Deep learning-based automatic recognition of water leakage area in shield tunnel lining. (October 2020)

• Record Type: Journal Article
• Authors: Xue, Yadong; Cai, Xinyuan; Shadabfar, Mahdi; Shao, Hua; Zhang, Sen
• Is Part Of: Tunnelling and underground space technology. Volume 104(2020)
• Publication Date: 2020-10
• Subjects: Mask R-CNN -- Data augmentation -- Transfer learning -- Cascade strategy -- Water leakage area; Tunneling -- Periodicals; Underground construction -- Periodicals; Tunnels -- Periodicals; Underground areas -- Periodicals; 624.193
• DOI: 10.1016/j.tust.2020.103524
• Languages: English
• ISSNs: 0886-7798

Abstract

Highlights: A deep learning-based model is constructed for the pixels segmentation of water leakage. Three methods are utilized to improve segmentation performance and nearly double the average accuracy of the original model. A calibration relationship is obtained between the number of pixels and true area. Using the calibration relationship, the safety situation of shield tunnel can be evaluated by means of one image. Abstract: Difficulties in visual inspection of metro shield tunnels induced by various on-site installations (e.g., lighting instruments and pipelines) have been resolved by two-stage structural damage detection techniques based on deep learning known as the state-of-the-art on surface defects recognition. However, thresholds of the Intersection over Union (IoU) are set extremely low in two-stage segmentation models to obtain more positive samples to suppress overfitting during the training stage. Therefore, the pixel segmentation results generally contain lots of noise and are difficult to meet the requirements of engineering. Accordingly, this study adopted (1) data augmentation to increase the number of positive samples, (2) transfer learning to improve the robustness of convolutional layers, and (3) cascade strategy to enhance the quality of samples. The segmentation results of the improved model on the validation set demonstrated that the above methods achieved high precision pixel segmentation of water leakage (AP0.5 = 0.806), which was greater than the classical deep learning model (i.e., Mask R-CNN with AP0.5 = 0.530). Given the fact that the number of water leakage pixels cannot be regarded as reliable metrics to evaluate the security situation of the shield tunnel, a series of field experiments were conducted to obtain the calibration relationship between the number of target pixels and true areas. The average error rate of the fitted curve was also by only 2.59%, which is within the tolerance of the engineering. Consequently, the proposed method could automatically and accurately calculate the water leakage area from the dataset of images. All the water leakage images used in this study can be downloaded freely from https://doi.org/10.17632/xz2nykszbs.1 .