Abstract
Accurate rock fragmentation and size distribution analysis are essential in mining operations for optimizing efficiency, reducing operational costs, and enhancing safety. Proper fragmentation management minimizes overbreak, reduces the need for secondary blasting, and improves the efficiency of downstream processes like loading, hauling, and crushing. This, in turn, optimizes energy usage and reduces maintenance costs.
Traditionally, image-based methods have been used for rock fragmentation analysis. However, these methods are limited to 2D perspectives and depend on lighting conditions, making them less effective in low-visibility environments, such as underground mines. To overcome these limitations, our study introduces a 3D point cloud-based approach using laser scanning and deep learning-driven point cloud segmentation. This allows for a comprehensive, lighting-independent assessment of rock fragmentation in three dimensions.
One key challenge in applying deep learning models to mining is the lack of labeled data for training. To address this, we created a synthetic dataset generation platform that produces labeled 3D point clouds of rock fragments, enabling precise instance segmentation and size analysis. This synthetic data supports the development of deep learning models capable of delivering highly accurate fragmentation assessments and detailed volumetric analysis critical for optimizing resource utilization. The model's accuracy was tested against a blasted rock pile scanned by a LiDAR scanner, which was used as ground truth.
The incorporation of multi-point scanning is shown to be crucial for achieving high accuracy in size distribution analysis. Scanning from only one side of the rock pile, similar to taking a picture from a single perspective, results in significant information loss, either omitting entire objects or capturing only partial views of visible rock fragments. This limitation leads to a notable reduction in fragmentation accuracy. We showed that a single-sided scan produced poor alignment with the ground truth rock size distribution. However, by adding scans from multiple points, the accuracy of rock size distribution can be significantly improved, achieving results that closely match the ground truth distribution.
In conclusion, our multi-point, 3D point cloud-based method provides an advanced, accurate solution for rock fragmentation analysis, overcoming the limitations of conventional techniques and supporting efficient, data-driven mining operations.