표제지
목차
요지 15
ABSTRACT 17
제1장 서론 20
1.1. 연구 배경 20
1.1.1. 국내 교량의 현황 22
1.1.2. 국내 교량의 안전진단 현황 25
1.1.3. 스마트한 교량 진단 기술 현황 29
1.2. 연구 목적 33
1.3. 연구 내용 34
제2장 개선된 교량 해석모델 기반 손상 탐지 36
2.1. 교량의 손상 탐지 개요 36
2.2. 계측 응답 기반 개선된 유한요소모델 구현 39
2.2.1. 계측응답 기반 유한요소모델 개선에 대한 연구 현황 39
2.2.2. RCGA 최적화 기법을 통한 유한요소모델 개선 41
2.3. 교량의 동적 응답에 대한 Signal analysis 50
2.3.1. 동적 응답에 대한 Signal analysis 50
2.3.2. 단시간 푸리에 변환 기법 Spectrogram 53
2.4. Deep learning을 통한 교량의 손상 탐지 61
2.4.1. Deep learning을 통한 손상 탐지 학습 연구 동향 61
2.4.2. CNN를 통한 이미지 학습 64
2.4.3. CNN 기반 교량 Spectrogram 이미지를 위한 학습 모델 설계 68
제3장 실측 기반 개선된 교량 유한요소모델 구현 69
3.1. 선정된 교량 소개 69
3.1.1. 교량 제원 69
3.1.2. 의사정적재하시험에 따른 계측 응답 70
3.1.3. 교량의 정적 변위 추출 73
3.2. 보완된 RCGA 모델업데이팅 기법을 통한 개선된 유한요소모델 75
3.2.1. 교량에 대한 초기 유한요소모델 75
3.2.2. 보완된 RCGA 모델업데이팅 기법 적용 77
3.3. 개선된 유한요소모델 검증 84
3.3.1. 교량의 계측 변위와의 비교 84
3.3.2. 교량의 고유진동수와의 비교 90
3.4. 소결 91
제4장 교량 손상 탐지를 위한 학습데이터 구축 92
4.1. 교량 손상 탐지 기법에 대해 가정한 사항 92
4.1.1. 한 종류의 차량 사용 92
4.1.2. 차량 1대만 등속도로 주행 94
4.1.3. 손상 시뮬레이션에 적용한 차량 94
4.2. 개선된 유한요소모델을 통한 손상 시뮬레이션 95
4.2.1. 선정된 차량에 대한 동적 재하 방식 95
4.2.2. 손상 시나리오 구성 98
4.3. 교량 손상에 대한 Spectrogram 학습 데이터 구축 101
4.3.1. 해석 응답을 통한 Spectrogram 이미지 101
4.3.2. 손상 학습을 위한 Spectrogram 이미지 데이터 구축 113
4.4. 소결 115
제5장 CNN 기반 손상 탐지 학습 모델 설계 116
5.1. Spectrogram의 학습용 이미지 데이터 분석 116
5.2. 손상 탐지를 위한 CNN 기반 Global 학습 방법 118
5.2.1. CNN 기반 Global 학습 모델 설계 118
5.2.2. Global 학습 모델 결과 121
5.3. 손상 탐지를 위한 CNN 기반 Local 학습 방법 123
5.3.1. CNN 기반 Local 학습 모델 설계 123
5.3.2. Local 학습 모델 결과 126
5.4. 소결 137
제6장 교량 손상 탐지 학습 모델 현장 적용 138
6.1. 실제 교량의 손상 138
6.2. 손상으로 인한 교량의 계측 응답 변화 139
6.2.1. 교량의 손상에 따른 정적 계측 응답 139
6.2.2. 교량의 손상에 따른 의사정적 계측 응답 143
6.3. 손상 탐지 학습 모델에 계측 응답 적용 146
6.3.1. 계측 응답에 대한 Spectrogram 이미지 데이터 146
6.3.2. CNN 기반 Local 학습 모델에 계측 응답 적용 148
6.3.3. 계측 응답 적용 전 해석 응답을 통한 손상 탐지 150
6.3.4. 계측 응답 적용 후 손상 탐지 결과 156
6.4. 소결 168
제7장 결론 169
7.1. 연구 결과 169
7.1.1. 개선된 유한요소모델 169
7.1.2. Spectrogram을 사용한 손상 특성 분석 170
7.1.3. Local 학습 방법을 통한 손상 탐지 170
7.1.4. Local 학습 모델 현장 적용 결과 170
7.2. 연구 성과의 의의 172
7.3. 향후 연구 방향 173
7.3.1. 연구 범위 외 교량 손상 탐지 방법 173
7.3.2. 계측 센서에 대한 추가 연구 173
7.3.3. 교량의 손상 탐지 기법을 통한 보수 및 보강 시기 결정 174
참고문헌 175
Table 1.1.1. Facility safety level 25
Table 1.1.2. Safety inspection cycle of facilities 27
Table 3.1.1. Axle weights of the truck 71
Table 3.1.2. Specifications of the displacement sensor used in the pseudo-static load test 72
Table 3.1.3. Static vertical displacement through pseudo-static load test 74
Table 3.2.1. Elastic modulus in element stiffness of the initial finite element model 76
Table 3.2.2. Boundary conditions of the initial finite element model 76
Table 3.2.3. Range of elastic modulus CEi for the 1st RCGA case[이미지참조] 78
Table 3.2.4. Range of boundary condition Nj for the 1st RCGA case[이미지참조] 78
Table 3.2.5. Result of CEi for the 1st RCGA case[이미지참조] 79
Table 3.2.6. Result of Nj for the 1st RCGA case[이미지참조] 79
Table 3.2.7. Range of elastic modulus CEi for the 2nd RCGA case[이미지참조] 81
Table 3.2.8. Range of boundary condition Nj for the 2nd RCGA case[이미지참조] 81
Table 3.2.9. Result of CEi for the 2nd RCGA case[이미지참조] 82
Table 3.2.10. Result of Nj for the 2nd RCGA case[이미지참조] 82
Table 3.3.1. Result of vertical displacement for each finite element model 86
Table 3.3.2. Result of SE and E for each finite element model 89
Table 3.3.3. Result of Ef for each finite element model[이미지참조] 90
Table 4.3.1. Change of the spectrogram for time resolution range from 0.4 to 0.6(from Fig 4.3.2) 105
Table 4.3.2. Change of the spectrogram for time resolution range from 0.1 to 0.3(from Fig 4.3.2) 106
Table 4.3.3. Change of the spectrogram for leakage range from 0.6 to 1(from time resolution 0.2 in Table 4.3.2) 107
Table 4.3.4. Result of spectrogram applying reassign from Table 4.3.3 108
Table 4.3.5. Final parameters of spectrogram for learning damage detection 109
Table 4.3.6. Result of vertical acceleration spectrogram at s2 position for LC1 110
Table 4.3.7. Result of vertical displacement spectrogram at s2 position for LC1 111
Table 4.3.8. Result of longitudinal strain spectrogram at s2 position for LC1 112
Table 5.2.1. Accuracy results of CNN-based learning model for 3 type responses 121
Table 5.3.1. Accuracy of each learning model using vertical analysis acceleration for G1 damage case 128
Table 5.3.2. Accuracy of each learning model using vertical analysis acceleration for G2 damage case 129
Table 5.3.3. Accuracy of each learning model using vertical analysis acceleration for G3 damage case 130
Table 5.3.4. Accuracy of each learning model using vertical analysis acceleration for G4 damage case 131
Table 5.3.5. Accuracy of each learning model using vertical analysis acceleration for G5 damage case 132
Table 6.2.1. Static vertical displacement response measured at the target bridge 139
Table 6.2.2. Static longitudinal strain response measured at the target bridge 141
Table 6.2.3. Specifications of each sensor used in the pseudo-static load test 144
Table 6.2.4. Measured maximum vertical displacement range of each load case 145
Table 6.2.5. Measured maximum longitudinal strain range of each load case 145
Table 6.3.1. Spectrogram image data of the responses measured at the mid-span of G1 when LC1 147
Table 6.3.2. Damage prediction results for G5 using undamaged analysis vertical acceleration 151
Table 6.3.3. Probability of damage prediction for undamaged case with vertical acceleration 152
Table 6.3.4. Probability of damage prediction for G5 10% damage case with vertical acceleration 153
Table 6.3.5. Probability of damage prediction for G5 20% damage case with vertical acceleration 154
Table 6.3.6. Damage prediction results for G5 using undamaged measured vertical acceleration 158
Table 6.3.7. Probability of predicting damage for undamaged measured vertical acceleration 159
Table 6.3.8. Probability of predicting damage for G5 tendon 2/4 damage case measured vertical acceleration 160
Table 6.3.9. Probability of predicting damage for G5 tendon 4/4 damage case measured vertical acceleration 161
Table 6.3.10. Probability of predicting damage for undamaged measured vertical displacement 162
Table 6.3.11. Probability of predicting damage for G5 tendon 2/4 damage case measured vertical displacement 163
Table 6.3.12. Probability of predicting damage for G5 tendon 4/4 damage case measured vertical displacement 163
Table 6.3.13. Probability of predicting damage for undamaged measured longitudinal strain 164
Table 6.3.14. Probability of predicting damage for G5 tendon 2/4 damage case measured longitudinal strain 165
Table 6.3.15. Probability of predicting damage for G5 tendon 4/4 damage case measured longitudinal strain 166
Table 6.3.16. Maximum vertical analysis displacement of each load case 167
Table 6.3.17. Maximum longitudinal analysis strain of each load case 167
Fig 1.1.1. Status of all facilities and road bridges in Korea 22
Fig 1.1.2. Number of road bridges managed by each institution 23
Fig 1.1.3. Annual status of facilities managed by each institution 23
Fig 1.1.4. Number of bridges according to period of use 24
Fig 1.1.5. Safety level status of bridges according to period of use 26
Fig 1.1.6. Continuous displacement estimation technology through contact and non-contact sensors 30
Fig 1.1.7. Steel Strand for Prestressed Concrete with Built-in FBG Sensor 31
Fig 2.1.1. The process of damage detection in a bridge 38
Fig 2.2.1. Bayesian-based model updating for pedestrian truss bridge 40
Fig 2.2.2. Bayesian-based nonlinear model updating method 40
Fig 2.2.3. Process of RCGA 42
Fig 2.2.4. The process of RCGA model updating with FEM analysis 45
Fig 2.3.1. Identification of natural frequency of a railway bridge through wavelet transform 52
Fig 2.3.2. Damage detection of structures through wavelet transform images 52
Fig 2.3.3. Results of spectrum and spectrogram using MATLAB 54
Fig 2.3.4. Results of spectrogram when time resolution changes 56
Fig 2.3.5. Results of spectrogram when leakage changes 58
Fig 2.3.6. Results of spectrogram according to reassign 60
Fig 2.4.1. Gas identification using spectrogram and CNN 62
Fig 2.4.2. Diagnosis pancreatic cancer using 1D-CNN and LSTM 62
Fig 2.4.3. Spectrogram generated through the acceleration reponse 63
Fig 2.4.4. Principle of convolution layer in CNN 65
Fig 2.4.5. Process of applying convolution layer using padding 65
Fig 2.4.6. Principle of max pooling layer 67
Fig 2.4.7. Principle of flatten layer 67
Fig 2.4.8. Deep learning model based on AlexNet CNN 68
Fig 3.1.1. The appearance of the target bridge 69
Fig 3.1.2. Tendons reinforced at the bottom of the outer girder of the target bridge 70
Fig 3.1.3. Axle spacing of the truck for pseudo-static load test 71
Fig 3.1.4. The truck driving path in pseudo-static load test 71
Fig 3.1.5. Displacement sensor position of A-A cross section in the target bridge 72
Fig 3.1.6. Results of applying filter to dynamic displacement response 73
Fig 3.1.7. Displacements after removing dynamic components through smoothing 74
Fig 3.2.1. Initial finite element model of the target bridge 75
Fig 3.3.1. Load cases of updated finite element model for displacement analysis 85
Fig 3.3.2. Vertical displacement of mid-span at the bridge for each load case 87
Fig 4.1.1. Comparison of vertical displacement according to damage 93
Fig 4.2.1. The truck load path for generating learning data in updated finite element model 96
Fig 4.2.2. The extraction location of analysis response in updated finite element model(bottom view) 97
Fig 4.2.3. Cases for damage location in updated finite element model(bottom view) 99
Fig 4.2.4. Total number of training data generated from damage scenarios 100
Fig 4.3.1. Comparison of vertical acceleration spectrogram at s2 position in G1 102
Fig 4.3.2. Results of the spectrogram for frequency range from 0 to 30Hz(from Fig 4.3.1) 103
Fig 4.3.3. Common parts of spectrogram graphs 113
Fig 4.3.4. Training image data for CNN extracted from spectrogram graph 114
Fig 5.1.1. Matrix size and RGB components of a training image data 117
Fig 5.1.2. Each RGB matrix pixel value of a training image data 117
Fig 5.2.1. Layer configuration diagram of the global CNN-based learning model 120
Fig 5.2.2. Loss results of CNN-based learning model for each response 122
Fig 5.3.1. Structure of CNN-based local learning model method for damage detection 124
Fig 5.3.2. Composition of data used in one CNN-based local learning model 125
Fig 5.3.3. Loss according to epoch in one CNN-based local learning model 126
Fig 5.3.4. Results of the CNN-based local learning model for vertical acceleration response 134
Fig 5.3.5. Results of the CNN-based local learning model for vertical displacement response 135
Fig 5.3.6. Results of the CNN-based local learning model for longitudinal strain response 136
Fig 6.1.1. External tendon failure to the target bridge 138
Fig 6.2.1. Static vertical displacement in the mid-span of the target bridge for damage cases 140
Fig 6.2.2. Static longitudinal strain in the mid-span of the target bridge for damage cases 142
Fig 6.2.3. Location of response measurement during pseudo-static load test 143
Fig 6.3.1. The process of creating spectrogram image data from measured responses 146
Fig 6.3.2. Total number of measurement responses 148
Fig 6.3.3. The process of inputting image data from measurement responses into the CNN-based Local learning model 149