Feature extraction and prediction of granular flow under obstacle influence

Abstract

When encountering obstacles, the flow direction of particles will change. For instance, during the landing of a probe on an asteroid, the motion of the planetary surface's weathered layer will be influenced by large boulders. We can observe these changes through certain feature regions, and extracting and predicting these feature regions can help us make timely adjustments in practical applications. In this study, we conducted impact experiments with different obstacles in granular bed and identified two types of characteristic regions. Using a variant of the UNet++ architecture, we extracted these characteristic regions under various obstacle conditions. The average accuracy of the extraction results was approximately 80% to 87%, demonstrating good generalization and robustness. The extracted characteristic regions were further subjected to prediction using ConvLSTM, resulting in predicted images that exhibit a high structural similarity to the actual images.