Fault Diagnosis of Backfill and Contact Status Using CNN and CWT
- 주제(키워드) Backfill , Continuous Wavelet Transform , Convolutional Neural Networks , Non-destructive Testing
- 주제(DDC) 510
- 발행기관 아주대학교
- 지도교수 최수영
- 발행년도 2022
- 학위수여년월 2022. 8
- 학위명 석사
- 학과 및 전공 일반대학원 수학과
- 실제URI http://www.dcollection.net/handler/ajou/000000031963
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
The structural collapse has led to many casualties in recent years. For this reason, significant research interest has been directed to evaluating the safety conditions of structures, an extension of the structural lifecycle, and the prevention of human casualties. This study examines socially required skills such as saving, safety, efficiency, and accuracy of structures. Typically, non-destructive testing (NDT) is used to inspect defects without modifying the state of the structure. Moreover, we examine the technology that incorporates the impact-echo method, which is by incorporating the NDT technique into machine learning. Among many structures, the defects are analyzed for the inner wall of the tunnel. With the help of Prof. Suyoung Choi of Ajou University and HBC Inc., we collected signal data and converted it into electrical signals, using an oscilloscope to measure physical vibrations caused by a mechanical impact of the structure. To prevent the regularity of the data according to the hitting position, the hitting position was randomly performed, and the data ratios of the hitting positions A, B, and C were 2:3:5. First of all, the study showed a performance of approximately 70% when Convolutional Neural Network (CNN) was analyzed with this one-dimensional time series data. We use two-dimensional image data obtained through continuous wavelet transformation (CWT) to improve the model performance. The image data showed a difference, according to the presence or absence of backfill, unobserved in the one-dimensional data. We distinguished between area C without a backfill and area A with a backfill, and the difference between areas A and B with a backfill was minimal. However, the CNN model showed some predictability. The final performance of our built model achieved 95.38%. The F1-score, which determined the data imbalance problem with the hitting position, achieved 94.21%.
more목차
1 Introduction 1
2 Preliminaries 2
2.1 Impact-echo Method 2
2.2 Fourier & Wavelet Transforms 2
2.2.1 Fourier Series 2
2.2.2 Fourier Transform 4
2.2.3 Short Time Fourier Transform 4
2.2.4 Continuous Wavelet Transform 5
2.3 Convolutional Neural Networks 6
2.4 Cross Validation 9
2.4.1 K-fold Cross Validation 9
2.4.2 Stratified K-fold Cross Validation 9
2.5 Confusion Matrix 10
3 Experimental Results 12
3.1 Data Description 12
3.2 Results 15
3.2.1 Different wavelets 15
3.2.2 Neural network variations 15
3.2.3 Different K values 16
3.2.4 Model Summary 17
3.3 Conclusion 18
References 19
