跳转至

基于确定性几何与受控智能体视觉语言优化的结构图纸转模型方法

文章背景与核心概要

将结构框架图纸转换为可编辑的有限元模型草图，在历史上一直是一项耗时且易出错的工作。传统的图纸理解系统通常依赖于针对特定任务训练的神经检测器，而现有的结构工程语言模型智能体往往仅处理文本或模型数据，无法直接解读视觉图纸本身。

本文提出了一种全新的框架，通过引入智能体视觉语言层，直接从结构框架图纸（PDF）中进行构件检测与模型草图绘制，且无需针对特定任务进行检测器的训练或微调。该框架结合了确定性几何提取与受控的智能体修正机制，在保证几何精度的同时，利用大模型的推理能力进行纠错，为结构工程自动化提供了高效的解决方案。

摘要概要

Converting structural framing plans into editable finite-element model drafts is historically labor-intensive and error-prone. Traditional drawing-understanding systems rely on task-specific trained neural detectors, while existing language-model agents in structural engineering operate solely on text or model data rather than the visual drawings themselves.

This paper introduces a novel framework that applies an agentic vision-language layer directly to structural component detection and model drafting from framing-plan PDFs—without requiring task-specific detector training or fine-tuning.

关键框架架构

1. 确定性阶段：

  • 提取底层图元。
  • 通过尺寸比例共识估计比例尺。
  • 利用形式化绘图语法识别五类实体。
  • 组装可编辑的结构布局。
  1. Deterministic Stage:
  2. Extracts underlying primitives.
  3. Estimates scale via dimension-ratio consensus.
  4. Recognizes five entity classes utilizing a formal drafting grammar.
  5. Assembles an editable structural layout.

2. 智能体阶段：

  • 提出严格受限于确定性候选对象的类型化修正建议。
  • 采用特定操作的准入测试、变更级别审查以及故障关闭（fail-closed）事务处理。
  1. Agentic Stage:
  2. Proposes typed corrections strictly constrained by deterministic candidates.
  3. Employs operation-specific admission tests, change-level reviews, and fail-closed transactions.

评估与结果

该框架使用作者生成的包含 100 张图纸的基准数据集进行了评估： * 开发集（一半）： 在创建过程中用于指导每一项规则的修订。 * 留出集（另一半）： 种子不重叠，且在规则冻结后生成，仅进行严格的一次性评估。

The framework was evaluated using an author-generated benchmark comprising 100 plans: * Development Half: Informed every rule revision during creation. * Held-Out Half: Seed-disjoint and generated after the rules were frozen, evaluated strictly once.

性能亮点

  • 比例尺估计： 每张图纸的精度均在生成器参考值的 \(0.1\%\) 以内。
  • 实体指标（召回率 / 精确率）：
  • 柱 (Columns): 0.922 / 0.997
  • 梁 (Beams): 0.886 / 0.990
  • 墙 (Walls): 1.000 / 1.000
  • 支撑 (Braces): 1.000 / 1.000
  • 洞口 (Openings): 1.000 / 0.964
  • 受控损坏研究： 在三张开发图纸上重复进行了三次损坏测试，共九次试验。校准通过了所有九次试验；构件修复在九次试验中有五次满足了所有严格的最终状态谓词。受控审查成功地在明确的操作范围内纠正了遗漏的框架和错误的标记。

Performance Highlights

  • Scale Estimation: Accurate within \(0.1\%\) of the generator reference for every drawing.
  • Entity Metrics (Recall / Precision):
  • Columns: 0.922 / 0.997
  • Beams: 0.886 / 0.990
  • Walls: 1.000 / 1.000
  • Braces: 1.000 / 1.000
  • Openings: 1.000 / 0.964
  • Controlled Corruption Study: Repeated two corruptions three times across three development drawings. Calibration passed all nine trials; member repair met every strict end-state predicate in five of nine trials. Guarded reviews successfully corrected missed framing and false marks within explicit operational bounds.

注：由于留出数据集与开发数据集共享相同的生成器，当前研究范围排除了独立绘制的图纸、栅格评估、分析连接性检查以及求解器验证。

Note: Because the held-out dataset shares the development generator, the current scope excludes independently drafted plans, raster evaluations, analytical connectivity checks, and solver validations.