Adaptive detection of equipment components in MEP construction drawings via graph-enhanced siamese vision

Abstract

Building Information Models (BIMs) of mechanical, electrical, and plumbing (MEP) systems are critical for facility operation and maintenance. Previous studies have explored the automatic MEP BIM reconstruction from construction drawings, yet detecting equipment components (ECs) remains challenging due to heterogeneous drawing styles and enterprise-specific conventions. To address the challenge, this paper proposes a novel self-adaptive solution for automatic EC detection in real-world MEP drawings using graph-enhanced self-supervised computer vision models. Both query objects (i.e., EC samples) and drawings are transformed into graph-based representations, with multiple features extracted. A graph-based retrieval process with Principal Component Analysis filtering is then employed to localize regions of interest, after which ECs are identified via a hybrid similarity measure that integrates graph distance with a self-supervised Siamese network. The method is evaluated on a total of 291 EC instances across 18 object types from real-world MEP drawings. For the 13 object types (192 instances) included in training, it achieved a precision of 88.6%, recall of 89.5%, and F1 score of 0.89. For the 5 unseen object types (99 instances), the method maintained strong performance with a precision of 98.9%, recall of 90.9%, and F1 score of 0.95, showing that the proposed approach generalizes well to new EC categories and diverse drawing conventions. The research contributes to MEP BIM theory by formalizing an adaptive detection process for automatically reconstructing equipment objects from engineering drawings, reducing manual modeling time and providing reliable digital models for facility management.