BERTopic-Based Latent Hazards Mining and Topic Database Construction for Gas Pipeline Accidents

As critical energy infrastructure, gas pipeline networks play an essential role in maintaining stable energy supply and supporting energy infrastructure sustainability. However, accident reports are characterized by dispersed data sources, highly unstructured narratives, and inconsistent terminology, limiting systematic hazard identification and knowledge utilization for gas pipeline safety. To address these, this study proposes a latent hazard mining framework integrating a domain-specific lexicon with semantic topic modeling. A multi-source corpus was established using gas pipeline accident investigation reports. Subsequently, a specialized dictionary containing 379 domain-specific terms was developed through a new-word discovery method based on mutual information and adjacent entropy, enabling standardized word segmentation and structured text representation. Building upon the corpus, the BERTopic model was employed for deep semantic topic mining and benchmarked against the latent Dirichlet allocation (LDA) model. Eight highly discriminative hazard topics were identified, covering critical risk categories such as third-party encroachment, inadequate safety clearance, crossing and parallel installations, and pipeline structural defects. Furthermore, a structured hazard-factor framework comprising six categories and 60 specific hazard factors was established, facilitating the transformation of unstructured narratives into a standardized dataset. The results demonstrate that this work enhances hazard identification accuracy and semantic representation capability. It provides an effective tool for extracting safety knowledge from accident reports and offers support for gas pipeline safety assessment and risk management.