关键词: 隧道支护, 合规性, 智能审查, 检索增强生成, 多模态, 思维链

Abstract: Manual verification of tunnel support parameter compliance suffers from high subjectivity, low efficiency, and a high risk of missed or erroneous judgments. To address these challenges, an intelligent tunnel support compliance review method is proposed based on multimodal retrieval-augmented generation (RAG). First, based on the compliance review requirements for railway and highway tunnel support, external private knowledge bases containing multimodal information (text, formulas, and tables) are constructed. Regular expressions are designed to associate tables with citation metadata, and a recursive loading strategy combining depth-first and breadth-first approaches is implemented to maximize restoration of the original document structure and internal and external knowledge associations in standards and specifications. Second, focusing on proprietary knowledge comprehension and design requirement interpretation, the generation performance of four large language models (ChatGPT-4o, DeepSeek-V3, ERNIE Bot 4.0 Turbo, and GLM-4-Plus) under three chain-of-thought (CoT) strategies is compared to identify the optimal generator for the multimodal RAG framework. In addition, the enhancement effect of multimodal RAG on generator performance is evaluated. Finally, from the perspective of user-friendly human-machine interaction, the effectiveness of multimodal RAG in autonomously reviewing and generating design schemes, regardless of the CoT strategy, is validated. The F₁ scores for autonomous initial parameter compliance review and compliant parameter generation reach 0.971 4 and 1.000, respectively, whereas those for assessing the necessity of reasoning steps reach 0.966 7 and 0.941 2. Application of the proposed method to the autonomous review of primary support parameter compliance and secondary lining reinforcement schemes for the Maoxian-Wenchuan fault section of the Maoxian Tunnel on the Chengdu-Lanzhou Railway shows that nine parameters are compliant, whereas four are appropriately strengthened in unfavorable geological sections according to independent design principles. The reinforcement scheme satisfies the requirements for sectional strength, durability, and constructability, with no necessary reasoning steps omitted or redundant steps generated.

Key words: tunnel support, compliance, intelligent review, retrieval-augmented generation, multimodal, chain of thought (CoT)