Deep Learning-Based Estimation of Muckpile Fragmentation Using Simulated 3D Point Cloud Data

This research introduces an innovative technique for estimating the particle size distribution of muckpiles, a determinant significantly affecting the efficiency of mining operations. By employing deep learning and simulation methodologies, this study enhances the precision and efficiency of these vital estimations. Utilizing photogrammetry from multi-view images, the 3D point cloud of a muckpile is meticulously reconstructed. Following this, the particle size distribution is estimated through deep learning methods. The point cloud is partitioned into various segments, and each segment's distinguishing features are carefully extracted. A shared multilayer perceptron processes these features, outputting scores that, when consolidated, provide a comprehensive estimation of the particle size distribution. Addressing the prevalent issue of limited training data, this study utilizes simulation to generate muckpiles and consequently fabricates an expansive dataset. This dataset comprises 3D point clouds and corresponding particle size distributions. The combination of simulation and deep learning not only improves the accuracy of particle size distribution estimation but also significantly enhances the efficiency, thereby contributing substantially to mining operations.