PSI - Issue 44

2022 Agnese Natali et al. / Procedia Structural Integrity 44 (2023) 2020–2027 Agnese Natali et al. / Structural Integrity Procedia 00 (2022) 000–000 3 Each task needs to be treated separately in the development of the tool, although they are strongly interconnected, implying for each one: A.  Capability of catching the global configuration and assembly of the structure. Capability of distinguishing one component from the other one (beams, pillars, supporting devices, etc.) and capability of recognizing the critical ones, which are component that are prone to fragile failures that can induce collapse of the whole structure or of a part of it if hardly damaged (as dapped ends). Capability of associating to the element its material, so capability of individuating the right composing material. B.  Capability of recognizing each type of damage, depending on the material previously recognized (point A). C.  Capability of characterization of each detected damage (point B) in terms of extent (the size of the damage in proportion to the dimension of the component along which the damage propagates) and intensity (the level of development of damage in terms of degradation). D.  Considering the structural scheme (point A), the possible presence of critical components (point A) and the damages with their characterization and distribution (points B and C), capability of recognizing the presence of critical conditions, meaning that the damages detected may induce sudden or close strong damage or collapse (sudden in case of strong damage in critical components, close if damage is relevant but still monitorable to proceed with deeper analyses, as evaluation of the level of safety of the structure). The development of the tool is currently dealing with the tasks A and B, which are showed in the following. 3. Datasets and AI tool development The AI tool being developed uses a data driven approach. Existing works in this direction include Mundt et al. (2019), Özgenel (2019). In the initial phase of development, we relied on the datasets introduced by these works. The CODEBRIM dataset in Mundt et al. (2019) contains 1590 high resolution (6000x4000) images acquired from 30 unique bridges at different scales and locations. The defects included in this dataset are Corrosion Stain, Cracks, Efflorescence, Exposed Bars and Spallation. This is a challenging dataset because (a) one type of defect can overlap with another, (b) there are graffities that may partially overlap the defects, and moreover, (c) there are defect-free areas with dirt and markings that can be mistaken for defects. The dataset also provides random-sized cropped patches with various defects and the corresponding class labels that can be used to train a classifier. Examples of the cropped patches from this dataset are shown in Fig. 2(b). The second dataset that we consider is the Middle East Technical University (METU) dataset in Özgenel (2019) includes concrete images having cracks collected from various METU campus buildings. This dataset is generated from 458 high-resolution (4032x3024) images from which 227x227 sized patches are provided for crack and no crack classes, each having 20000 patches. The defect-free areas with dirt markings makes this dataset a challenging one for crack/no-crack classification. Examples of the cropped patches from this dataset are shown in Fig. 3.

Fig. 1. CODEBRIM dataset in Mundt et al. (2019). (a) Example of images available in the dataset, (b) cropped patches with various defects.

Fig. 2. Sample patches from METU dataset in Özgenel (2019) corresponding to the two classes, viz. cracks and no-cracks.