AI-based damage detection in prestressed concrete beams: a vision-integrated deep learning framework for crack localization and severity classification

Abstract Aging prestressed concrete (PC) structures, particularly those utilizing unbonded tendons, are susceptible to long-term deterioration mechanisms such as creep, shrinkage, and prestress loss, which manifest in complex damage forms including surface cracking and FRP debonding. Traditional inspection techniques are often labor-intensive and insufficiently accurate in detecting early-stage or subvisible defects. This study proposes a novel artificial intelligence (AI)-driven framework for automated defect detection and classification in prestressed concrete beams based on visual and structural response data. This work examines progressive damage in newly cast laboratory prestressed concrete beams under controlled loading, using synchronized high-resolution vision and vibration measurements, and evaluates specimen-disjoint generalization and edge-oriented deployability. A hybrid deep learning architecture combining YOLOv8 for crack localization, Swin Transformer for damage severity classification, and CNN–Transformer models for time-series vibration analysis is developed. The results demonstrate that the proposed system achieves a crack detection accuracy of 96.3%, a severity classification F1-score of 93.1%, and can localize damage with a mean IoU of 0.82. This work presents an integrated, aging-aware multimodal AI framework for damage assessment in prestressed concrete beams with unbonded tendons, combining real-time vision-based crack/debonding localization with synchronized vibration/strain analysis to provide a scalable alternative to traditional SHM methods. Graphical Abstract