AI-Powered Damage Segmentation for Automated Condition Assessment of Historic Masonry-Lined Tunnels

Historic masonry tunnels form a substantial part of the underground railway infrastructure globally. To ensure their stability, regular condition inspections must be undertaken that typically involve a lengthy and subjective manual defect labelling process. Surface damages can be extensive on the lining of historic masonry tunnels. These are the multiple different types of surface damage such as spalling, efflorescence, degraded mortar and biological growth. This thesis investigates the potential of automated methods to improve the manual procedure typically employed within industry for masonry lined tunnel condition assessment.
Masonry spalling severity, defined by the depth of spalling, is a key indicator of a masonry tunnel’s condition. This study proposes an automated workflow for identifying spalling severity from 3D tunnel lining point cloud data obtained by lidar. The workflow combines using deep learning for masonry joint and damage semantic segmentation with a geometric undamaged masonry plane fitting procedure to recreate the tunnel lining without surface damages. The study first uses synthetic data to investigate the workflow’s limitations before analysing the most effective algorithms for masonry joint and damage segmentation. Supervised convolutional neural networks were selected for masonry joint and damage segmentation. Training with topological loss functions and a transformer-based encoder were shown to improve performance. Using masonry joint segmentation as a proxy for the automated workflow’s performance, possible solutions to the challenges of generalising the method between different tunnels, quantifying uncertainty and determining the optimal training method are presented. Training a method on different tunnels alongside targeted tunnel specific training is shown to achieve the best performance.
The study finally evaluates the trained spalling severity segmentation algorithm on multiple real-world tunnels, demonstrating effective masonry spalling localisation. The workflow’s robustness enables it to always provide a useful indication of potentially damaged areas that a human assessor can then analyse in more depth. The presented work shows how an automated method using deep learning can be integrated into routine masonry tunnel condition assessments to reduce analysis time and generate more comprehensive spalling maps.