Four-Dimensional Fire Safety Knowledge-Enhanced Large Language Model for BIM Compliance Checking

Fire safety compliance checking is essential for BIM-based design review, yet it remains largely manual in practice, leading to substantial time costs. Although automated checking approaches (e.g., NLP-based methods) have been explored, two major challenges remain: (1) in practice, operationalization is largely limited to semantic requirements, whereas clauses involving geometric and topological structures are difficult to execute; (2) automation remains limited because NLP-based methods often require manual mapping tables due to terminology mismatch and unclear condition parsing, and the extracted rules are not sufficiently stable. To address these issues, this paper proposes the Framework of Four-Dimensional Fire Safety Knowledge-Enhanced Large Language Model (4FKE-LLM). First, this paper constructs a four-dimensional fire safety knowledge graph-Semantic, Geometric, Topological, and Scenario, which organizes and structurally represents regulations across the four dimensions to support subsequent LLM-based compliance checking. Second, this paper proposes an LLM-based two-layer compliance checking mechanism. The first layer pre-filters regulatory and model information to identify applicable regulations and relevant objects, and the second layer performs dimension-specific regulation matching and reasoning across the four categories. Experiments on a dormitory building case study show that the 4FKE-LLM framework achieves F1 = 0.863 and Accuracy = 0.885, and outperforms the two selected baselines on this case in terms of classification performance and checkable coverage.