Dual-image attention convolutional network for monitoring the shape of embankment materials during dam construction

Abstract

This research explores the influence of particle roundness on the macro-mechanical behavior of soils, emphasizing the need for effective classification methods. Traditional approaches, including Wadell’s 2D-based roundness and computational geometry (CG) techniques, are hindered by inefficiency, subjectivity, and sensitivity. To address these challenges, the study introduces a novel solution using a dual-graph attention convolution network (DGACN) for 3D point cloud classification. A dataset of 2400 soil particles, scanned via X-ray computed tomography, is utilized to train and evaluate the DGACN model. The results demonstrate an accuracy of 90.1%, showcasing the model’s robustness to defective data and its ability to accurately classify six roundness classes. Furthermore, the DGACN approach outperforms traditional CG methods in computational efficiency, being 53 times faster. This work establishes deep learning as a powerful and efficient tool for soil particle characterization, offering valuable contributions to geotechnical engineering and materials science research.