Enhancing delamination detection and monitoring of concrete bridges through infrared thermography, deep learning, field data, and numerical simulations

Concrete bridges are essential for a safe, efficient transportation infrastructure, yet they are vulnerable to deterioration over time. One of the most critical forms of damage is delamination, in which layers of concrete separate due to corrosion of the embedded steel reinforcement. Traditional detection methods, such as hammer-sounding or chain-dragging, are time-consuming, subjective, and error-prone. In response, this dissertation proposes a multi-faceted strategy that combines infrared thermography (IRT), deep learning, field data, numerical simulations, and augmented reality (AR) to achieve faster, more reliable, and more objective delamination detection and maintenance planning.The study begins by collecting both field data and numerically simulated images under varied geometries, environmental conditions, delamination properties, locations, image processing techniques, and augmentation approaches. This extensive dataset serves to build a robust object detection–based deep learning model that removes the need for subjective external inputs—a key limitation of many earlier image segmentation techniques. The models evaluated here, Mask R-CNN and YOLOv5, benefit from the inclusion of simulation-generated images, thereby overcoming significant obstacles in the literature related to limited training data, minimal exploration of deep learning methods for IRT-based delamination detection, and incomplete datasets that result in overfitted models resulting in missed or misclassified defects.These numerical simulations not only supply diverse training examples but also allow for an investigation into how different climate zones and seasonal changes across the United States affect optimal IRT inspection times. Many prior studies on ideal detection windows have been limited to isolated locations and times, leading to conflicting or incomplete conclusions due to the difficulty and impracticality of filed data collection. This research details optimal delamination detention conditions and timeframes by simulating and examining detection accuracy throughout the day and across multiple seasons.The final stage introduces an AR-integrated platform that displays, in real-time, the delamination regions revealed during the data analysis process. This tool enhances both collaboration and data management during inspections by integrating current findings with earlier inspection records and maintenance logs, generating a complete history of bridge health.In field trials on a custom-built mockup slab and an in-service bridge, supplementing real IRT data with simulation-based images significantly improved deep learning outcomes, enabling accurate delamination detection across several conditions, most notably deep delamination detection – a significant limitation of current NDE techniques. For diverse U.S. climate zones, midday to early evening emerged as a generally favorable window for inspection. The AR-based system successfully consolidates multiple nondestructive evaluation (NDE) results and historical data into a single, user-friendly interface, suggesting strong potential for streamlining bridge inspection and maintenance practices.