Intelligent Semantic Matching Model of Production Safety Hazard and Regulations in the Oil and Gas Industry

Addressing the significant semantic gap between the colloquial expressions of on-site hazard descriptions and the professional terminology of safety regulations in the oil and gas industry-which hinders efficient matching-this paper proposes an intelligent semantic matching model based on enhanced contrastive learning. First, a specialized dataset of hazard-regulation pairs was constructed using real-world manual inspection data from oilfields. To tackle the challenges of insufficient negative sample diversity and slow convergence in the late stages of traditional contrastive learning, an initial negative sample selection method based on vector retrieval was designed. Furthermore, a Dynamic Hard Negative Mining (DHNM) mechanism is proposed to optimize the negative sample pool in real-time by monitoring the model's score fluctuations during iterative training. Architecturally, a deep encoder based on a bidirectional attention mechanism is employed to extract robust text embedding features, combined with the InfoNCE loss function to optimize the vector space distribution for precise mapping between informal hazard descriptions and professional regulatory clauses. Case study results demonstrate that the proposed model performs excellently across key metrics such as Recall@K, providing critical technical support for automated compliance auditing and intelligent safety decision-making in oil and gas enterprises.

Keywords: Oil and gas production safety, Hidden danger identification, Safety regulations, Semantic matching, Contrastive learning, Sentence embedding, Bidirectional attention, Compliance management.