Bolt Looseness Identification using Faster R-CNN and Grid Mask Augmentation

Conference proceedings article

Authors/Editors

Strategic Research Themes

Publication Details

Author list: Natchapon Panmatharit; Yuttapong Jiraraksopakun; Anek Siripanichgorn; Punnarai Siricharoen

Publication year: 2022

Start page: 1632

End page: 1637

Number of pages: 6

ISBN: 9786165904773

URL: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85146275720&doi=10.23919%2fAPSIPAASC55919.2022.9980142&partnerID=40&md5=f869be4c003fb2b3863c85a37a24bcd0

Languages: English-Great Britain (EN-GB)

View on publisher site

Abstract

The quality of bolted connections is important for modular steel construction. This paper presents an automated looseness detection of bolted joints using faster region-convolutional neural networks (Faster R-CNN) with grid mask augmentation. Faster R-CNN has an ability to accurately detect an object and identify the object class. Our application leverages this ability for distinguishing the looseness levels into visually tight, loose, and unidentified bolted connections. Grid mask and flipping augmentation is used for the model to be more accurate with small dataset of bolted joints. Faster R-CNN with ResNeXt-101 backbone with grid mask augmentation shows promising mean average precision of detecting joints which is increased by 9.6% from without augmentation. Compared with YOLOF, our Faster R-CNN with ResNeXt-101 backbone has higher mAP by 0.94% and the integrated system with cloud computing allows our system to practically work real time. © 2022 Asia-Pacific of Signal and Information Processing Association (APSIPA).

Keywords

bolted joint, bolt looseness, faster r-cnn, gridmask, quality identification