Deep Learning-Based Detection of Structural Components and Data from Legacy Engineering Drawings

Farhan Tanvir Nabil; Gursans Guven

Abstract:

The lack of digitization and efficient utilization of legacy data stored in paper-based engineering drawings pose a significant challenge in the maintenance and management of aging infrastructure, as these documents are critical for assessing structural conditions and planning repairs. Extracting and organizing information from a large volume of drawings is traditionally a manual, time-consuming, and error-prone process. This paper presents an ongoing research on developing an automated asset inventory framework designed to extract, structure, and utilize data from legacy engineering drawings using a state-of-the-art Deep Learning (DL)-based object detection model (YOLOv11). The proposed methodology integrates YOLOv11 to identify and extract geometric and non-geometric data from scanned engineering drawings. The dataset, collected from real-world electrical substations, includes key structural assets information such as asset names, ages, cross-sectional views, plan views, dimensions, and general notes. Preliminary findings demonstrate that the model accurately detects and localizes critical structural components and their attributes, achieving a mean average precision (mAP) of 0.953 and an F1 score of 0.92 at a 0.344 confidence level, validating its effectiveness in automatically extracting information with high accuracy. By automating an image-based asset inventory creation from legacy engineering drawings, this framework facilitates structural condition assessment at electrical substations. Additionally, it assists informed decision-making for repairs, material reuse, and lifecycle optimization. By streamlining data extraction from legacy engineering drawings, this research contributes to the development of more efficient, data-driven infrastructure asset management practices.