Underwater inspection of bridge substructures using sonar and deep convolutional network. (April 2022)

• Record Type: Journal Article
• Authors: Hou, Shitong; Jiao, Dai; Dong, Bin; Wang, Haochen; Wu, Gang
• Is Part Of: Advanced engineering informatics. Volume 52(2022)
• Publication Date: 2022-04
• Subjects: Underwater bridge substructures -- Scour depth -- Apparent damage measurement -- Semantic segmentation networks -- Side-scan sonar; Computer-aided engineering -- Periodicals; Engineering -- Data processing -- Periodicals; 620.00285
• DOI: 10.1016/j.aei.2022.101545
• Languages: English
• ISSNs: 1474-0346

Abstract

Abstract: Bridges are vital structures for worldwide physical infrastructure networks, and efficient inspection methods are needed to reveal and evaluate the health conditions of damaged critical bridges. In contrast to the superstructure inspection, underwater inspection methods are still time-consuming and labor-intensive tasks. As the underwater environment is invisible and hard to access, scouring and damage are easily neglected in traditional human diving. To address this problem, this paper presents a rapid underwater inspection framework using a sonar device and a deep convolutional network to provide quantitative measuring results for scour depth and apparent damage. The side-scan sonar and fixing platform were used to collect underwater images, and a modified semantic segmentation was designed based on the U-Net architecture. The pretrained residual building blocks and a designed feature fusion connection called Respath were utilized to replace the original encoder and skip connection. Comparative experiments showed that this modified architecture achieved the best performance in multiple-class segmentation for sonar images in which the mean intersection over union (mIoU) and IoU for specific apparent damage can reach up to 0.918 and 0.63, respectively. An on-site test was also conducted to validate the applicability of the proposed method, and the scour depth and two points of damage were identified based on the pixelwise segmentation images. Therefore, further structural evaluation can proceed based on these accurate measuring results. Graphical abstract: Highlights: An underwater inspection framework is proposed for scour depth and apparent damage measurement. An U-Net-based convolutional network is proposed for accurate object detection and segmentation. The mIoU and IoU for damage can reach up to 0.918 and 0.63, and an on-site test was implemented.