Deep Learning-Based Classification of Defective Tunnel Linings

Summary This study introduced a novel method for identifying defective tunnel lining sections through an enhanced EfficientNetV2 model, which has been optimised with EMA and transfer learning techniques to improve the distinction between highly similar scenarios. Ablation experiments demonstrated that the accuracy of defective tunnel classification has been improved to 92.9%. With the proposed method, the tunnel lining GPR measurements are gone through a classification process, in which defective radargram segments are identified and the healthy ones are excluded. This approach reduces the number of radargrams requiring interpretation, significantly lowering the workload and improving diagnostic efficiency. Additionally, it helps avoid artefacts and pseudo-noise that intepreation may introduce in healthy tunnel sections.