Knowledge-driven 3D damage mapping and decision support for fire-damaged reinforced concrete structures using enhanced deep learning and multi-modal sensing

Abstract

Rapid and precise damage assessment of fire-damaged reinforced concrete (RC) structures is critical for structural safety decisions. To overcome limitations of existing 2D methods in spatial localization and real-time deployment, an integrated knowledge-driven framework is proposed. Multi-modal sensing is combined with an enhanced FastSAM-P deep learning network for automated 3D damage mapping. Three core innovations are introduced: (1) Deformable Spatial-Channel Reconstruction Convolution (DSCConv) dynamically adjusts receptive fields to capture fine-grained damage features; (2) Receptive Field Block (RFB) module optimizes multi-scale feature extraction; (3) Pyramid Pooling Shuffle Attention (PPSM) enhances robustness in noisy environments through contextual fusion. The framework achieves 92.0 % mean Intersection-over-Union (mIoU) for segmenting concrete spalling and rebar exposure, with inference at 65.11 FPS on GPU. Validation across five public datasets (roads, bridges, buildings) confirms generalization capability. Deployment on Jetson TX1 edge devices demonstrates operational feasibility (123.4 ms latency). Integration with photogrammetric 3D reconstruction enables damage localization within ± 15 mm accuracy. This approach establishes a scientifically rigorous pipeline from data acquisition to decision support, significantly advancing automated post-fire assessment for knowledge-intensive engineering tasks.