Intelligent identification of force chain particles within hang-up system of caved ore and rock based on graph neural networks

Hang-ups frequently occur during caving mining operations. Force chains are essential for bearing and transmitting stresses, making them crucial for characterizing hang-ups. However, efficiently and accurately identifying force chain particles directly from the particle system presents a significant challenge. To address this issue, this study developed six deep learning models and evaluated their performance on 1206 datasets under varying conditions, including the ratio of drawpoint size to average particle size, the particle friction coefficient, and the overburden stress. The models were compared based on their performance in terms of convergence, accuracy, and feature extraction and analysis, following force chain particle identification. This comparison led to the identification of the most effective neural network model for processing hang-up data in the particle system of caved ore and rock. The results indicate the following: (1) Incorporating graph structural information within the same network architecture resulted in slower training and testing speeds; however, it improved model accuracy by 5% to 10%, with graph neural networks generally outperforming traditional neural networks. (2) The Graph Attention Network (GAT) model exhibited convergence that was 8% to 71% better than the others, with its identification results on the test set being 1% to 65% superior, and its extraction results demonstrating the highest consistency with actual conditions. (3) In the results from the GAT model, identification errors for particle size and contact force were both below 3%, and its self-attention mechanism effectively captured particle size characteristics while also delineating contact relationships between particles.