Assessment of beam-column joints in reinforced concrete and precast concrete structures based on CNN

In this study, a CNN (Convolutional Neural Network) based image recognition model is proposed to address the challenges in diagnosis and inspection of deteriorated buildings. With approximately 42.6% of buildings nationwide classified as aging, regular inspections are critical, yet current visual assessments are prone to a shortage of specialized personnel. While existing deep learning studies focus primarily on surface defects, this research targets the failure modes of beam-column joints which are critical elements for overall safety of structural system. Based on data collected from existing literature, a dataset was constructed by classifying the failure modes of beam-column joints in reinforced concrete and precast concrete structures according to the crack patterns at the joints. Using libraries such as TensorFlow and Grad-CAM++, the model was trained, and its performance was evaluated. The classification of joint failure modes based on the ACI 352R-02 code resulted in an accuracy of approximately 64%. In contrast, the 5-fold cross-validation results showed an accuracy of 77% and AUC (Area Under the Curve) of 80%, demonstrating the potential to develop a system that enables even non-experts to easily assess the damaged structures.