Abstract:
Affected by obstacles such as reservoir water coverage and plankton,the underwater concrete structural defects of water-related buildings are highly concealed,making it difficult to accurately identify them with conventional manual inspection and engineering geophysical prospecting methods.As a carrier of high-dimensional data information,video images contain a large amount of important information closely related to the evolution of structural behavior such as the health status of the structure,mechanical evolution rules,and response to load and environmental effects.Based on an in-depth analysis of the causes and types of underwater concrete structural diseases,this paper proposes and constructs an engineering structural defect image database SD-ImageNet by integrating on-site collection,open source information acquisition,physical model testing,and other means.Combining inter-domain and intra-domain transfer learning strategies,a two-stage fusion transfer learning strategy was studied and proposed to achieve universal image feature extraction of concrete structures.On this basis,combined with the deep residual network ResNet50,weakly supervised dynamic visualization theory,and two-stage hybrid transfer learning strategy,a vision-driven underwater defect identification method for wading concrete structures is proposed to achieve multi-category defect identification and weak supervision of damaged areas position.Taking a roller-compacted concrete gravity dam as an engineering example,the adaptability and effectiveness of the proposed method are evaluated from both qualitative and quantitative dimensions.