This study aims to address the challenges of concealed corrosion damage, limited samples and strong environmental variability in metal pressure-resistant structures of underwater equipment by proposing a corrosion localization method integrating data augmentation and deep learning to enhance accuracy and reliability under small-sample conditions.
The geometrical configuration of this structure is analyzed, and the geometrical modeling method is proposed. When necessary parameters are determined, such as the structural span, length, vault rise, longitudinal and lateral giant grid number and section height to top chord length ratio of the lattice member, the structure geometrical model can be generated.
Numerical simulations and experimental results yielded a global generalization RMSE of 0.04 (on normalized, unitless data) in ambient air. In a targeted underwater validation for a corrosion defect, a physical RMSE of 2.38 mm was obtained (after denormalization). These findings indicate the robustness and effectiveness of the proposed approach in aquatic environments.
This study presents a novel framework that integrates dual-domain data augmentation with deep learning for corrosion localization under small-sample conditions, providing a high-precision and generalizable approach for structural health monitoring of underwater equipment with significant theoretical and engineering value.
