Structure-Aware Semantic Segmentation and Automated Geometric Analysis Framework for Steel Plate Girder Point Clouds

Accurate geometric information of steel plate girders is essential for structural assessment and lifecycle management, yet reliable as-built data is often unavailable due to missing drawings or discrepancies between design and in-service conditions. This study proposes a unified framework for automated component-level characterization of steel plate girders from raw point clouds, covering both in-plane dimensions and out-of-plane web initial imperfections. A structure-aware semantic segmentation strategy embeds the vertical hierarchy of key load-bearing components directly into the loss function, improving structural consistency without modifying the network architecture and remaining compatible with state-of-the-art backbones. Subsequently, geometric parameters are extracted through noise filtering, principal component analysis–based alignment, and percentile-based three-dimensional bounding box computation. For web imperfection analysis, segmented stiffener labels define panel boundaries automatically, a global neutral surface is estimated using Huber-robust polynomial fitting, and per-panel deviation fields are represented by two-dimensional Fourier sine series to derive chord-referenced amplitudes. Experiments on real-world data demonstrate reliable segmentation and accurate dimension estimation. The extracted imperfection amplitudes are consistent with fabrication tolerances, supporting drawing-free inspection and digital reconstruction of steel bridges.