Abstract
This study investigates the use of advanced convolutional neural networks (CNNs) to analyze and classify the fracture behavior of U-shaped concrete modified with polyurethane (PU) under repeated drop-weight impact loads. A total of 17 U-shaped specimens were tested under multiple drop-weight impact loads for each PU binder content (0%, 10%, 20%, and 30%) by weight of cement. By integrating digital image correlation (DIC) with dynamic and static mechanical testing, this research evaluates the concrete's impact resistance and flexural behavior with varying PU binder content. Three CNN architectures, InceptionV3, MobileNet, and DenseNet121, were trained on a dataset comprising 1655 high-resolution crack images to classify the failure stages into no crack, initial crack, and advanced failure. Experimental results revealed that 20% PU content optimally enhances impact resistance and flexural strength, while mechanical properties declined significantly with 30% PU content. The strain localization in DIC analysis indicated reduced matrix cohesion, which was measured by the extent of strain concentration in the material, highlighting the importance of maintaining PU content below 20% to avoid compromising structural integrity. Among the models, InceptionV3 demonstrated superior accuracy (96.67%), precision, and recall, outperforming MobileNet (94.56%) and DenseNet121 (90.03%). The combination of DIC and deep learning offers a robust, automated framework for crack assessment, significantly improving accuracy and efficiency over traditional methods such as visual inspections, which are time-consuming and reliant on expert judgment.