PSI - Issue 84

Giuseppe Santarsiero et al. / Procedia Structural Integrity 84 (2026) 313–320

314

Nomenclature AI

Artificial Intelligence

CoA Class of Attention CSLLPP High Council of Public Works (Italy) CV Computer Vision DL Deep Learning FN False Negative I-PRC I-shaped Prestressed Reinforced Concrete IoU Intersection over Union LG2020

Italian Guidelines for Risk Classification and Management of Existing Bridges

LLM MIT PRC

Large Language Model

Ministry of Infrastructure and Transport (Italy)

Prestressed Reinforced Concrete

RC

Reinforced Concrete

ReLUIS YOLO

Network of University Laboratories for Seismic and Structural Engineering

You Only Look Once

1. Introduction The ageing of existing bridge infrastructure and the increasing attention to structural safety require road authorities and owners to implement systematic inspection, monitoring, and maintenance strategies. In Italy, this need has been strongly reinforced by the introduction of the Italian Guidelines for Risk Classification and Management of Existing Bridges (LG2020), which define a multi-level, multi-risk assessment framework (MIT 2020, updated in MIT 2022). At the first level, the process is strongly dependent on visual inspections, during which large quantities of photographic material and inspection records are collected and subsequently processed to derive the so-called Class of Attention (CoA), used as an indicator of potential structural risk. However, this procedure is still largely manual, time-consuming, and subjective, since the recognition and classification of defects rely on the experience and judgment of individual inspectors. Moreover, the photographic material is typically not acquired with the specific purpose of automated processing, leading to heterogeneous quality, perspectives, lighting conditions, and framing, which further increases interpretation variability. In parallel, Deep Learning (DL) and Computer Vision advancements offer promising opportunities to support engineers in these tasks (Dogan et al. 2023). In particular, object detection networks such as YOLO (You Only Look Once) (Redmon et al. 2016) have shown strong performance in identifying and localizing visual patterns in real-world scenes, including civil engineering structures. Their ability to operate in real time and on large image datasets makes them suitable candidates for assisting inspection workflows by automatically recognizing surface damage and linking it to code-based defect taxonomies. However, a few researchers already attempted to develop multiclass damage detection algorithms devoted to operate on concrete structures (e.g. Iraniparast et al., 2025; Liu et al., 2025) or historical structures (Karimi et al., 2024). At the University of Basilicata, a broader AI-assisted inspection platform is currently under development to support bridge inspectors through a combination of deep learning models for visual damage detection on RC/PRC structural elements, and engineering-constrained AI tools for the identification and assessment of bridge bearings, based on an integrated DL + LLM (Large Language Model) approach. These two components are conceived as complementary modules of a unified framework, sharing the same ultimate objectives: • reduce inspector subjectivity; • standardize defect terminology and classification; • enable structured data extraction aligned with LG2020 inspection forms;