Simulation‐driven deep learning for rebar clutter elimination in ground‐penetrating radar images to detect backfill grout defects in segment linings

Inspecting the backfill grout behind segment linings using ground‐penetrating radar (GPR) is essential for the maintenance of shield tunnels. However, reinforcing rebars embedded in the segment linings generate strong clutter in GPR data, which obscures the detection of defect signals within the backfill grout. In addition, acquiring sufficient and consistent GPR data to train deep learning models is challenging due to restricted site access and variability in tunnel environments. To address these limitations, this study proposed a simulation‐driven deep learning network for clutter elimination and defect detection in GPR images. A training database was constructed exclusively through finite‐difference time‐domain numerical simulations to model segment linings containing backfill grout defects. This configuration provides a standardized and well‐controlled dataset for training and evaluating the network. Several architectures within the encoder–decoder framework, including U‐Net 3+, were employed to develop models for eliminating rebar clutter. The performance of the reconstructed GPR B‐scans was assessed using image quality and quantitative metrics, with U‐Net 3+ demonstrating the highest accuracy. The findings confirm that realistic GPR signal characteristics can be learned and generalized through simulation‐based data without relying on extensive field data. Finally, GPR B‐scans collected from a full‐scale tunnel lining segment were reconstructed using the proposed network to verify its practical applicability. This study demonstrates the potential feasibility of transfer learning from simulation‐only data to real‐world engineering applications, enabling more effective backfill grout inspection and supporting efficient maintenance.