Abstract
Abstract
To explore the feasibility of hyperspectral machine vision for detecting stress states of underground structures, this study utilized a hyperspectral camera to acquire spectral datasets from the surfaces of concrete and sandstone specimens under different stress conditions. Machine learning models were established to predict stress states based on raw spectral data and data pre-processed using the Savitzky-Golay (S-G) method. The results indicated that satisfactory outcomes were obtained with no pre-processing and S-G pre-processing. In addition, the hyperspectral response characteristics of concrete and sandstone under different stress states were investigated. The hyperspectral dataset of sandstone was observed to yield higher predictive accuracy than that of concrete. Finally, a comprehensive analysis of the performances of the principal component regression, partial least squares regression, and least-squares support vector machine models was performed over various datasets in terms of computed model evaluation metrics.