This study designs and implements a generative AI–based fire investigation support pipeline that can be operated in an on-premise environment, with the objective of systematically utilizing unstructured fire incident overviews produced during fire i...
This study designs and implements a generative AI–based fire investigation support pipeline that can be operated in an on-premise environment, with the objective of systematically utilizing unstructured fire incident overviews produced during fire investigation processes. Existing fire investigation information systems rely primarily on structured fields and keyword-based retrieval methods, which limits their ability to adequately capture the contextual and structural similarities embedded in overview descriptions written by investigators.
To address these limitations, this study proposes an integrated pipeline that takes investigators’ original fire incident overviews as input and sequentially performs text preprocessing, similar case retrieval and selection, ignition cause inference, investigation report generation, and result evaluation. In the preprocessing stage, sentence correction and summarization are applied to reduce variations in expression while preserving the original content, and keywords relevant to fire investigation are extracted for use as search queries. In the similar case retrieval stage, semantic-based retrieval and keyword-based retrieval are combined, and cases are selected based on the structural consistency of incident components, enabling stable identification of similar cases that are suitable for investigative reasoning.
In the inference stage, ignition factor categories are estimated based on the selected set of similar cases, thereby examining whether the quality of the retrieved cases can serve as valid grounds for investigative reasoning. Subsequently, in the report generation stage, investigation reports are automatically generated by integrating fire incident overviews, preprocessing results, and structured metadata in accordance with prescribed investigation record formats. The generated reports are then evaluated in terms of structural consistency, content completeness, and clarity of expression to assess their suitability for use as official investigation records. This process is intended to provide collaborative support for investigators by organizing and refining investigation results rather than replacing human judgment.
In addition, this study considers the risk of model file exposure arising from internal access in on-premise deployment environments and applies an encryption-based weight protection scheme to mitigate unauthorized model use. The computational overhead introduced by the proposed protection method is quantitatively evaluated to assess its feasibility in practical operational settings. This study treats fire investigation support as a structural problem that requires joint consideration of unstructured data utilization, reasoning support, workflow integration, and operational constraints, and suggests that the proposed approach can be extended to other investigation and analysis domains with similar characteristics.