Multi-modal NDE data fusion with machine learning for reinforced concrete corrosion damage classification

Early detection of corrosion-induced deterioration in reinforced concrete (RC) is critical for effective maintenance of aging infrastructure. This study presents a machine learning (ML)-driven framework integrating multiple nondestructive evaluation (NDE) techniques—electrochemical, electromagnetic, and acoustic—for automated classification of corrosion-induced damage in RC. Three concrete mixes were subjected to accelerated corrosion, and NDE data were collected between exposure cycles. Damage states were categorized into three levels: intact, internal deterioration, and visible surface cracking (≥ 0.1 mm). ML models using fused multi-modal NDE inputs outperformed single-technique models, particularly in detecting early-stage damage. The best-performing model achieved an accuracy of 0.936 and a Cohen’s Kappa of 0.783 through feature-level fusion of five NDE parameters—half-cell potential, relative permittivity, electrical resistivity, impact-echo peak frequency, and ultrasonic surface wave velocity—using the Subspace KNN classifier. Robustness was validated using independent specimens with distinct mixture properties. Additionally, a decision-level fusion model employing weighted soft voting was developed to algorithmically exploit the complementary sensitivities of individual NDE parameters, achieving a synergistic enhancement of diagnostic performance. A simplified configuration using only half-cell potential and ultrasonic surface wave velocity achieved 0.832 accuracy and 0.607 Kappa, with a class-wise Kappa of 0.674 for incipient internal deterioration, the highest among all model combinations. Overall, the proposed framework improves early-stage corrosion detectability and supports more reliable maintenance decisions for aging concrete structures.