaying information about risks and mitigation strategies to the target operator.

This thesis hence explores the feasibility of using large language models, particularly GPT-4, for generating safety texts in automatic hazard detection systems within smart manufacturing environments. The study focuses on enhancing existing hazard detection outputs into clear, human-readable safety instructions. The overall goal being to optimize this step of the hazard detection procedure.

Through a case study, the research demonstrates that GPT-4 can produce high-quality, fluent, and comprehensive safety texts that address identified hazards effectively. Additionally, a synthetic dataset is created to quantitatively measure how robust GPT-4 is incovering addressed hazards, revealing that it frequently covers all risks. While the generated texts showed high coverage of hazards and were generally well-received by experts, the study also identified areas for improvement, such as the need for better steerability and the inclusion of specific hazard prevention documentation.

Finally, we note that GPT-4 has a high tendency of adding superfluous information that was not requested when generating summaries.