Bridges constitute vital elements of Taiwan’s transportation infrastructure; nevertheless, numerous structures have surpassed their intended service life, thereby posing safety hazards and amplifying the necessity for prompt maintenance. Conventional inspection methodologies, which predominantly depend on visual assessments, are labor-intensive, time-consuming, costly, and prone to human error. To address these limitations, this study introduces a multi-stage automated inspection and reporting framework that integrates unmanned aerial vehicle (UAV) sensing, computer vision, large language models (LLMs), and metaheuristic optimization. UAVs initially capture high-resolution imagery of bridge components to ensure exhaustive data collection while reducing inspector exposure to hazardous conditions. A Vision Transformer (ViT) is utilized for image classification. In contrast, the YOLO (You Only Look Once) model executes object detection and instance segmentation to localize deterioration, evaluate its severity, and estimate repair costs. The framework is further refined by incorporating a multimodal LLM, fine-tuned on a dataset of common deterioration types and structural components derived from bridge inspection records in Taiwan, to automatically generate descriptive reports that document deterioration conditions, recommend repair strategies, and maintain clarity and technical precision. Additionally, the hyperparameters of the integrated system are optimized through a newly developed metaheuristic algorithm, thereby ensuring enhanced efficiency and accuracy. Validation via cross-validation and language model performance metrics affirms the system’s robustness and capability for generalization. By integrating UAV sensing, advanced computer vision, LLM-driven report generation, and optimization algorithms, the proposed framework markedly enhances the efficiency, accuracy, and safety of bridge inspections while equipping management agencies with cost-effective and sustainable decision-support tools for long-term infrastructure maintenance.

Keywords

Automated Bridge Inspection; Unmanned Aerial Vehicles; Computer Vision; Large Language Models; Metaheuristic Optimization; Structural Health Monitoring.