Machine Learning-Based Indexing of 2D MEP Drawings

Mechanical, electrical, and plumbing (MEP) drawings created after project completion are valuable data sources that can be revisited to save time by avoiding the repetition of similar drawings and design situations. This study reports a use case in which machine learning (ML) methods were employed to index and retrieve contextually similar vector-based 2D MEP drawings, thereby improving workflow speed. Our approach uses ray-shooting algorithms to analyze vector-based 2D drawings and generate feature vectors capturing the geometric patterns of MEP drawings. These feature vectors are then clustered and visualized using Autoencoder (AE), constituting a qualitative navigation tool for AEC offices to index MEP drawings and their sub-elements within a structured database. Furthermore, after training the AE framework, a k-d tree model was used to identify and localize similar MEP layouts and aid the retrieval of similar geometric patterns, thereby helping create a machine-learning-based shape-indexer of MEP drawings. The method was applied to data from a completed project, showing that ML-based indexing can enhance project workflows, reduce manual workload, and improve MEP documentation. Further use cases for 2D vector-based searches of Building Information Modeling (BIM) or 2D MEP drawings hold the potential for expanding the capabilities of automated MEP drawing management. Future research will focus on refining ML models, expanding dataset diversity, and integrating the approach into BIM systems for broader AEC industry adoption.