Cavity detection under road based on deep learning models with ground penetration radar data|RISS 상세보기

국문 초록 (Abstract)

최근 다양한 원인에 의해 발생하고 있는 공동현상은 도로 및 인도의 지반침하로 이어져 국민의 안전과 생명, 물적 피해를 동시에 발생시키며 사회적 재난의 한 형태로 인식되고 있다. 이러한 사회적 재난으로부터 사전에 예방하기위한 조치로 국토교통부는 지하안전관리에 관한 특별법이 지난 2016년 개정하여, 2018년 1월 1일부터 시행되었다. 안전관리항목으로 계측관리방안과 지하탐사항목으로 도로 하부공동 및 지하매설물 위치파악에 지표투과레이더(GPR; Ground Penetration Radar)탐사 방법 등이 명시되었다(국토교통부, 2022). 하지만 GPR 탐지 후 수동처리(인력에 의한 분석)는 노동 집약적이며, 오류가 발생하기 쉽다(Hou, Feifel, et al., 2021). 최근까지도 전문분석가에 의한 경험 및 데이터 분석을 통해 공동의 의심·추정·판단 및 그 결과를 도출하는데 많은 시간이 소요되고, 탐지의 오류가 발생할 수 있다.
본 연구는 GPR 데이터를 통해 공동을 탐지하는 딥러닝 모델을 제안하는 것으로, 정확한 탐지성능과 인력분석에 소요되는 시간 및 공통탐지 오류를 최소화하고, 공동탐지에 따라 서울시 공동등급 분류기준(2019)에 따른 공동등급 및 관리방안의 매뉴얼에 따라 신속한 보수 보강의 의사결정을 할 수 있는 GPR 데이터를 활용한 딥러닝(Deep learning) 기반의 도로 하부 공동탐지 모델에 관한 연구의 필요성이 국내에서도 대두되고 있다. 연구 방법으로 최근 지하에 설치된 지하매설물 탐사분석에 있어 해외에서는 CNN 기반의 다양한 모델을 적용한 연구가 활발히 진행되고 있으며, 교량 상판의 철근 배근 확인 등 딥러닝 모델을 통해 연구가 계속되고 있으며, 국내에서도 Faster R-CNN을 통해 도로포장 균열 자동탐지 소프트웨어 개발과 같은 연구를 진행하고 있다.
본 연구에서는 딥러닝 모델을 통해 공동을 탐지하기 위해서는 많은 양의 공동데이터를 학습해야 탐지성능을 높일 수 있기에, 확보된 소량의 GPR 공동데이터를 증강하기 위해 AutoEncord와 Dual-GAN을 통해 증강한 결과 Dual-GAN의 품질이 우수한 증강을 나타낸 Dual-GAN을 적용하였다. 딥러닝 모델을 제안하기 위해 첫 번째 모델로 분석속도가 우세한 One stage Architecture의 YOLECT(You Only Look At CoefficienTs)과 두 변째로 정확도에서 우세한 Two stage Architecture의 R-CNN 계열인 Faster R-CNN과 세 번째로 Faster R-CNN 모델의 확장된 모델인 Mask R-CNN을 적용하여, 공동을 탐지하는 실험을 통해 Mask R-CNN이 가장 우수한 탐지성능을 보인 것을 확인할 수 있었다. 신속하게 판독 및 공동의 유무를 판정하는 딥러닝(Deep Learning) 기반의 탐지모델을 제안하였다. GPR 데이터 구축은 향후 인공지능 학습용 데이터 구축 사업에 필요한 후보 과제이며, 4차 산업혁명의 핵심기술인 딥러닝(Deep Learning) 모델을 적용하여 지하의 공동탐지를 하였다는 것은 기술의 융합이란 점에서 학술적으로 의미가 있으며, 향후 산업계에서도 탐지를 위한 활용방안으로 많은 도움이 될 것이라고 기대한다.

번역하기

최근 다양한 원인에 의해 발생하고 있는 공동현상은 도로 및 인도의 지반침하로 이어져 국민의 안전과 생명, 물적 피해를 동시에 발생시키며 사회적 재난의 한 형태로 인식되고 있다. 이러...

다국어 초록 (Multilingual Abstract)

With recent active pace of urban development, cavity phenomenon arising from various causes leads to ground subsidence of roads and sidewalks, posing risks to people's safety, life as well as material damage at the same time, and is recognized as a type of social disaster. In an effort to establish preventive measures against such social disasters, the Ministry of Land, Infrastructure and Transport(MOLIT) amended the Special Act on Underground Safety Management in 2016 and the amendment took effect as of January 1, 2018. In the clause on safety management, measures of monitoring and management were stipulated, and in the clause on underground surveying, methods of Ground-Penetrating Radar(GPR) surveying were specified for locating cavities underneath urban roads and underground facilities(MOLIT, 2022). However, manual data processing(analysis by manpower) of GPR scan data after GPR detection is labor-intensive and prone to human errors(Hou, Feifel, et al., 2021). Until recently, it has been still time-consuming to derive results from data suspected of cavities based on experience of and data analysis by human expert analysists, there may still be detection errors.
This study proposes a deep learning model for cavity detection based on GPR data, aiming to achieve accurate detection performance and minimize time required for analysis by human inputs and errors in cavity detection. In South Korea, there has been increasing emphasis on necessity of research on deep learning-based models for cavity detection underneath urban roads using GPR data that enables prompt decision-making on repair and rehabilitation for detected cavities based on cavity grades according to Cavity Grade Classification Standards, City of Seoul(City of Seoul, 2019) and manual of management methods. In terms of research methodology, in the field of inspection and analysis of underground facilities, studies using various CNN-based models have been actively conducted in overseas cases as well as continuous research with application of deep learning model such as detecting rebars in bridge decks. In the case of South Korea, relevant studies include research on software development for automated crack detection in road pavements using faster R-CNN.
In this study, in order to perform cavity detection using a deep learning model, the model needs to be trained with a large amount of cavity data to improve detection performance. Therefore, to perform data augmentation with a small amount of cavity data acquired with GPR, autoencoder and dual-GAN were used, and dual-GAN that showed high quality data augmentation performance was applied. In order to propose a deep learning model, cavity detection experiments were performed with three different models for comparison of detection performance: The first model was YOLACT(You Only Look At CoefficienTs) with one-stage architecture with superior analysis efficiency; the second model was faster R-CNN, a type of R-CNN model with two-stage architecture with superior accuracy; and the third model was Mask R-CNN, an extended version of faster R-CNN. The results of cavity detection experiments showed that Mask R-CNN had the best detection performance out of the three models. In this way, we proposed a deep-learning based cavity detection model, capable of fast reading and classification of cavity status. Construction of GPR dataset is a candidate task required for future projects on construction of data for AI training, and the representative achievement of this study, which is performing underground cavity detection with application of deep learning models, a core technology in the era of the 4th industrial revolution, has academic significance in terms of technological convergence. Furthermore, it is expected that the findings of this study will be of great use for cavity detection for industrial applications in the future.

번역하기

목차 (Table of Contents)

Chapter 1. Introduction 1
1.1 Research Background and Necessity 1
1.1.1 Research Background 1
1.1.2 Research Necessity 13
1.2 Research Objectives, Methodology, and Composition 15

Chapter 1. Introduction 1
1.1 Research Background and Necessity 1
1.1.1 Research Background 1
1.1.2 Research Necessity 13
1.2 Research Objectives, Methodology, and Composition 15
1.2.1 Research Objectives 15
1.2.2 Research Methodology 17
1.2.3 Research Composition 18
Chapter 2. Background and Preceding Review 20
2.1 Ground Penetration Radar(GPR) exploration 20
2.1.1 Theoretical Consideration on GPR exploration 20
2.1.2 Method of exploration with GPR 25
2.1.3 GPR equipment 30
2.2 Cavity 33
2.2.1 Theoretical Consideration of underground cavity 33
2.2.2 Characteristics of GPR cavity signals 39
2.2.3 Cavity grade classification by Seoul Metropolitan Government 42
2.3 Ground subsidence 45
2.3.1 Theoretical Consideration of ground subsidence 45
2.3.2 Domestic and overseas cases of ground subsidence incidents 56
2.4 Artificial Intelligence Image Recognition Technology 68
2.4.1 Theoretical Consideration of image recognition technology with AI 68
2.4.2 One Stage Architecture 75
2.4.3 Two Stage Architecture 77
2.4.4 AutoEncoder 83
2.4.5 Generative Adversarial Network(GAN) 86
Chapter 3. Research Methodology 92
3.1 Overview of Research Methodology 92
3.2 Data Augmentation Model 94
3.3 Data Annotation Processing 96
3.3 Deep Learning based Cavity Detection Model 97
Chapter 4. Case Analysis and Result 102
4.1 Data Collection 102
4.2 Data Augmentation 110
4.3 Data Annotation Processing 115
4.4 Cavity Detection Model Performance Analysis 116
Chapter 5. Conclusions 125
5.1 Summary of Research Results 125
5.2 Academic Implications 127
5.3 Plans for Industrial Utilization 130
5.4 Limitations of Research and Future Research Directions 132
References 135
Korean Abstract 153

참고문헌 (Reference)

1. Squeeze-and-Excitation Networks, Hu, J., Sun, G., Shen, L., 7132-7141, , 2018

2. Image segmentation by clustering, Andrews, H. C., Coleman, G. B., 67(5), 773-785, , 1979

3. Yolact Real-time instance segmentation, Lee, Y. J., Bolya, D., Xiao, F., Zhou, C., 9157-9166, , 2019

4. Groundwater flow Analysis Using MODFLOW in the Tunnel, Kim, J. H., Ji, H. K., Heo, C. H., Lee, S. T., 36(1), 129-142, , 2004

5. Safety Management for Ground Excavation and Groundwater, Kim, J. O., Kim, J. H., Park, M. C., Lee, J. H., 69(1), 18-25, , 2021

6. Deriving Strategic Agenda for Response of Road Sink Phenomenon, Choi, B. L., Sung, J. H., Park, W. J.,, Lee, J. K., 31(6), 99-104, , 2016

7. Automatic analysis of GPR images: A pattern-recognition approach, Farid, M., Pasolli, E., Massimo, D., 47(7), 2206-2217, , 2009

8. DEM Simulation on the Initiation and Development of Road Subsidence, Kim, Y. H., Park, S. Y., 33(7), 43-53, , 2017

9. Sinkholes, West-Central Florida, Land subsidence in the United States, Thiansky, A. B, 1182, 121-140, , 1999

10. A Study on GPR Image Classification by Semi-supervised Learning with CNN, Bae, H. L., Kim, H. M., 6(1), 197-206, , 2021