A recent article published in Buildings proposed an automatic approach for generating building information modeling (BIM) of mechanical, electrical, and plumbing (MEP) systems based on two-dimensional (2D) drawings.
Semantic segmentation was utilized to identify MEP components in the drawings, trained on a custom-made MEP dataset, achieving a mean intersection over union (mIoU) of 92.18%.
Image Credit: Gorodenkoff/Shutterstock.com
Background
Many existing buildings face aging MEP (mechanical, electrical, and plumbing) systems, which can lead to a range of problems—from decreased energy efficiency in HVAC equipment to leaks in water and drainage systems and, more critically, deteriorating electrical wiring. These issues not only affect building performance but also raise safety concerns.
Traditional methods for operating and maintaining MEP systems still rely heavily on paper-based documentation—such as as-built 2D drawings and maintenance logs—that often fail to capture the current state of the building accurately. This makes troubleshooting and planning more difficult for facility teams.
Building Information Modeling (BIM) offers a more efficient approach to MEP system operation and maintenance. By providing a clear, interactive 3D model, BIM allows O&M teams to visualize better and understand system layouts.
Additionally, BIM enables the integration of historical and real-time operational data, making it possible to identify patterns, anticipate potential failures, and plan preventive maintenance more effectively. This proactive strategy helps minimize unexpected breakdowns and keeps systems running more smoothly.
Interested in how digital tools are reshaping facility management? You might also want to explore topics like digital twins, IoT integration in smart buildings, or predictive maintenance technologies.
Methods
This study adopted a hybrid methodology that combines semantic segmentation with computer vision techniques to extract key parameters—such as coordinates, classification, and pixel count—from MEP 2D drawings.
The process began by developing a diverse and comprehensive MEP dataset using images of the components to be identified. MMSegmentation was then utilized to train several neural networks on this dataset. The model achieving the highest evaluation score, measured by mean Intersection over Union (mIoU), was selected for image prediction.
Following segmentation, contour and bounding box detection were applied to each component. From the bounding box’s four corner coordinates, essential modeling information—such as the center point and rotation angle for equipment or the two endpoints for ducts—was derived.
Next, pixel counts and annotation data were matched against a pre-established MEP component dictionary to determine the specific model type of each component. Pixel counts were converted to real-world dimensions to ensure accurate model matching. For ductwork, OCR (optical character recognition) was employed to extract annotation data.
Finally, Dynamo was used to automatically generate 3D Revit models of the MEP components based on the extracted coordinates and dimensions from the dictionary. The entire workflow was validated through real-world experiments, and five different neural networks were trained using the MMSegmentation library to benchmark performance.
Curious about how AI is streamlining building design and modeling? You might also be interested in exploring topics like AI-driven CAD tools, automated clash detection, or machine learning in construction planning.
Results and Discussion
Five neural networks—DeeplabV3+, Segformer, K-Net, PSPNet, and Fast-SCNN—were trained and tested using the MMSegmentation library. The mean Intersection over Union (mIoU) was used as the primary evaluation metric to determine which network weights to retain.
After 40,000 training iterations, K-Net delivered the highest mIoU, outperforming the others in evaluation accuracy. Except for the lightweight Fast-SCNN, all models achieved an mIoU above 89%, underscoring the quality and consistency of the custom-built dataset.
However, some challenges emerged during prediction. Because the input images were cropped, certain images only contained small portions of a component, making it difficult for the model to distinguish between similar equipment types.
In some cases, relevant features occupied only a small area of the image, leading to incorrect classifications. To address this, data augmentation techniques—such as rotation, scaling, and mirroring—are recommended. These operations would help the neural networks better recognize subtle visual differences and improve overall classification accuracy.
Early in training, when the dataset was still being developed, certain components like fittings and ducts were occasionally missed in predictions. This was particularly true for ducts, which were often represented by two thin, parallel lines—making their features harder to detect.
To improve detection, the dataset was enhanced by significantly increasing the number of annotated images for these components.
Another issue arose with the predicted semantic segmentation masks: some boundaries did not perfectly align with the actual contours of the component drawings. Though the mismatch was typically within 1–2 pixels, it introduced slight shifts in the calculated component coordinates during bounding box extraction.
Looking to boost segmentation accuracy even further? Exploring advanced post-processing techniques, boundary refinement algorithms, or higher-resolution image inputs could be worthwhile next steps.
Conclusion and Future Prospects
Overall, the researchers addressed the pressing challenges of aging building infrastructure within the context of digital modernization. By integrating semantic segmentation, image processing techniques, OCR, and Dynamo for automated model generation, they developed a comprehensive approach to support MEP system operation and maintenance.
The proposed methodology focuses on efficiently identifying and processing various duct and equipment components from 2D MEP drawings—accurately extracting their coordinates and classifying each component. This not only streamlines model creation but also lays the groundwork for smarter facility management workflows.
Looking ahead, the team plans to enhance the method’s real-world applicability by expanding the range of component classes and model types in the dictionary. They also aim to improve the neural network’s ability to recognize components in more complex and cluttered 2D drawings.
Beyond immediate MEP applications, the resulting BIM models could serve as valuable data sources for virtual and augmented reality environments, opening the door to more immersive facility management and training tools.
Interested in where this technology is headed next? Consider exploring how AI-driven modeling integrates with digital twins or how VR and AR are being used in building inspection and maintenance workflows.
Journal Reference
Wang, D. & Fang, Y. (2025). Automatic BIM Reconstruction for Existing Building MEP Systems from Drawing Recognition. Buildings, 15(6), 924. doi: 10.3390/buildings15060924. https://www.mdpi.com/2075-5309/15/6/924.
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.