Data-Driven Safety: Leveraging Historical Accident Data and Generative AI to Generate Workplace Hazard Scenarios

Workplace safety is a critical concern, with preventable hazards posing significant risks to workers and operations. An effective way to reduce hazard-related accidents is to use data-driven methods. These methods leverage historical workplace accident data to identify and mitigate safety hazards. Implementing such approaches is essential to keep pace with modern and dynamic workplaces. However, a gap in data-driven methods remains related to the lack of labeled data to train such models. In this work, we present a method for utilizing historical accident data to generate images of accident-prone scenarios in the workplace. Our methodology uses large language models to analyze historical accidents, identify causes, and generate alternative scenarios with similar outcomes, then employs diffusion models to visually represent these scenarios as pre-accident images. These images, supplemented by detailed descriptions of the corresponding hazards, constitute a comprehensive dataset for training vision models specifically engineered to identify subtle and overt workplace hazards.