PSI - Issue 62

Sergio Ruggieri et al. / Procedia Structural Integrity 62 (2024) 129–136 Author name / Structural Integrity Procedia 00 (2022) 000 – 000

131

3

transitability, operativity) and structural monitoring is recommended. In the case of low risk, the bridge is currently safe and can be checked through the procedures at level 1 after some time (guidelines suggest an interval period between consecutive inspections of 6 months). It is clear that, already from the first levels, to faithfully employ the approach proposed by guidelines, several conditions should be respected. First, for level 0, documentations could be available for newest structures, but for older bridges the original drawings are usually missing in the archives. In addition, structural modifications and interventions could not be traced by the management companies, which started their work after some years from the construction of the bridge. When going to level 1, it is worth considering that usually few well-instructed surveyors are involved in the inspections of hundreds of bridges to check in a short time, which implies high costs for this task. As a matter of fact, to perform a careful inspection, a surveyor needs o f a working day for checking “only” 4 -5 bays. This situation raises other several issues, as the lapses in attention in the inspection campaigns, which increases the subjectivity by surveyors in expressing defects extent and severity. In addition, it is worth considering the inaccessibility of some bridge components, such as the supports. In light of these problems, it is evident that a support to the inspection operations is necessary. With this regard, the paper presents a study on the use of a new technology AI-based, aimed at detecting typical defects in RC bridges and exploiting a DL approach. In detail, the paper discusses the use of a single-stage detector technique, named YOLO, for purpose of defect identification within images. First, a description of the training/test/validation dataset is provided, which comprises a set of real-life images of old RC bridges. After, several YOLO-based architectures were tested, using different typologies of the version 5. The experimental campaign reveals the architecture reporting the best performance and some examples in terms of prediction are provided. Pros and cons of the proposed approach are provided, showing the possible margins of improvement that a YOLO-based technique can apport in the field of bridge visual inspections. 2. State-of-the-art: AI and DL in the field of defect detection in existing RC bridges In the recent years, several new computer-aided techniques were involved in structural engineering problems, aimed at solving complex problem and establish new input-output relationships. Among these new approaches, DL was widely used, trying to retrieve information from images. As an example, some of the authors of this paper proposed VULMA (Ruggieri et al., 2021), a tool for defining a simplified vulnerability index calibrated on typological features retrieved from imagery of the existing building stock. With this purpose, a dataset was collected (Cardellicchio et al., 2022a) via Google Street View service and about 13 features were labelled in order to train a cascade of CNNs to automatically recognize features (for the case at hand, transfer learning and fine tuning were adopted). Analogously, also in the field of existing bridges, DL techniques were employed for developing interesting tools, especially aimed at defect detection. The earlier approaches regarded the implementation of standard image processing algorithms. For example, Yang et al. (2015) developed an end-to-end framework for predicting cracks based on image rectification. Authors tested the approach on a full-scale RC bridge and identified thin cracks, up to 0.2 pixels. Li et al. (2014) proposed a three-step method for long-distance image acquisition and crack identification. Image processing techniques were used to extract cracks and distance algorithm was used to compute the width of each crack. The approach was tested on about 1000 images, and a misclassification rate of 6.58% was achieved. Modern approaches implemented CNNs, trying to reduce the computational efforts by using other techniques (e.g., fine tuning, transfer learning). Talking about the use of CNNs for purpose of defect detection, Zhang et al. (2017) proposed CrackNet, a tool for identifying asphalt surface cracks, by employing a special type of CNN that did not pooling layers. Results showed an accuracy of about 90% in a dataset of few images (around 2000). Li et al. (2018) proposed DDLNet, a tool able to achieve a detection accuracy of around 80%, using a small dataset composed by 823 images. Cha et al. (2017) used different CNNs on a dataset of 40000 images, with the aim of predicting five typologies of cracks. Zhu et al. (2020) adopted transfer learning to train CNNs models for predicting different defects on few images. Authors achieved an accuracy of about 98% by using only 134 images. Recently, Cardellicchio et al. (2023a) proposed an extensive defect recognition, aiming at predicting six common typologies, by using a first dataset investigated in Cardellicchio et al. (2022b). Eight types of CNNs were employed, showing the performances of each network, and providing some novel metrics for purpose of explainability. However, interesting outcomes can be achieved by using object detection approaches, which employs a different concept from the CNNs, because specific defects are