Multicategory fire damage detection of post‐fire reinforced concrete structural components

Abstract

This paper introduces an enhanced you only look once (YOLO) v5s‐D network customized for detecting various categories of damage to post‐fire reinforced concrete (RC) components. These damage types encompass surface soot, cracks, concrete spalling, and rebar exposure. A dataset containing 1536 images depicting damaged RC components was compiled. By integrating ShuffleNet, adaptive attention mechanisms, and a feature enhancement module, the capability of the network for multi‐scale feature extraction in complex backgrounds was improved, alongside a reduction in model parameters. Consequently, YOLOv5s‐D achieved a detection accuracy of 93%, marking an 11% enhancement over the baseline YOLOv5s network. Comparison and ablation tests conducted on different modules, varying dataset sizes, against other state‐of‐the‐art networks, and on public datasets validate the resilience, superiority, and generalization capability of YOLOv5s‐D. Finally, an application leveraging YOLOv5s‐D was developed and integrated into a mobile device to facilitate real‐time detection of post‐fire damaged RC components. This application can integrate diverse fire scenarios and data types, expanding its scope in future. The proposed detection method compensates for the subjective limitations of manual inspections, providing a reference for damage assessment.