Cumulative gangue mixing ratio prediction model for image-based in situ coal/gangue identification

Image-based in situ coal/gangue identification has emerged as a pivotal tool for monitoring instantaneous gangue mixing ratios (IGMR) in fully mechanized top coal caving operations. However, intelligent coal caving control requires dynamic optimization based on the "top coal recovery rate–cumulative gangue mixing ratio (CGMR)" curve. This study establishes a predictive framework linking IGMR to CGMR through numerical simulations and machine learning. The authors proposed a particle swarm optimization–random forest (PSO–RF) hybrid model that outperforms conventional RF, achieving R² values of 0.937 (advancing direction) and 0.962 (layout direction). Feature importance analysis reveals scraper speed, coal caving position, and sequential/interval caving strategies as dominant factors influencing CGMR. Physical experiments validate the model's robustness, demonstrating a 56% reduction in prediction error compared to baseline methods.