Evaluation of Rust Condition in Weathering Steel by Optical Spectra and Machine Learning

Weathering steel reduces the corrosion rate by forming a dense protective rust layer on its surface. Monitoring the rust condition is essential to maintain corrosion resistance. Currently, weathering steel bridges are evaluated through visual inspections conducted by engineers. However, such inspections are qualitative and heavily dependent on inspector experience, presenting significant challenges. Thus, quantitative methods for rust evaluation are needed. One promising approach is the use of optical spectra in the visible to near-infrared region. Previous studies extracted four characteristic wavelengths and applied principal component analysis (PCA) to obtain two principal components for rust classification. However, this method was limited in evaluating the thickness of specific rust types, resulting in insufficient separation. To address this, the present study extracted three principal components, achieving improved classification accuracy. Using the k-nearest neighbor (kNN) method, an evaluation accuracy of 96.6% was attained. Furthermore, a deep learning approach employing a recurrent neural network (RNN) was explored to estimate the rust condition directly from spectral waveforms. Although the RNN-based method did not achieve high accuracy, it indicated potential for application in rust condition assessment.