In a groundbreaking study published in Applied Sciences, researchers have proposed a deep learning-based Scan-vs-BIM framework for the automated inspection of steel structures. This innovative approach compares as-built scan data with as-planned BIM, Building Information Modeling, data to assess structural integrity and identify errors, offering a more efficient and accurate alternative to traditional inspection methods. The framework, which relies on deep neural networks, represents a significant leap forward in construction technology, enabling faster, more reliable structural evaluations.
Ensuring the integrity of structures is paramount in construction, particularly for tasks such as facility management, new construction, safety inspections, repairs, and remodeling. Traditionally, structural integrity evaluations have been carried out manually, relying heavily on subjective assessments and the expertise of skilled workers. However, this method often leads to incomplete or inaccurate results, increasing the risk of overlooked structural issues.
To overcome these challenges, the construction industry has increasingly turned to Information and Communication Technologies (ICT), such as laser scanning, to collect accurate data for as-built verification. Laser scanning allows the collection of point cloud data, which is then compared with the as-planned BIM data to identify discrepancies. However, the majority of point-cloud-based verification methods still depend on manually constructed models of real-world structures, necessitating the development of more automated and efficient processing tools.
The proposed distance-DNN-based Scan-vs-BIM framework offers a streamlined process for automated structural inspection. It consists of three key stages: preprocessing, Scan-vs-BIM, and distance-DNN. The framework operates by evaluating the integrity of a single structural object at a time, repeating this process for every object in the structure to produce a comprehensive evaluation of the entire building or structure.
In the preprocessing stage, a bounding space for each BIM object is created, serving as a reference for isolating and sampling points from the 3D scanned data. This ensures accurate comparisons between the as-built and as-planned data during the Scan-vs-BIM step, where geometric relationship data is extracted from the point cloud and BIM models.
To address the inefficiencies of traditional methods, which require extensive computational power, this framework focuses on the distance and index data, allowing for direct comparison without the need for complex models. The distance-DNN process then analyzes these data points to evaluate structural integrity and classify error types.
The entire framework is implemented using the Python programming language, with TensorFlow for deep learning, and NumPy, Pandas, and Matplotlib for data manipulation and visualization.
To validate the framework, the researchers used real project data, which included 26,500 datasets from actual steel structures and 65,000 datasets from virtual simulations. These datasets were used to train the DNN model, enabling it to learn to classify structural objects and error types.
The target structure for the case study was a steel structure with 184 structural objects across a 423 m² area and 10-meter height. Approximately 10 million points of 3D scan data were collected during the remodeling phase of the structure.
Using the SIE (Structural Integrity Evaluation) and SETA (Structural Error Type Analysis) networks, the integrity of 20 structural column objects was evaluated in just 42 milliseconds. The result showed that the evaluation rate for each column object was just 2.1 milliseconds, far outpacing traditional Scan-vs-BIM methods, which typically took around eight hours to process the entire dataset.
The SIE network predicted an average accuracy of 94.68% for the columns, while the SETA network detected a significant number of "Error" labels, attributed to additional equipment and pipes in the structure. This demonstrated the ability of the framework to accurately identify and classify structural issues.
The proposed framework demonstrated exceptional computational efficiency compared to traditional methods. While the traditional Scan-vs-BIM method took approximately eight hours to process the data for the entire structure, the new framework completed the same task in just three minutes, reducing processing time by over 95%. This substantial improvement in speed, coupled with its accuracy, makes the proposed framework a game-changer for automated structural inspection.
The SIE network of the distance-DNN model achieved an accuracy of 95.77%, while the SETA network achieved 68.97% accuracy. The corresponding loss rates for the networks were 0.03 and 0.04, respectively, indicating minimal prediction errors. The model’s overall accuracy in evaluating the structural integrity of an actual steel structure was 94.2%, showcasing its robustness and reliability.
Despite the promising results, the study highlighted some limitations. The framework was validated primarily on linear structural components, and its generalizability to more complex structures remains a challenge. Additionally, the SETA network requires further refinement for more accurate classification of error types.
The Scan-vs-BIM framework, powered by deep learning and distance-DNN models, offers a breakthrough in the automated inspection of steel structures. By comparing as-built scan data to as-planned BIM data, this system can evaluate structural integrity and identify error types with remarkable speed and accuracy. This study provides a significant step towards fully automating the inspection process, reducing manual labor, and improving the efficiency of structural evaluations in the construction industry.
