| Keywords |
Fire-protection design automation; Architectural floor plans; Automated recognition; Machine learning; ; AutoCAD |
| Abstract |
This study develops and validates an algorithm that automates fire-protection design from architectural drawings. Building-summary tables and annotations are parsed with optical character recognition (OCR) and OpenCV to normalize occupancy, gross floor area, story count, and floor height. Doors and columns are detected by a You Only Look Once v4 (YOLO v4) convolutional neural network, and room boundaries are reconstructed to form space-level metadata. A rule engine derived from the National Fire Safety Code (NFSC) determines installation or exemption for each system and computes equipment placement and wiring. Outputs are written as computer-aided design (CAD) entities through the AutoCAD application programming interface. In tests, table structure and text recognition reached accuracy 0.91, precision 0.89, recall 0.99, F1 Score 0.94, and intersectionover-union 0.83. Average detection confidence was 0.89 for doors and 0.86?0.93 for columns. Checklist comparison yielded about 97% normal outputs, and repeated runs reproduced coordinates and connections, indicating reliable end-to-end automation from image inputs to CAD deliverables. |