Simplified Scan-vs-BIM Frameworks for Automated Structural Inspection of Steel Structures

Abstract

This paper presents a deep learning-based Scan-vs-BIM methodology for evaluating structural integrity through the extraction of features from As-Built scan and As-Planned Building Information Modeling (BIM) comparison data. Traditional Scan-vs-BIM frameworks often rely on Scan-to-BIM processes to generate point cloud-based mesh models for comparison, which significantly impairs computational efficiency. In contrast, the proposed streamlined Scan-vs-BIM framework incorporates a deep neural network (DNN) model consisting of two neural networks: one for structural integrity assessment and another for error type analysis. The model evaluates the structural integrity of individual components in a sequential manner, repeating the process across all elements to comprehensively assess the entire structure. Rather than converting point cloud data into mesh models for comparison, this approach directly measures the spatial discrepancies between the As-Built point cloud and As-Planned BIM, analyzing the distribution tendencies of these distance values. Experimental validation on actual steel structures demonstrated that the proposed method effectively predicts structural integrity, providing significant improvements in both accuracy and computational performance.