Non-Destructive Testing and Machine Learning in Inspection of Post-Tensioned Concrete Bridges

Abstract

Bridges are critical infrastructures essential for transportation and economic activities yet face accelerated deterioration from environmental factors and increased usage. Post- tensioned concrete bridges offer potential solutions but recent failures underscore deficiencies in inspection protocols. In Norway, where harsh climates and marine environments exacerbate deterioration, effective Non-Destructive Testing (NDT) methods are crucial. This thesis explores the integration of Machine Learning (ML) with NDT techniques for post-tensioned bridge inspection. Mock-up testing revealed significant variability among NDT devices, emphasizing the need for expert interpretation. Training ML models, including the Roboflow model, demonstrated initial efficacy but highlighted the ongoing need for fine- tuning and optimal dataset distribution. Model testing across various scenarios showed the Roboflow model's superiority in detecting structural features such as ducts and rebars, albeit with challenges in distinguishing subtle features. Custom models also showed potential, emphasizing the need for further refinement and adaptation to real-world conditions. In conclusion, effective real-time detection of structural features in post-tensioned bridges requires advanced ML models integrated with expert NDT interpretation. While promising, ongoing optimization and validation against real-world datasets are essential. This research contributes to enhancing inspection protocols, aiming to improve infrastructure safety and longevity in civil engineering practices. Keywords: Non-destructive testing, Machine Learning, Post-tensioned Concrete, Roboflow, Custom Models