MACHINE LEARNING TECHNIQUES FOR THE CREATION OF BRIM/FEM MODELS APPLIED TO BRIDGES

Bridge and infrastructure management increasingly relies on accurate, up‑to‑date digital representations to support inspection, maintenance planning and structural assessment. Modern surveys based on terrestrial laser scanning (TLS), Unmanned Aerial Vehicle (UAV) photogrammetry and, when available, mobile mapping systems (MMS) provide dense 3D point clouds that describe the “as‑is” condition with metric rigor, but they are inherently unstructured: geometry is available, while semantic information and interoperable model entities are not. This thesis addresses the transformation of raw point clouds into structured, semantically meaningful information that can be consumed by Bridge Information Modelling (BrIM/InfraBIM) and Finite Element Method (FEM) workflows. A complete, computable pipeline is proposed and formalised, covering: (i) acquisition and pre‑processing (registration, georeferencing and quality control); (ii) scalable data structures for neighbourhood queries; (iii) multi‑scale extraction of geometric descriptors; (iv) supervised semantic classification of point clouds; and (v) downstream modelling and export towards IFC‑based BrIM and FEM environments. The classification stage is built around a Random Forest (RF) model trained on local geometric features computed at multiple scales. The feature set includes spectral descriptors derived from eigenvalues (linearity/planarity/sphericity, omnivariance), curvature‑related measures, verticality, and height‑based attributes (e.g., z‑height / height‑above‑ground). To reduce manual labelling effort while preserving control on ground truth quality, a semi‑automatic iterative workflow is adopted: a manual seed is produced in CloudCompare, an initial RF is trained, predictions are propagated with confidence thresholds, and residual ambiguities are corrected in successive iterations. Class imbalance is handled through balanced class weights and controlled sampling. The workflow is validated on two bridge case studies: a training project (P3) and an independent validation project (P4). On the training bridge, the RF model reaches overall accuracy OA = 0.892 with macro‑F1 = 0.859 and Cohen’s kappa = 0.841, showing stable discrimination for the main structural classes (e.g., piles, deck slab, longitudinal beams). On the independent bridge, the model generalises without re‑tuning, achieving Accuracy = 0.86, macro‑F1 = 0.84, kappa = 0.83 and mean IoU = 0.78. Error analysis through confusion matrices highlights typical failure modes related to thin elements, local occlusions and scale mismatch in neighbourhood definition. Feature‑importance rankings confirm the dominant role of height‑based attributes and of multi‑scale spectral descriptors in separating context (ground/vegetation) from structural components and in resolving planar/linear patterns. After semantic segmentation, the thesis discusses how classified point clouds can be converted into modelling primitives suitable for BIM and FEM. Two complementary modelling strategies are compared: Track A, a section‑based approach in Rhinoceros/Non Uniform Rational B-Spline (NURBS) oriented to controllable, “clean” solids/surfaces and robust IFC export; and Track B, a parametric approach in Grasshopper oriented to rapid updates, variant generation and mesh‑ready geometries for FEM. Practical quality‑assurance checks (overlay and cloud‑to‑model comparisons, class‑to‑IFC mapping, and go/no‑go criteria) are proposed to ensure metric coherence and traceability throughout the conversion chain. Overall, the thesis contributes an end‑to‑end, reproducible and interoperable pipeline from survey data to BrIM/FEM‑ready models, together with operational guidelines to balance accuracy, effort and export reliability when moving from point‑based geometry to information‑rich digital representations of bridges. Point cloud; TLS; UAV photogrammetry; supervised classification; multi‑scale geometric features; Random Forest; BrIM/InfraBIM; IFC 4.3; FEM; Rhinoceros/Grasshopper.