AI-driven Defect Detection and Multi-Agent Disaster Response for Civil Infrastructure

Abstract

Rapid and reliable assessment of civil infrastructure is essential for effective disaster response. Traditional inspection relies on manual, expert-driven surveys, which are time-consuming, resource-intensive, and difficult to scale under limited accessibility and evolving hazards. Although computer vision enables automated damage detection, most approaches focus on isolated perception tasks and lack support for system-level reasoning.

This thesis develops an AI-driven framework that integrates robust local-scale perception with structured global-scale reasoning. At the local scale, efficient vision models are designed for defect detection under realistic conditions. A Transformer-based crack segmentation model with average pooling improves robustness in noisy environments, while a Robust Feature Knowledge Distillation (RFKD) framework transfers noise-resilient representations from teacher to lightweight student models for deployment. Vision Mamba architectures are further explored to improve scalability for high-resolution inspection.

At the global scale, a multi-agent framework transforms perceptual outputs into actionable insights through role-based task decomposition, information integration, and collaborative reasoning, enabling coherent situational awareness with human-in-the-loop support. Experiments demonstrate improved robustness and scalability over conventional methods.

Challenges remain in handling uncertainty and ensuring reliable long-horizon reasoning. Future work will focus on uncertainty-aware models and tighter integration with sensing and decision-support systems.