For efficient maintenance of existing bridges, quantitative evaluation of residual load capacity using numerical analysis with FE models is widely adopted. Among various FE models, the fiber-based model, composed of center lines and cross-sectional geometries, is particularly effective for analyzing entire bridges. However, as-built drawings are often unavailable, and bridge conditions inevitably change over time, necessitating modeling methods that do not rely on design drawings. Point cloud data, capable of capturing as-is 3D geometry, have therefore attracted increasing attention. The authors have previously developed a method to semi-automatically construct fiber-based models from point clouds of entire steel truss bridges. However, a mismatch remains between the dimensional reproducibility required for load-bearing capacity analysis and the limitations imposed by measurement equipment and site conditions. While millimeter-level accuracy is required, measurement conditions are often constrained, resulting in incomplete or low-quality data. To address this issue, this study proposes an extended method incorporating deep learning. A measurement simulation tool is introduced to generate training data that reflect realistic laser scanner conditions, improving member recognition accuracy even in regions with poor measurement quality. Furthermore, deep learning enables the estimation of flange and web dimensions as well as plate thickness. As a result, the proposed method improves dimensional reproducibility and represents a step forward in capturing the overall structural configuration of entire bridges.
Hidaka et al. (Sat,) studied this question.