Automated detection of underwater dam damage using remotely operated vehicles and deep learning technologies

Abstract

Underwater damage poses significant risks to the safe operation of dams, making timely detection critical. Traditional manual inspection methods are hazardous, time-consuming, and labor-intensive. This paper introduces an automated detection system integrating remotely operated vehicles (ROVs) and enhanced deep-learning technologies. The proposed YOLOv8n-DCW model incorporates deformable convolution networks, coordinate attention mechanisms (CoordAtt), and an improved loss function to boost detection performance. Trained on an underwater dam damage dataset, the model achieved an 84.5 % mean average precision. Ablation studies validated the effectiveness of these enhancements, while comparative experiments demonstrated the superiority of YOLOv8n-DCW over existing models and CoordAtt's advantage among attention mechanisms. The developed detection software, integrated with the ROV, was tested in a laboratory pool, confirming its practicality and efficiency. This system offers a safer, faster, and cost-effective solution for underwater dam damage detection, addressing limitations of traditional methods and providing a robust tool for engineering applications.