Date of Award
Master of Science (MS)
We collected and developed an SDNET2021 dataset, a rare combination of annotated non-destructive evaluation (NDE) data, from five in-service bridge decks in Grand Forks as a feed for validating, benchmarking, developing, training, and testing artificial intelligence (AI) models to evaluate, monitor, and assess bridge conditions. The developed dataset, which serves as ground truth, contains sound concrete as class 1 and delaminated sub-surface concrete as classes 2 and 3. The SDNET2021 dataset consists of 488 delaminated (class 2 and 3) and 1,448 sound (class 1) Impact Echo (IE) signals, 214,943 delaminated (class 2 and 3) and 448,159 sound (class 1) Ground Penetrating Radar (GPR) signals, and 1,718,083 delaminated (class 2 and 3) and 2,862,597 sound (class 1) pixels from Infrared thermography (IRT). SDNET2021 is publicly freely available. The dataset was annotated autonomously to reduce human errors and increase reliability. This study presents an adaptative image processing-based model for bridge deck sub-surface delamination evaluations. The proposed method adopts the IRT dataset and annotated ground truth generated from the in-service bridge decks. The model was developed by iterating the sensitivity (s) parameter and optimized by selecting s-values based on performance evaluation metric interactions. The evaluated 2- and 3-clustered optimized s-values ranged from 0.365 to 0.38 and 0.459 to 0.486, respectively. An average accuracy of 69% was obtained for the model. The study revealed that several factors, such as delamination depth and spatial dimensions, ambient weather conditions such as wind speed, temperature, and humidity, and mosaic image quality affect the IRT model’s performance.
Ichi, Eberechi Orie, "Validating NDE Dataset And Benchmarking Infrared Thermography For Delamination Detection In Bridge Decks" (2021). Theses and Dissertations. 4170.