Automated delamination detection in concrete bridge decks using 1D-CNN and GPR data
The adoption of deep learning models for ground penetrating radar (GPR) data analysis has great potential for advancing the field of infrastructure condition monitoring, especially in the early detection of bridge deck distresses. This work presents a deep learning approach to detect delamination in...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-07-01
|
| Series: | Case Studies in Construction Materials |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2214509524013263 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The adoption of deep learning models for ground penetrating radar (GPR) data analysis has great potential for advancing the field of infrastructure condition monitoring, especially in the early detection of bridge deck distresses. This work presents a deep learning approach to detect delamination in concrete bridge decks using GPR data, employing one-dimensional convolutional neural networks (1D-CNN). The experiment uses GPR data from the SDNET2021 dataset containing five in-service bridge decks. The main objective is to classify each GPR A-scan as either 'sound' or 'delaminated', thus allowing efficient and timely detection of subsurface structural problems. The proposed method incorporates a contextual information approach to enhance the accuracy and reliability of delamination detection. Two techniques were evaluated and compared to identify the optimal approach. The results demonstrate that the A-scan data, when combined with the average filter, significantly improves the detection performance with a 0.9940 weighted average F1 score compared to the raw A-scan only with a 0.7735 weighted average F1-score. Moreover, a real case study is introduced with a transfer learning approach. The detection results achieved 92.6 % accuracy when a pre-trained model was fine-tuned with 5 % of the labels from the new data. The findings of the research contribute to the advancement of non-destructive testing methodologies by providing the first approach to benchmark and work with the GPR data of the SDNET2021 dataset. |
|---|---|
| ISSN: | 2214-5095 |