AI Automates Steel Inspection with Scan-vs-BIM

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

A recent article published in Applied Sciences proposed using as-built scan and as-planned building information modeling (BIM) comparison data to prepare a deep learning-based Scan-vs-BIM framework for automated inspection of steel structures. The framework comprised a deep neural network (DNN) model with two neural networks, one each for structural integrity evaluation (SIE) and structural error type analysis (SETA).

Background

Integrity evaluation is critical to ensure the quality and safety of construction projects, particularly for facility management, new construction, safety inspections, repairs, and remodeling. However, traditional integrity evaluation in the construction industry is a manual process, depending on subjective observation, judgment, and expertise of skilled workers, often resulting in incomplete and inaccurate reporting.

To mitigate these issues, more efficient integrity evaluation methodologies have been developed using information and communication technologies. For instance, laser scanning has been widely adopted to collect data for integrity evaluations in construction projects. It is generally used to reconstruct as-built data, which is subsequently compared to as-planned data.

However, most point-cloud-based as-built verification methods depend on manually designed models of real-world structures. Consequently, there is an increasing demand for automated processing tools for enhanced “as-planned vs. as-built” comparison efficiency.

Methodology

The proposed distance-DNN-based Scan-vs-BIM framework involved the following steps: preprocessing, Scan-vs-BIM, and distance-DNN. The framework evaluated the structural integrity of a single structural object per process, which was repeated for all structural objects to assess the integrity of the entire structure. 

During preprocessing, a parameter-based bounding space of a BIM object was created for single-structure evaluation. This bounding space was the reference for separating the BIM shape model and sampling the points in a single operation. In the subsequent scan-vs-BIM process, the geometric relationship data of the point cloud and BIM were collected.

To overcome the slow speed of traditional processes, the proposed framework of this study performed integrity evaluation by collecting only distance and index data for direct comparison. Accordingly, in the distance-DNN process, the tendency of distance and index data was analyzed for SIE and SETA. A DNN model was constructed to perform these two functions.

Python programming language was used to implement the proposed framework. Meanwhile, the TensorFlow deep learning library was used to develop and train the distance-DNN model. Data preprocessing and manipulation were performed using NumPy and Pandas libraries, and the results were visualized using Matplotlib.

Real project data (26,500 datasets) and virtual data (65,000 datasets) were collected to train the distance-DNN model; these were used to train the function of real structural objects and structural error types, respectively. The performance of the model was evaluated in terms of the “accuracy” and “loss” based on the learning results.

Application Results

The Distance-DNN-based Scan-vs-BIM framework was used for the structural integrity evaluation of actual steel structures. The target steel structure had 184 structural objects covering a 423-m² area with a 10-m height. Three-dimensional scanning data (approximately 10 million points) were collected during the remodeling of this structure.

The integrity evaluation of the 20 structural column objects through the SIE and SETA networks took 42 ms. Accordingly, the evaluation rate for one object was 2.1 ms, faster than the traditional Scan-vs-BIM framework despite using fewer computational resources.

The SIE network predicted that all columns were 94.68% accurate (on average). Meanwhile, the SETA network revealed a high proportion of the “Error” label, attributed to the addition of equipment and pipes to the target structure. Therefore, the SIE and SETA networks could accurately evaluate the structural integrity.

The computational efficiency of the proposed framework was demonstrated by comparing its processing time to that of the traditional Scan-to-BIM method under the same conditions. Notably, the traditional method took approximately eight hours to process the entire structural dataset, while the proposed framework completed the same task in approximately three minutes, demonstrating a reduction of over 95% in processing time.

Conclusion

Overall, the researchers successfully developed a distance-DNN model for structural integrity evaluation through the comparison of as-built scan data and as-planned BIM data. In contrast to the traditional Scan-vs-BIM methods involving constructing comparative models and employing complex transformations to process them, the proposed framework integrated two unique neural networks (SIE and SETA) for a comprehensive structural integrity assessment.

During training, the SIE network attained an accuracy of 95.77%, while the SETA network achieved an accuracy of 68.97%. Their respective loss rates were 0.03 and 0.04, demonstrating minimal prediction errors. Additionally, the trained distance-DNN model demonstrated an accuracy of 94.2% in the structural integrity evaluation of an actual steel structure.

Despite the promising results, this study has some limitations. Primarily, the proposed framework has been validated on linear structural components, hindering its generalizability and practical usefulness. Additionally, the SETS network needs further improvement for the accurate classification of structural error types.

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.