Generative AI Applications For Occupational Safety Engineering: Successful Implementations And Lessons Learned

Occupational safety performance highly relies on integrating lessons from near-misses and past accidents. Artificial Intelligence, and in particular Generative AI, provides significant opportunities to reanalyze massive corpuses of data related to occupational safety, and contributes to the transformation in the safety/HSE engineering domain. This paper examines various approaches implemented since early 2023 using Generative AI. The internal TotalEnergies digital solution, ‘SafeTE’, has demonstrated to be an effective platform for proving concepts, benchmarking capabilities, and creating new user experiences by leveraging historical safety data. The solution successfully covers several use cases targeted at different persona and operational processes: Learn by experience: Based on the accidents and near-misses database: GenAI summarizes and finds related incidents from accidents and near-misses. This is the backbone of SafeTE. The objective is to access and associate faster cases with operations, to learn from their causes, assess and study the correctives and remediations actions. Based on the anomalies, observation cards and stop cards records: topic modelling with GenAI continuously models and tracks trends in field reports. This allows weak signal detection and pair comparison between Operational sites. By connecting between planned activities and past accidents/near-misses cases database, GenAI is leveraged to interpret connects the context of planned activities with relevant past incidents occurred in a similar context, to typically support toolbox talks and job preparation. After over a year of use, the application shows that generative AI is well-suited for occupational safety data enhancement and elaboration of safety performance-oriented use cases. Not only the Large Language Models perform well in this domain (where technical jargon is limited and where situational comprehension is generally accessible to these types of technologies), but user experience is enhanced with respect to previous processes, where reliance on memory was critical for accident recall, as well as a limitation. The key findings are: LLMs / Agents, semantic vectorization and topic modelling do work very well for occupational safety, where technical jargon is not critical to situational understanding. Establishing benchmark procedures enables us to track the quality / performance of LLMs over time, as new models become available. Process safety remains more challenging as we write this paper. The user research shows it is a very different use case, but and jargon, technicity, lower heterogeneity of situations makes it a still-to-explore domain. This paper explores the use cases, technical implementation, and lessons learned from applying Generative AI for the integration of textual and document data in occupational safety engineering. The good users and stakeholders adoption confirms the relevance of this innovative technique in defeating information overload and increasing situational relevance.