Multimodal NDT-integrated neural network with cGAN augmentation for predicting reinforcement corrosion in RC structures

초록

Reinforcement corrosion is a major deterioration mechanism that compromises the long-term safety of reinforced concrete (RC) structures. However, existing non-destructive testing (NDT) techniques struggle to detect early-stage corrosion and suffer from limited data availability, constraining the reliability of data-driven models. To address these gaps, this study proposes a novel multimodal artificial neural network (ANN) framework that integrates material, environmental, and NDT features while employing a conditional generative adversarial network (cGAN) for data augmentation. This integration allows the model to learn corrosion-related patterns even under data-scarce conditions. Controlled RC specimens exposed to inland and coastal environments were monitored to establish ground-truth corrosion levels. The augmented model achieved a root mean square error (RMSE) of 3.55 and an R2 of 0.96, outperforming conventional models by maintaining predictive stability with only 30% of real data. Furthermore, the predicted corrosion probability and deterioration levels derived from the proposed multimodal ANN + cGAN framework can be directly linked to infrastructure maintenance decision-making. Structures or components with a higher predicted corrosion risk can be prioritized for detailed inspection or preventive repair, while low-risk areas can be scheduled for routine monitoring. This AI-based prioritization enables more efficient allocation of inspection resources and maintenance budgets, supporting condition-based maintenance rather than traditional time-based inspection cycles.

키워드