Damage detection and localization of structural cracks based on dynamic attention based transformer

Purpose This article addresses the need for a comprehensive model for structural crack detection in the context of structural health monitoring. The main innovation of this research is the introduction of a dynamic attention-based transformer model that significantly enhances the accuracy and efficiency of detecting and localizing cracks in structures. This study seeks to overcome previous limitations and contribute to advancements in structural health monitoring practices. Design/methodology/approach The research focuses on three primary computer vision tasks: classification, object detection and semantic segmentation applied to crack detection in concrete, brick and asphalt structures. The proposed approach employs transformer encoders with dynamic attention mechanisms to assess the severity and extent of damage accurately. Findings In this study, we propose a dynamic attention-based transformer model for structural crack detection, achieving a remarkable accuracy of 99.38% and an impressive F1 score. Our method demonstrates superior performance compared to existing techniques, such as the fusion features-based broad learning system and deep convolutional neural networks, while also significantly reducing execution time, highlighting its efficiency and potential for practical applications in structural health monitoring. Originality/value This research introduces a novel framework for crack detection, leveraging recent advancements in deep learning technology, with significant implications for the field of civil engineering and maintenance.