基于边缘计算的代理式检索增强生成技术:用于联邦公路管理局(FHWA)桥梁检测合规性
文章背景与核心概要
美国联邦公路管理局(FHWA)监管着全美超过60万座桥梁,要求其严格遵守《国家桥梁清单(NBI)记录与编码指南》。传统的合规性验证高度依赖人工,不仅耗时耗力且易出错,在网络连接受限的偏远现场环境下更是难以开展。
本文提出了 BridgeGuard 系统,这是一种完全离线、基于边缘计算的代理式检索增强生成(RAG)系统,专为自动化的桥梁检测合规性而设计。BridgeGuard 可在无需外部网络连接的通用边缘硬件上本地运行,通过编排多步 ReAct 规划循环,将针对监管指南的向量搜索与针对 NBI 表格数据的结构化 SQL 查询相结合,实现了高效、精准的现场合规性评估。
论文信息
arXiv ID: arXiv:2608.20372
学科分类: 信息检索 (cs.IR); 人工智能 (cs.AI); 多智能体系统 (cs.MA)
作者: Viraj Nishesh Darji, Hemaliben Rakeshkumar Darji
提交日期: 2026年6月23日(预印本已提交至 ASCE Journal of Computing in Civil Engineering)
资源: 代码与数据仓库 | 查看 PDF
arXiv ID: arXiv:2608.20372
Subjects: Information Retrieval (cs.IR); Artificial Intelligence (cs.AI); Multiagent Systems (cs.MA)
Authors: Viraj Nishesh Darji, Hemaliben Rakeshkumar Darji
Submitted: June 23, 2026 (Pre-print submitted to the ASCE Journal of Computing in Civil Engineering)
Resources: Code and Data Repository | View PDF
执行摘要
美国联邦公路管理局(FHWA)监管着全美超过60万座桥梁,要求其严格遵守《国家桥梁清单(NBI)记录与编码指南》。传统的合规性验证高度依赖人工,不仅耗时耗力且易出错,在网络连接受限的偏远现场环境下更是难以开展。
为了解决这些挑战,本文介绍了 BridgeGuard,这是一个完全离线、基于边缘计算的代理式检索增强生成(RAG)系统,专为自动化的桥梁检测合规性而设计。BridgeGuard 可在无需外部网络连接的通用边缘硬件上本地运行,通过编排多步 ReAct 规划循环,将针对监管指南的向量搜索与针对 NBI 表格数据的结构化 SQL 查询相结合。
The Federal Highway Administration (FHWA) regulates over 600,000 bridges across the United States, requiring strict adherence to the Recording and Coding Guide for the National Bridge Inventory (NBI). Traditionally, compliance verification has been manual, labor-intensive, error-prone, and severely restricted in remote field environments with limited network connectivity.
To solve these challenges, this paper introduces BridgeGuard, a fully air-gapped, edge-based agentic Retrieval-Augmented Generation (RAG) system engineered for autonomous bridge inspection compliance. BridgeGuard runs locally on commodity edge hardware without external network access, combining vector search over regulatory guides with structured SQL queries over NBI tabular data through an orchestrated multi-step ReAct planning loop.
关键创新与方法论
- 离线边缘架构: 完全在通用边缘硬件上本地运行,无需依赖云端连接即可实现可靠的现场合规性检查。
- 混合数据集成: 无缝结合了针对 FHWA 记录与编码指南的非结构化向量搜索,以及针对 NBI 表格数据的结构化 SQL 执行。
- 基于章节感知的切片算法: 保留了层级化的监管条目边界,实现了 94.2% 的切片完整性(相比之下,传统的固定大小切分仅为 28.4%)。
- 有状态的多步 ReAct 规划: 采用自主代理循环来精确编排检索、查询和监管推理步骤。
- Air-Gapped Edge Architecture: Operates entirely locally on commodity edge hardware, enabling reliable field compliance checks without relying on cloud connectivity.
- Hybrid Data Integration: Seamlessly combines unstructured vector search over the FHWA Recording and Coding Guide with structured SQL execution over NBI tabular data.
- Section-Aware Chunking Algorithm: Preserves hierarchical regulatory item boundaries, achieving a 94.2% chunk integrity rate (compared to just 28.4% for naive fixed-size splitting).
- Stateful Multi-Step ReAct Planning: Employs an autonomous agentic loop to orchestrate retrieval, querying, and regulatory reasoning steps accurately.
性能与评估结果
BridgeGuard 在大规模公共桥梁数据集上进行了严格测试: * 特拉华州 2023 NBI 清单(874 座桥梁): 在识别结构性缺陷桥梁方面达到了 99.77% 的分类准确率。 * 德克萨斯州样本数据集(200 座桥梁): 达到了 100.0% 的分类准确率。 * 引用准确性: 在评估中保持了完美的 100.0% 引用准确率。 * 处理速度: 在边缘硬件上本地评估的处理速度为 每小时 197.0 座桥梁。 * 消融实验结果: 证实了向量搜索组件和多步代理循环对于维持稳健的合规性推理都是必不可少的。
BridgeGuard was rigorously tested on large-scale public bridge datasets: * Delaware 2023 NBI Inventory (874 bridges): Achieved 99.77% classification accuracy for identifying Structurally Deficient bridges. * Texas Sample Dataset (200 bridges): Achieved 100.0% classification accuracy. * Citation Accuracy: Maintained a flawless 100.0% citation accuracy across evaluations. * Processing Speed: Evaluated at a rate of 197.0 bridges per hour locally on edge hardware. * Ablation Findings: Confirmed that both vector search components and the multi-step agentic loop are strictly necessary to maintain robust compliance reasoning.
引用与元数据
- 引用格式:
arXiv:2608.20372 [cs.IR] - DOI: 10.48550/arXiv.2608.20372
- ACM 分类: I.2.7; H.3.3
- MSC 分类: 68T50
- Cite as:
arXiv:2608.20372 [cs.IR]- DOI: 10.48550/arXiv.2608.20372
- ACM Classes: I.2.7; H.3.3
- MSC Classes: 68T50