Abstract This research proposes a Ground Penetrating Radar (GPR) data processing method for non-destructive detection of tunnel lining internal defects, called defect segmentation. To perform this critical step of automatic tunnel lining detection, the method uses a CNN called Segnet combined with the Lovász softmax loss function to map the internal defect structure with GPR synthetic data, which improves the accuracy, automation and efficiency of defects detection. The novel method we present overcomes several difficulties of traditional GPR data interpretation as demonstrated by an evaluation on both synthetic and real data—to verify the method on real data, a test model containing a known defect was designed and built and GPR data was obtained and analyzed.
Introduction
Maintaining tunnel safety is a high-stakes engineering challenge. Tunnel linings often develop internal defects like cracks, voids, or "lining-rock separation" (gaps between the concrete and the surrounding earth) due to aging and geological stress. Ground Penetrating Radar (GPR) is the gold standard for non-destructive testing because it is fast and penetrates deep into structures. However, GPR does not produce a direct image of the interior; instead, it generates a "B-scan" filled with overlapping hyperbolic curves that represent reflected electromagnetic waves. Interpreting these scans currently requires highly trained experts, is prone to human error, and makes it difficult to determine the exact shape and location of a defect—especially when multiple defects or reinforcing rebars are present.
The Defect Segmentation Framework
The researchers developed an automated "defect segmentation" method that uses a Convolutional Neural Network (CNN) to transform complex GPR signals into clear, pixel-level maps of tunnel lining defects. Unlike traditional classification methods that merely identify the presence of a defect, this approach provides a pixel-by-pixel map of the defect's geometry.
Methodology: SegNet and Lovász Softmax
The authors chose SegNet, an encoder-decoder CNN, for this task. Unlike the popular U-Net, SegNet stores the indices of its max-pooling layers to perform non-linear upsampling. This helps the model maintain high-frequency structural integrity, which is vital for mapping the sharp boundaries of internal defects.
A critical innovation in this study is the use of the Lovász softmax loss function. Standard cross-entropy loss functions often fail when dealing with "imbalanced" data—for example, when a tiny crack (a few pixels) is surrounded by a massive block of concrete (thousands of pixels). By optimizing the Mean Intersection over Union (MIoU), the Lovász softmax loss significantly improved the model's ability to detect small, high-resolution features like rebars and thin cracks that are often "washed out" by standard training methods.
Data Generation and Simulation
Because real-world labeled data of tunnel defects is scarce, the team used Finite Difference Time Domain (FDTD) simulations to generate a massive synthetic training set of 124,800 models. These models covered nine categories, including rebar, concrete, rock, and various types of defects (cracks, voids, separations) both with and without water. The accuracy of these simulations depends on the physical properties assigned to the media.
To verify the results, a physical concrete testbed (4.4m x 2m x 0.7m) was built to simulate a real tunnel lining. PVC pipes were embedded to create known voids and separations, and real GPR data was collected using a 600MHz antenna.
Key Results and Discussion
The proposed SegNet with composite loss achieved a Mean Pixel Accuracy (MPA) of 93% and a Mean Intersection over Union (MIoU) of 90% on synthetic data. It significantly outperformed U-Net, which struggled with overfitting and failed to accurately classify complex overlapping signals, such as separations and cracks occurring at different depths.
One significant finding was the quantification of the rebar shielding effect. Rebars, due to their high conductivity and dielectric constant, create a shielding effect that can hide or distort defects located beneath them. The study showed that while the CNN can still classify these defects, the accuracy of their predicted shape drops significantly in the presence of rebar.
On the physical test model, the CNN correctly predicted the classification and location of water-bearing voids. While the predicted shape was less precise than in simulations due to real-world noise, the method proved its feasibility for field engineering.
Conclusion
This research demonstrates that deep learning-based semantic segmentation can effectively automate the interpretation of GPR data for tunnel maintenance. Future work will focus on improving noise sensitivity, as real-world GPR data is much noisier than FDTD simulations. The authors suggest that collecting more diverse "background signals" from various real-world environments will be essential to creating a more robust, "plug-and-play" model for different tunnel types.
Summary updated