基于知识图谱的梁桥检测报告自动抽取与分析

Knowledge Graph-based Automatic Extraction and Analysis Method for Girder Inspection Reports

摘要:

桥梁健康监测是桥梁运维管养的重要技术手段之一，目前大多应用于大跨度桥梁，而中小跨度桥梁的性能评估主要仍依靠人工巡检，并以具有专业性、规范性的文字记录形成报告。大部分中小跨径桥梁为梁桥，设计服役年限在五十年以上，每年至少产生一份检测报告，因此，形成了数量庞大、语言风格相对统一的桥梁检测报告。随着桥梁数量的增加，需要对大量历史检测报告的信息进行重新梳理，以提取桥梁性能的历史演变规律及区域内多座桥梁间的时空关联特性，这使得人工阅读获取有效信息的分析手段十分耗时耗力。为提高提取检测报告中有效信息的效率，提出了一种基于知识图谱的适用于桥梁专业领域的信息自动抽取与分析方法，在现有大规模预训练模型上继续开发，根据“预训练+继续预训练+微调”的研究范式，训练了一个针对桥梁领域的预训练模型BriBERT，对检测报告中的病害描述语句进行信息提取，以知识图谱的形式存储历年信息，构建桥梁病害发展数据库，为分析、研究和预测桥梁状态提供数据支持。

Abstract:

Bridge health monitoring serves as a crucial technical approach for bridge operation and maintenance. While widely applied to large-span bridges， the performance evaluation of small- and medium-span bridges predominantly relies on manual inspections， which are documented in professional and standardized text reports. Most small- and medium-span bridges are girder bridges with a design service life exceeding fifty years. Since at least one inspection report is generated annually， a massive volume of bridge inspection reports with a relatively uniform language style is accumulated. With the increasing number of bridges， the information within massive historical inspection reports must be reorganized to extract the historical performance evolution patterns and the spatiotemporal correlation characteristics among multiple regional bridges. Consequently， manual reading for effective information extraction becomes a highly time-consuming and labor-intensive analytical method. To enhance the extraction efficiency of effective information from inspection reports， a knowledge-graph-based automatic information extraction and analysis method tailored for the bridge domain was proposed. Based on existing large-scale pre-trained models， a bridge-domain pre-trained model， BriBERT， was trained following the "pre-training + continued pre-training + fine-tuning" research paradigm. Through this method， information is extracted from the defect description sentences within the reports， and historical information is stored via knowledge graphs. Ultimately， a bridge defect development database is constructed to provide data support for bridge condition analysis， research， and prediction..