Analysing roadwork zone incursions and accidents using a sustainable GenAI system

Roadwork zone incursions, which refer to unauthorized or unintended entries into restricted work areas, present significant safety hazards and often lead to road accidents. While these incidents account for a fraction of total crashes, their potential severity makes them a critical concern for road safety management. Analysing incursion data and accident records is essential for identifying risk factors, understanding patterns, and improving prevention strategies. However, the data spans structured, semi-structured, and unstructured formats, growing in complexity over time, making traditional analysis methods unsustainable and difficult to scale. These methods also fail to account for the relational and hierarchical dependencies within data and lack natural language (NL) interaction, limiting user engagement and the ability to derive insights. A sustainable solution must be flexible, scalable, and adaptable to diverse data formats, with intuitive interactions for long-term efficiency. Recent generative artificial intelligence (GenAI) advancements offer solutions for analysing multi-format data by incorporating relational dependencies. However, standalone large language models (LLMs) struggle with integrating private accident datasets. To address this challenge, this study proposes a GenAI-powered information system (IS) that leverages LLMs and retrieval-augmented generation (RAG) for context-aware accident analysis through NL interactions. The system analyses real-world incursion data and associated accidents, identifying key risk factors and emerging patterns. Evaluated using the Llama3.2 and DeepSeek-R1 models, the system exhibits scalability, efficiency, and sustainability in roadwork safety management, contributing to improved risk mitigation and accident prevention.