Classifying corrosion damage in reinforced concrete building using machine learning based clustering of electrochemical data - CANCELLED (PRES2026-00097)
As the inventory of corroding reinforced concrete (RC) structures is increasing, it is important to develop technologies to reliably assess the condition of such structures. Conventional electrochemical methods, such as half-cell potential and surface resistivity, classify corrosion risk using code-based thresholds. However, these thresholds often fail to capture corrosion in highly resistive concrete or when exposure conditions vary within members. In this study, a machine learning–based approach was adopted for classifying the corrosion damage on structural members in a building with varying temperate climate and moisture conditions. Instead of relying solely on raw measurements, additional electrochemical indicators such as the gradient of half-cell potential(delta) and conductance per unit area were derived. This derived information provided a more direct representation of corrosion. Principal Component Analysis was applied to these derived and measured parameters, and the most critical features were selected for clustering the data. Using these features, K-means clustering was employed to classify the electrochemical data into three levels of corrosion (i.e., low, moderate, and high corrosion). The classification was verified using photographs of deteriorated RC members and the results show that clustering-based classification, when supported by meaningful electrochemical indicators, provides a practical and more reliable alternative to conventional threshold-based classification.
