<<
[an error occurred while processing this directive] | [an error occurred while processing this directive] >>
Study on a Deep Learning-based Model for Detecting Apparent Defects in Shield Tunnel Lining
(1. China Energy Engineering Group Jiangsu Power Design Institute Co., Ltd., Nanjing 211102; 2. Department of Geotechnical Engineering, College of Civil Engineering, Tongji University, Shanghai 200092; 3. Key Laboratory of Geotechnical and Underground Engineering of the Ministry of Education, Tongji University, Shanghai 200092)
Abstract In this paper, a detection model for many apparent defects of shield tunnel lining based on deep learning modular design is proposed to quickly and accurately identify defects such as water seepage, cracking, falling blocks, and mud & sand leaking on the surface of shield tunnel lining. The model is divided into four modules: data loading, network structure, loss function and post-processing, training and evaluation. Combining the detection principle and data set characteristics of SSD (Single Shot MultiBox Detector) and YOOv4 (You Only Look Once), it is proposed to comprehensively evaluate the matching degree between the prior boxes of the model and this dataset by using two indicators: fitness and maximum possible recall rate. Based on the distribution of defect labeling boxes of the dataset, the K-means method is used to cluster and obtain a set of prior boxes with the highest matching degree.The structure of the SSD model is optimized by considering the structural characteristics of the YOLOv4 model. The results show that the optimized model has a detection accuracy of 0.623, which is nearly 70% higher than that of the original SSD model (0.373). The detection speed has been increased from 40 FPS to 50 FPS, fully proving the rationality of the optimized model.
Abstract：
In this paper, a detection model for many apparent defects of shield tunnel lining based on deep learning modular design is proposed to quickly and accurately identify defects such as water seepage, cracking, falling blocks, and mud & sand leaking on the surface of shield tunnel lining. The model is divided into four modules: data loading, network structure, loss function and post-processing, training and evaluation. Combining the detection principle and data set characteristics of SSD (Single Shot MultiBox Detector) and YOOv4 (You Only Look Once), it is proposed to comprehensively evaluate the matching degree between the prior boxes of the model and this dataset by using two indicators: fitness and maximum possible recall rate. Based on the distribution of defect labeling boxes of the dataset, the K-means method is used to cluster and obtain a set of prior boxes with the highest matching degree.The structure of the SSD model is optimized by considering the structural characteristics of the YOLOv4 model. The results show that the optimized model has a detection accuracy of 0.623, which is nearly 70% higher than that of the original SSD model (0.373). The detection speed has been increased from 40 FPS to 50 FPS, fully proving the rationality of the optimized model.
WU Gang1 LUO Wei2,
3 WANG Xiaolong1 ZHU Jingjing1 JIA Fei2,
3 XUE Yadong2 etc
.Study on a Deep Learning-based Model for Detecting Apparent Defects in Shield Tunnel Lining[J] MODERN TUNNELLING TECHNOLOGY, 2023,V60(4): 67-75
2023, Vol. 60 