Multiclass damage detection in concrete structures using a transfer learning‐based generative adversarial networks

Abstract

A large amount of the world's existing infrastructure is reaching the end of its service life, requiring intervention in the form of structural rehabilitation or replacement. A critical aspect of such asset management is the condition assessment of these structures to evaluate their existing health and dictate the scheduling and extent of required rehabilitation. It has been demonstrated that human-based manual inspections face logistical constraints and are expensive, time extensive, and subjective, depending on the knowledge of the inspection. Recently, autonomous vision-based techniques have been proposed as an alternative, more accurate method for the inspection of deteriorating structures. Convolutional neural networks (CNNs) have demonstrated state-of-the-art accuracy with respect to damage classification for concrete structures and are often implemented to process images taken from vision-based sensors such as cameras, smartphones, and drones. However, these archetypes require a large database of annotated images to train the network to an accurate level, which is not readily available for real-life structures. Moreover, CNNs are limited to the extent by which they are trained; they are often only trained for binary damage classification of a singular material model. This paper addresses these challenges of CNNs through the application of a generative adversarial network (GANs) for multiclass damage detection of concrete structures. The proposed GAN is trained using the SDNET2018 dataset to detect cracking, spalling, pitting, and construction joints in concrete surfaces. Moreover, transfer learning is implemented to transfer the learned features of the GAN to a CNN architecture to allow for accurate image classification. It is concluded that, for a 0%-30% reduction in the amount of labeled data used, the proposed GAN method has comparable accuracy to traditional CNNs.

A literature review: Generative adversarial networks for civil structural health monitoring

“…In the civil SHM literature, there are also others [ (Wang et al, 2019;Rastin et al, 2021;Yuan et al, 2021;Yang et al, 2022;Huang et al, 2020;Sathya et al, 2020;Sun et al, 2022;Dunphy et al, 2022)] (a total of 10 studies) who are observed that do not fit a category and address different problems. It is observed that the subject of "damage detection after increased resolution" has one study (Sathya et al, 2020) in the literature where the researchers use SRGAN to increase the resolution of the images to improve the performance of the damage classifier.…”

“…There is one study on the topic of "track irregularity estimation" on acceleration data (Yuan et al, 2021), one study on the topic of "damage identification" via acceleration data (Rastin et al, 2021), and three studies on the topic of "Annotation reduction via transfer learning" [ (Huang et al, 2020); Dunphy et al, 2022)] which aim to reduce the need of annotating data through transfer-learning are other instances of GAN applications in civil SHM.…”

“…The authors observed significant improvement in the displacement calculation after the increased resolution of the images as increased resolution with GAN enables more feature points to be detected. Dunphy et al (Dunphy et al, 2022) discussed the importance of a large amount of annotated data points to train the DL networks, which is not readily available in real life. The study investigated the multiclass crack identification of concrete structures utilizing sparse amounts of annotated images.…”

Performance Comparison of Multiple Convolutional Neural Networks for Concrete Defects Classification

“…In contrast, multiple detections are essential to comprehend the actual scenario of damage condition of any structure. Even though some studies have worked with different types of damages [26,27], the image dataset is minimal for some cases, i.e., spalling, and rebar exposure. Moreover, most prior research lacks multiple CNN model analyses and detailed sensitivity analyses of hyper-parameters.…”

An ensemble approach for enhancing generalization and extendibility of deep learning facilitated by transfer learning: principle and application in curing monitoring

“…Wang et al [28] utilized TL to determine damage types caused by the repeated bearing of mechanical operations and natural factors, enhancing the service life of a pressure container. Dunphy et al [29] applied TL and generative adversarial networks for multiclass damage detection within infrastructures. Chamangard et al [30] reported successful TL application in accurately diagnosing invisible damage within complex structures.…”