摘要

现有建筑信息模型（BIM）审查方法大多针对特定类型的设计缺陷，无法在同一流程中识别并修正多类与属性相关的 BIM 设计缺陷。为此，本文提出一种基于大语言模型（LLM）的集成框架，用以识别并修正 BIM 中多类属性相关设计缺陷。首先，提出采用构件均衡分块的 BIM 转文本方法，打通 BIM 数据与大语言模型之间的数据通路；其次，构建融合规则注入与少样本提示的提示学习方法实现缺陷识别，并搭建检索增强生成（RAG）方法输出修正建议；同时设计结合关键标识校验与词元长度阈值的幻觉抑制策略，提升系统可靠性。实验结果表明，该方法性能优于规则审查方法；经过微调的大语言模型缺陷识别精度提升 15%，修正建议合理度达 94%；幻觉抑制策略将准确率由 64% 提升至 85%，单轮可消除 92.5% 的模型幻觉。

Existing Building Information Modeling (BIM) checking methods focus on specific design defect types, while failing to identify and correct multi-type property-related BIM design defects in one loop. Therefore, this paper proposes an integrated framework to identify and correct multi-type property-related design defects in BIM via Large Language Models (LLMs). First, a BIM-to-Text method with component-balanced chunking is introduced to bridge BIM data and LLMs. Then, a prompt learning based method incorporating rule injection and few-shot prompting is proposed for defect identification, and a Retrieval Augmented Generation (RAG) method is established for correction suggestions. A hallucination control strategy combining key-identifier validation and token-length thresholds is developed to improve reliability. Experiments show that the proposed approach outperforms rule checking, and the fine-tuned LLM improves identification accuracy by 15%, with a 94% rationality rate for correction suggestions. Hallucination control increases accuracy from 64% to 85%, eliminating 92.5% of hallucinations in one round.

论文下载链接

预印本下载链接

The authors are grateful for the financial support received from the National Key R&D Program of China (No. 2023YFC3804600) and the National Natural Science Foundation of China (No. 52378306).

相关项目资助情况: