Advancing Structural Health Monitoring with Lightweight Real-Time Deep Learning-Based Corrosion Detection

Abstract

Structural Health Monitoring (SHM) is essential for preserving the safety and service life of industrial infrastructure. Corrosion, in particular, remains one of the most critical degradation phenomena, demanding timely and accurate detection to prevent structural failures and costly downtime. This study proposes a lightweight, real-time corrosion detection framework tailored for SHM applications. The framework integrates design elements inspired by the latest YOLOv11 and YOLOv12 architectures while incorporating task-specific optimizations for detecting small, irregular corrosion patterns under diverse environmental conditions. Two curated datasets, augmented with domain-specific transformations, are used to enhance model robustness and generalization. Comprehensive benchmarking against previous YOLO versions (YOLOv3, YOLOv5, YOLOv7, YOLOv8) demonstrates that our optimized YOLOv11m configuration achieves up to 7.7\% improvement in mAP@50 and 12.1\% in mAP@50–95 over YOLOv8m, while the YOLOv12s variant offers a competitive accuracy–speed trade-off. These findings highlight the potential of the proposed approach for deployment in edge-based SHM systems for real-time industrial monitoring.