Application of deep learning techniques for automated monitoring of railway infrastructure assets

• Autores: Javier Grandio González
• Directores de la Tesis: Belén Riveiro Rodríguez (dir. tes.), Pedro Arias Sánchez (dir. tes.)
• Lectura: En la Universidade de Vigo ( España ) en 2025
• Idioma: inglés
• Materias:
  • Ciencias tecnológicas
    • Tecnología industrial
      • Ingeniería de mantenimiento
    • Tecnología de los ferrocarriles
      • Equipo ferroviario
• Enlaces
  • Tesis en acceso abierto en: Investigo
• Resumen

  • Transport infrastructure is a critical foundation of modern society, yet its maintenance and monitoring present significant challenges. This PhD research addresses these challenges by developing automated, data-driven methodologies for infrastructure digitalization and health monitoring. Central to these methodologies is the deployment of state-of-the-art deep-learning techniques specifically tailored to solve complex infrastructure-monitoring problems. The work is structured around two complementary thrusts: (1) automated 3D modeling of transportation assets (with a focus on railways) from LiDAR point cloud data using deep learning, and (2) AIbased structural health assessment for critical structures (bridges). In the first thrust, novel deep learning frameworks were created for semantic and panoptic segmentation of large-scale railway point clouds, enabling the identification of tracks, cables, masts, and other assets with high accuracy. A multimodal approach fusing image and point cloud data was introduced to improve efficiency, doubling the processing speed of previous methods and achieving over 90% classification accuracy in diverse scenarios. In the second thrust, a Point Transformer Network surrogate model was developed to predict bridge deflections from geometric data, yielding submillimeter accuracy (MAE < 0.0213 mm) while vastly reducing computation time compared to finite element simulations. Additionally, a machine learning-based early warning system was validated on an 1897 steel truss bridge, detecting damage scenarios with 90.8% recall. The findings demonstrate that combining advanced sensing (mobile mapping, vibration monitoring) with deep learning can greatly enhance infrastructure digital twins. This enables more proactive maintenance, improved safety, and integration with Building Information Modeling (BIM) workflows. Overall, the contributions of this thesis push the state-of-the-art in automated infrastructure modeling and health monitoring, providing a foundation for smarter asset management and future research in civil infrastructure digitalization.