An intelligent method for leakage segmentation and area quantification to evaluate the safety performance of tunnels

Abstract

Water leakage in tunnels affects the structural stability and operational safety of tunnels, and timely inspections and repairs are needed. Traditional manual inspection methods are less automated and less efficient, which cannot meet the rapid inspection of tunnels. In this study, YOLO v8-DBC is proposed for tunnel water leakage segmentation and area quantification. In the backbone, a new Cross Stage Feature-Dynamic Snake Convolution (C2f-DSC) module is designed to improve the ability of the network to extract features of complex structures. C2f-DSC adaptively focuses on meandering and slender local structures, which better captures information about the edges of water leakage. In the neck, Bi-directional Feature Pyramid Network (BiFPN) enhances the fusion of deep and shallow information features. Convolutional Block Attention Module (CBAM) enhances the attention of the model to local and spatial information and reduces the interference of noise and irrelevant features. The results showed that YOLO v8-DBC segmentation is better than YOLO v8, YOLO v6, and YOLO v5, with the mAP50 (S) of 0.881 and the mIoU of 0.823. The area of tunnel leakage is calculated by the binary segmentation technique of the image. The tunnel water leakage hazard level is evaluated based on the area of leakage and technical code for waterproofing of underground works. The proposed method applies to the leakage segmentation and safety performance evaluation of tunnels, offering a reference for construction personnel in decision-making.