Segmental-lining tunnels are critical components of urban infrastructure, requiring regular inspection and maintenance to ensure safety and longevity. Point clouds from LiDAR scans have emerged as an effective and scalable alternative for automated health monitoring of infrastructure. Identifying semantic information from point clouds is the first step towards comprehensive geometric and structural evaluation. Deep learning has demonstrated promising efficacy for segmentation tasks across various domains, including tunnel point cloud semantic segmentation [1]. However, its success relies on the availability of substantial high-quality annotated datasets, which are typically lacking in tunnel infrastructure. This research addresses this gap in data scarcity by introducing 'Tunnel Scanner': a novel, parameterisable simulator that automatically generates labelled synthetic tunnel point clouds with high-fidelity structural geometries and tunnel environments. To bridge the domain discrepancy between synthetic and real data, the study also explores transfer learning techniques for deep learning training. Results demonstrate that pre-training on synthetic data, followed by fine-tuning on real data, can be highly beneficial, particularly when real data is scarce or of low quality. The findings demonstrate the potential of synthetic data generation and transfer learning in overcoming data scarcity challenges. This work paves the way for a more generalised and efficient method of tunnel deformation detection, which can be integrated into digital twin inspection systems for safety and maintenance practices in tunnels.