Abstract
Segmental-lining tunnels are crucial underground infrastructure that face increasing inspection challenges due to urban expansion. While Terrestrial Laser Scanning (TLS) provides high-precision point clouds, leveraging these with deep learning (DL) for semantic segmentation is hampered by the scarcity of specialised annotated datasets and benchmarks. This research confronts this limitation by introducing Tunnel Scanner, a novel, automated, and parameterisable simulator that generates high-fidelity, labelled synthetic tunnel point clouds. Tunnel Scanner facilitates the creation of customised benchmarks and accelerates the generation of millions of labelled points, achieving over 95% time reduction compared to manual annotation. The research also demonstrates that pre-training DL models on this synthetic data, followed by fine-tuning on limited real data, substantially improves segmentation performance—boosting mean Intersection-over-Union (mIoU) and Overall Accuracy (OA) by approximately 20% compared to training solely on scarce real data. This work highlights the potential of point cloud synthesis and transfer learning in overcoming data limitations, paving the way for more generalised and efficient tunnel deformation detection methods suitable for integration into digital twin inspection systems.