01 · Automated safety interpretation of crack-aperture for concrete dams using mutation-aware visual-sensor fusion and knowledge-augmented reasoning

基于变异感知视觉-传感器融合与知识增强推理的混凝土坝裂缝开度自动安全解读

Research record	Details
Authors	Hao Gu, Jiale Wang, Yixiang Fang, Yifan He, Jun Shen, Chenfei Shao
Published	2026-10-10
Sources	openalex
Focus	crack monitoring, multimodal visual–sensor fusion, anomaly/mutation detection, GraphRAG (knowledge-augmented RAG), time-series prediction 裂缝监测, 多模态视觉-传感器融合, 异常/变异检测, GraphRAG（知识增强的检索增强生成（RAG））, 序列预测

English

Recommendation: skim. The paper is domain-specific but contains practical implementations of GraphRAG, multimodal triple extraction, and mutation-detection that may offer useful extractor designs or evaluation cues for the thesis; it is not central to core questions about agentic persistent graph memory or entity-resolution strategies.

中文

建议：略读。该论文具有领域专用性，但包含关于 GraphRAG、多模态三元组提取与变异检测的实用实现，可能为论文提供有价值的提取器设计或评估思路；但它并非关于智能体持久化图记忆或实体解析策略的核心参考。

Research synopsis

English

The paper addresses automated safety interpretation of concrete-dam crack apertures when crack-gauge records are incomplete and abnormal changes are hard to distinguish from normal evolution. It proposes a pipeline combining a deep visual extractor (MDS-YOLOv11n) to derive crack widths from images, alignment with gauge measurements to fill missing observations, and Autoformer for fused-sequence prediction. A dynamic three-sigma criterion and a Heaviside–EMA reconstruction identify and reconstruct mutation-induced increments. Finally, GraphRAG links mutation-analysis outputs to domain safety knowledge to produce interpretable recommendations; validation on a dam case shows improved prediction versus comparative models and GraphRAG recommendations consistent with inspection records.

中文

文章解决了在裂缝尺记录不完整且异常变化难以区分于正常演化时，混凝土坝裂缝开度的自动安全解读问题。提出了一个流水线：使用深度视觉提取器 MDS-YOLOv11n 从图像中提取裂缝宽度，和裂缝尺测量对齐以补全缺失观测，使用 Autoformer 对融合序列建模预测。通过动态三西格玛准则识别持久偏离，并用 Heaviside–EMA 重建变异引起的增量。最后，GraphRAG 将变异分析结果与坝体裂缝安全知识相连接以产生可解释的建议；在一个混凝土坝案例上的验证显示相比对比模型有更好的预测性能，且 GraphRAG 的建议与定期检查记录一致。

Thesis relevance

English

Overlap: the paper uses GraphRAG to connect analytic outputs to domain knowledge, showing a knowledge-augmented RAG application and a concrete example of turning measurements into knowledge-linked recommendations. Differences/limitations: it is a domain-specific monitoring system focused on multimodal sensor and image fusion and time-series mutation detection rather than on incremental, persistent graph memory constructed from agentic interactions. Complementarity: its visual-extraction, sequence-modeling, and mutation-detection modules could supply structured triples or state-change edges that a graph-memory system might ingest.

中文

重合点：该论文使用 GraphRAG 将分析结果与领域知识连接，展示了知识增强的 RAG 应用，并给出将观测转为知识驱动建议的具体范例。差异/限制：它是面向特定领域的监测系统，侧重多模态传感器与图像融合及序列变异检测，而不是基于智能体交互构建的增量、持久化图记忆。互补性：其视觉提取、序列建模和变异检测模块可生成结构化三元组或状态变更边，供图记忆系统摄取。

Writing and related work

English

Position this paper in related work as an applied GraphRAG instance combining multimodal extraction and domain knowledge for actionable recommendations. It is a useful citation for practical triple-extraction from images and for connecting analytic outputs to a knowledge-backed decision layer, but not as evidence for claims about agentic graph memory or entity-resolution strategies.

中文

在相关工作中将此论文作为一个应用型 GraphRAG 实例来定位，展示如何将多模态提取与领域知识结合以生成可操作建议。它是关于从图像提取三元组并将分析结果连接到知识驱动决策层的实用参考，但不能用来支持关于智能体图记忆或实体解析策略的论断。

Method and evaluation

English

Consider reusing the paper’s components as data-source-specific extractors: treat MDS-YOLOv11n outputs and aligned gauge measurements as entity-attribute assertions or time-stamped triples to insert into an evolving KG. Autoformer-driven mutation detection (dynamic three-sigma and Heaviside–EMA) can be used to create edges denoting state changes or confidence-weighted events. When integrating such signals into your evaluations, measure whether these structured, sensor-derived triples improve bridge-entity recovery and multi-hop path completeness compared with plain vectorized evidence.

中文

可以将论文的方法作为特定数据源的提取器重用：将 MDS-YOLOv11n 的输出和对齐后的尺测量视为实体属性断言或带时间戳的三元组，插入到演化的知识图谱（KG）中。基于 Autoformer 的变异检测（动态三西格玛与 Heaviside–EMA）可用于创建表示状态变化或带置信度的事件边。在将这些信号整合进你的系统时，应评估这些结构化、传感器生成的三元组是否比纯向量化证据更有助于桥实体恢复和多跳路径完整性。

Future directions

English

Map the paper’s mutation-detection outputs into incremental KG updates and evaluate how such sensor-derived state edges affect multi-hop retrieval and verification. Test entity-resolution approaches on sensor/image-derived entity mentions and assess latency/precision trade-offs when such extractors run on local models.

中文

将论文的变异检测输出映射为增量的知识图谱更新，并评估此类传感器来源的状态边如何影响多跳检索与可验证性。对来自传感器/图像的实体提及测试实体解析方法，并评估在本地模型上运行这些提取器时的延迟与精度权衡。

---