Deep learning-based photogrammetry monitoring of large-scale underground infrastructure at CERN

Simple item page
dc.check.date2031-09-30
dc.check.infoControlled Access
dc.contributor.advisorLi, Zili
dc.contributor.advisorexternalOsborne, John Andrew
dc.contributor.authorOuyang, Aohuien
dc.contributor.funderScience Foundation Ireland
dc.contributor.funderCERN
dc.contributor.funderIrish Centre for Applied Geoscience (iCRAG)
dc.date.accessioned2026-02-16T14:06:44Z
dc.date.available2026-02-16T14:06:44Z
dc.date.issued2025
dc.date.submitted2025
dc.descriptionControlled Access
dc.description.abstractIn many countries worldwide, thousands of miles of existing underground tunnel infrastructure are facing consistent structural degradation. The structural deterioration of existing tunnels has been increasingly evidenced by crack growth, water leakages and other defects arising from lining material ageing, drainage blockage, etc. One representative tunnel network is the underground infrastructure at the European Organization for Nuclear Research (CERN), which spans over 83km, hosting the largest particle physics underground laboratory in the world. Built in the 1970s, CERN’S underground infrastructure network shows clear signs of structural degradation, combined with complex geological conditions that pose a risk to its serviceability. As integral part of Tunnel Asset Management (TAM) tasks, periodic inspections and the documentation of tunnel defects can provide insights for early-stage structural remediation prior to escalation into more severe damage. However, some challenges, such as the vast scale of infrastructure, the radioactive environment, and limited inspection windows, pose great challenges to traditional, manual visual inspection. This thesis develops a computer-vision-based inspection method for monitoring the tunnel defects at CERN, addressing various scales from millimetre-scale cracks (at the pixel level) to tens-of-centimetre water leakages. To conduct the pixel level crack monitoring in large-scale CERN underground infrastructure, a remote and automated system is developed based on the inhouse CERNBot, using robot-mounted imaging technology. This system can collect crack images remotely and stitch them together to create a panorama image of the tunnel surface. Employing transfer learning, the benchmark semantic segmentation model, the DeepLab V3plus, is finetuned to detect cracks automatically. A novel smooth blending prediction method is implemented on the panorama to present long-distance tunnel crack distribution, alleviating misclassification problems encountered in high-resolution image inference. In addition, transfer learning, tailored loss functions, and the regularization techniques have been developed based on the CERN tunnel crack database characteristics to maintain high performance and generalization of the proposed method. The proposed monitoring system was applied to two typical tunnel sections, creating long tunnel panoramas over 100 meters. The resultant crack density distribution allows identification of critical crack-damaged tunnel sections. On the other hand, the detected crack spatial distribution reveals the long-term mechanism of deformation for different tunnel sections at CERN: the transversal shearing mode is prevalent in the Large Hardon Collider (LHC) tunnel section, whereas the vertical elongation of the tunnel cross section is the dominant deformation mode in an inclined TT1 tunnel. Towards demarcation of water leakage defects in the LHC tunnel, a computer vision-based method is developed based upon the Tunnel Inspection Monorail (TIM) platform. Equipped with 360-degree camera, the customized TIM is deployed for the remote video data acquisition. A customized Omnidirectional Image (ODI) processing pipeline is proposed to transform the gathered omnidirectional video into two-dimensional tunnel mosaics with units of 10-metre length. Furthermore, an ensemble deep learning strategy leveraging transfer-learning is employed to generate a scaled demarcation map of tunnel leakage areas. The proposed method was employed to inspect a 3.45km-long tunnel alignment, spanning from LHC point 4 to point 3. The spotted water leakage distribution identifies critical leakage zones along the extended tunnel alignment and reveals spatial leakage patterns primarily attributed to accumulated water pressure behind the lining, caused by drainage system clogging due to calcite deposits. Using the proposed system, the spatial distribution of the detected drainage plates suggests that the construction joints are particularly vulnerable to tunnel waterproofing degradation, after decades of operation. To achieve accurate quantification and documentation of refined tunnel defects, especially cracks, a digitalization method tool is also developed on a detailed level, enabling 3D visualization and logging into structured data. The method reconstructs sparse point clouds with Structure from Motion (SfM) and cleans out the irrelevant tunnel facilities using a two-stage filtering method. The denoised 3D point clouds are then fitted with customised meshes and textured into 3D reconstruction models. The flat scaled orthomosaic is generated by the cylindrical unrolling. Deep learning methods are also employed for pixel-level crack detection in this high-resolution image, enabling precise extraction of crack locations and quantification. Applied to four different CERN tunnel sections, the method presents the spatial crack distributions in 3D space around the tunnel circumference and quantifies crack dimensions. Furthermore, the digitalization method provides a statistical tool to investigate the crack geometry characteristics in a whole cracked section. Statistical analysis shows that the crack occurrences decrease with increasing crack lengths, largely following a log-normal distribution. The corresponding fitting parameters can serve as quantification indicators of crack length characteristics All these proposed methods offer visualization tools for engineers to inspect the tunnel asset remotely, presenting the tunnel lining surface in consecutive panoramas. The developed image-based deep learning algorithms facilitate the automated tunnel defect visual inspection and digitalize the tunnel defects into structured data format. In field applications, the proposed approaches enhance tunnel asset management (TAM) by identifying severely defect-affected tunnel sections, revealing the spatial distribution patterns of defects over a larger Field of View (FoV), and generating digital representations of deteriorated tunnel areas. Furthermore, the application of the proposed approaches offers potentials to investigate tunnel defects from a statistical angle. The subsequent statistical analysis reveals inherent defect characteristics and provides engineering insights in long term tunnel ageing behaviour.en
dc.description.statusNot peer revieweden
dc.description.versionAccepted Versionen
dc.format.mimetypeapplication/pdfen
dc.identifier.citationOuyang, A. 2025. Deep learning-based photogrammetry monitoring of large-scale underground infrastructure at CERN. PhD Thesis, University College Cork.
dc.identifier.endpage220
dc.identifier.urihttps://hdl.handle.net/10468/18530
dc.language.isoenen
dc.publisherUniversity College Corken
dc.relation.projectinfo:eu-repo/grantAgreement/SFI/Research Centres Programme::Phase 2/13/RC/2092_P2/IE/iCRAG_Phase 2/en
dc.rights© 2025, Aohui Ouyang.
dc.rights.urihttps://creativecommons.org/licenses/by-nc/4.0/
dc.subjectTunnel defect monitoringen
dc.subjectCrack detectionen
dc.subjectTunnel leakage detectionen
dc.subjectImage-based deep learningen
dc.subjectTunnel deteiroation patternen
dc.subjectPhotogrammetry modellingen
dc.subjectOmnidirectional imagingen
dc.subjectRobotic inspectionen
dc.subjectTunnel defect digitalizationen
dc.subjectTunnel defect statisticsen
dc.subjectTunnel visual inspectionen
dc.titleDeep learning-based photogrammetry monitoring of large-scale underground infrastructure at CERNen
dc.typeDoctoral thesisen
dc.type.qualificationlevelDoctoral
dc.type.qualificationnamePhD - Doctor of Philosophy
Files
Original bundle
Now showing 1 - 2 of 2
Thumbnail Image
Name:
OuyangA_PhD2026.pdf
Size:
12.06 MB
Format:
Adobe Portable Document Format
Description:
Full Text E-thesis
Thumbnail Image
Name:
OuyangA_PhD2026.docx
Size:
66.97 MB
Format:
Microsoft Word XML
Description:
Full Text E-thesis (Word)
License bundle
Now showing 1 - 1 of 1
Thumbnail Image
Name:
license.txt
Size:
5.2 KB
Format:
Item-specific license agreed upon to submission
Description:
Download
Collections
Research Theses
Civil and Environmental Engineering - Doctoral Theses
College of Science, Engineering and Food Science - Doctoral Theses