Deep learning-based structural element extraction and knowledge-graph-based validation of 2D bridge drawings

Abstract

The maintenance and management of existing bridge infrastructure rely on conventional two-dimensional construction drawings, which remain largely unstructured and difficult to integrate into digital workflows, including Building Information Modeling (BIM). In prestressed concrete bridges, tendons and anchorage systems are documented across multiple drawings and sectional views, making manual interpretation time-consuming and error-prone. This paper presents an automated pipeline for detecting and semantically structuring structural elements in 2D bridge drawings. The approach combines deep learning-based object detection, classical computer vision for identifier-frame detection, optical character recognition, and spatial matching to identify elements, recognize identifiers, and link elements with their labels. The extracted information is integrated into an ontology-based knowledge graph, enabling SPARQL-based validation of missing identifiers, cross-view relationships, and section-level quantities. On 29 bridge drawings, detection achieved mAP@50 values of up to 97.4%, and text recognition reached 93.0% exact-match accuracy with a 2.2% character error rate.