Digital twin · BIM + ML · Industry project

Crack detection to digital twin: concrete water tank

A concept workflow linking crack detection, Eurocode-based width classification and a Revit model inside Power BI, so engineers can triage tank condition remotely.

Context: AECOM × Ulster University industry project (BEN715)
Period: Feb – May 2025
My role: Project lead (individual): scoping with AECOM, Revit modelling, ML, dashboards

Code & documentation: water-tank-crack-digital-twin

Power BI: the 3D Revit model beside the live crack register and recommended actions.

70.0%
VGG16 crack classifier: saved run, 30 test images · 89.69% in the submitted report; 36.67%
Baseline custom CNN: first attempt: over-fitted; 0.3 mm
EN 1992 w_max threshold used: Table 7.1N, exposure XC2; 80%
Module mark: BEN715 Industry Project

Chapter 01

The engineering challenge

Cracks in concrete water tanks cause leakage and deterioration. Tanks are often remote or difficult to access, which makes manual inspection slow and hazardous.

AECOM framed the problem at the kick-off meeting: connect inspection evidence to the asset model and to maintenance decisions.

Chapter 02

Input data

AECOM drawings

2D drawings of the tank, modelled in Revit

Crack images

Public RC-wall crack datasets (Mendeley, Utah State), 80/10/10 split

Design standard

EN 1992-1-1 Table 7.1N crack-width limits

Revit site model with the tank on terrain — Revit site model of the tank, built from AECOM's 2D drawings.

Chapter 03

The AI & engineering workflow

Stage 1 of 5

Model

2D drawings → Revit → IFC

What I did

Built a Revit model from AECOM's 2D drawings, exported it to IFC, and linked crack data back into Revit with Dynamo.
Trained a crack classifier. A first custom CNN over-fitted (36.67% test accuracy); a pretrained VGG16 reached 70.0% on the saved 30-image test run (89.69% reported in the submitted report).
Estimated crack width from masks and classified it against EN 1992-1-1 Table 7.1N (w_max 0.3 mm, XC2) to assign a maintenance action.
Moved visualisation from a browser IFC viewer, which failed on complex BIM files, to Power BI with an embedded 3D Revit visual, a crack table and a mobile layout.

Autodesk RevitDynamoIFCPythonTensorFlow / Keras (VGG16)OpenCVPower BISpeckle

Input → output

Drag to compare.

Crack photoPixel mask

Training pair from a public RC crack dataset: the photograph and the binary mask used to estimate crack width.

Chapter 04

Technical result

Delivered a working prototype. Detected cracks are classified, carry a recommended maintenance action, and are visible on the 3D asset in Power BI.
Recorded an evidence-based lesson on tool choice: lightweight browser IFC viewers could not handle the model, and Power BI with a 3D visual could.

Limitations, stated plainly

No camera was installed on site. The live-camera and cloud stages are a proposed architecture, and the report says so.
Training images came from public crack datasets (Mendeley, Utah State University), not from the tank itself. They carry no physical scale, so the mm widths rely on an assumed calibration.
The submitted report states 89.69% for VGG16; the last saved notebook run shows 70.0% on 30 test images. Both are published in the repository.

Outputs & artefacts

Revit model of an elevated concrete water tank on a steel frame — Fig. 01Tank model in Revit, exported to IFC and linked to crack data with Dynamo.

Crack register table with crack type, image link, suggested maintenance and width — Fig. 02Crack register: EN 1992 width class and a recommended action for each crack.

Crack photos outlined in green with class labels such as Severe Crack — Fig. 03Width-based classes; severe (>0.3 mm) flagged for immediate repair.

Binary crack masks above the original crack photos — Fig. 04Pixel masks used for width estimation.

Accuracy and loss curves for the VGG16 classifier — Fig. 05VGG16 transfer learning: accuracy and loss.

Diverging validation loss for the first CNN — Fig. 06First CNN: validation loss diverged, prompting the switch to VGG16.

Phone mock-up showing the 3D tank view and a Live button — Fig. 07Mobile view concept, fed from the same dataset.

Chapter 05

Practical impact & relevance

Bridges the structural standard (crack-width limits) with data science and BIM, which is the kind of problem digital-delivery teams bring to consultants.

EvidenceWhere every figure on this page comes from

Each metric is taken from a primary record, not from a CV. Hover a metric to see its source. The underlying files are available for review at interview.

BEN715 industry-project report (AECOM × Ulster, 2025)
BEN715 project files: Revit, Dynamo, Power BI
BEN715 dashboard captures

Next case study

BIM coordination & automation: The Brinell Building