An Intelligent Detection Method for Precise Analysis of Shield Tunnel Lining Joints Through the Enhanced Faster Region-Based Convolutional Neural Network

Leakage between adjacent lining blocks is a common issue in shield tunnels, necessitating prompt joint inspection and repair to prevent secondary accidents. The challenge lies in swiftly and accurately locating these leakage joints. While Convolutional Neural Network (CNN)-based detectors are prevalent in tunnel object detection, Transformer-based detectors excel at extracting holistic image features. This study introduces an innovative framework for automatic and efficient detection of repaired and leakage joints in shield tunnels. Utilizing a Faster Region-Based Convolutional Neural Network (R-CNN) model predicts joint bounding boxes, and a novel backbone, integrating Swin Transformer and sandglass blocks, captures both holistic and regional features. To fuse these features, a Feature Pyramid Network (FPN) is employed, enhancing overall detection accuracy. Transfer learning mitigates data dependency in the Swin Transformer. Evaluation on a shield tunnel joints dataset demonstrates superior performance compared to original Faster R-CNN and one-stage networks. Ablation experiments analyse influence of various model settings on predictive performance. This framework not only addresses the challenge of rapid and precise leakage joint detection but also outperforms established models, showcasing its effectiveness in enhancing tunnel safety.