AI Automates Steel Inspection with Scan-vs-BIM

By Nidhi DhullReviewed by Susha Cheriyedath, M.Sc.Dec 17 2024

A recent study in Applied Sciences has introduced a deep learning-based Scan-vs-BIM framework for automating steel structure inspections. Using as-built scan and as-planned BIM data, the framework's deep neural network evaluates structural integrity (SIE) and analyzes error types (SETA) for streamlined assessments.

The Role of Structural Integrity Evaluations

Structural integrity evaluations are vital for maintaining safety and quality in construction projects, spanning facility management, repairs, new builds, and remodeling. Yet, traditional methods rely heavily on manual inspections, which often lead to inconsistent results and inefficiencies.

In recent years, laser scanning has emerged as a powerful tool to capture precise as-built data for comparison with as-planned models. However, current methods frequently involve manually designed models, making the process time-consuming and resource-intensive. This challenge has created a demand for automated tools that streamline the "as-planned vs. as-built" comparison process, enhancing accuracy and efficiency.

Framework Methodology

In the study, the proposed Scan-vs-BIM framework integrated a distance-based deep neural network (distance-DNN) to automate the evaluation process. It involved three main steps:

1. Preprocessing: A bounding space is created for each structural object based on BIM parameters. This serves as a reference for separating BIM shape models and sampling point cloud data for individual analysis.

2. Scan-vs-BIM: Geometric relationships between the point cloud and BIM models are captured, allowing the framework to analyze as-built versus as-planned discrepancies.

3. Distance-DNN: Unlike traditional methods, which rely on complex comparative models, this framework simplifies the process by analyzing only distance and index data. These inputs are processed through the DNN to evaluate structural integrity (SIE) and identify structural errors (SETA).

The framework was implemented using Python, with TensorFlow for deep learning model development. Libraries like NumPy and Pandas were used for data manipulation, and Matplotlib was employed for result visualization.

The training process utilized 26,500 real-world datasets and 65,000 virtual datasets to ensure the model could effectively evaluate real structural objects and error types. Performance was measured based on accuracy and loss during the learning phase.

Application and Results

The Distance-DNN framework was tested on an actual steel structure comprising 184 components across a 423 m² area and 10 m in height. A 3D scanning dataset of approximately 10 million points was collected during a remodeling project.

Key findings include:

• Speed: The framework evaluated 20 structural columns in just 42 milliseconds (2.1 milliseconds per object), significantly faster than traditional Scan-vs-BIM approaches, which took around eight hours for a similar dataset.
• Accuracy: The SIE network achieved an average accuracy of 94.68 % in structural integrity assessments.
• Error Analysis: The SETA network identified a high proportion of errors due to additional equipment and piping modifications in the structure.

This efficiency—achieved with fewer computational resources—represents a major advancement in structural inspection technology.

Broader Implications

The study highlights the potential of deep learning to transform structural inspections. By integrating the SIE and SETA neural networks, the framework provides a streamlined and precise approach to assessing structural integrity. During model training, the SIE network achieved 95.77 % accuracy with a loss rate of 0.03, while the SETA network reached 68.97 % accuracy with a loss rate of 0.04.

Despite these promising results, the framework has limitations. It has only been tested on linear structural components, which restricts its generalizability. Additionally, the SETA network requires further refinement to enhance its ability to classify structural error types accurately.

Final Thoughts

The Distance-DNN-based Scan-vs-BIM framework is a breakthrough for steel structure inspections. It simplifies the process, delivering fast and accurate results that save time and resources. Traditional inspections can be slow and inconsistent, but this framework evaluates structural integrity in minutes, enabling quicker, more reliable decisions.

Its streamlined approach focuses on analyzing straightforward data—distances and indexes—through a deep learning network. This makes it practical for various projects, from small-scale construction to large facilities, without overburdening computational resources.

While the framework is already showing impressive results, there’s potential for growth. Expanding its use to more complex structures and fine-tuning its error classification could unlock even greater possibilities. It’s an exciting step forward, making inspections smarter, faster, and safer for everyone involved.

Journal Reference

Kim, B., Jo, I., Ham, N., & Kim, J. (2024). Simplified Scan-vs-BIM Frameworks for Automated Structural Inspection of Steel Structures. Applied Sciences, 14(23), 11383. DOI: 10.3390/app142311383, https://www.mdpi.com/2076-3417/14/23/11383

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.