Artificial-intelligence-driven fragmentation assessment for safer block caving operations

Authors: Ziinaa, T; Enkhbayar, M-E; Sanaakhorol, M; Boldbaatar, A; Nasantogtokh, K; Ganbaatar, U; Soyolbaatar, B; Battulga, B Download Paper Open access courtesy of:

DOI https://doi.org/10.36487/ACG_repo/2645_30

Cite As:
Ziinaa, T, Enkhbayar, M-E, Sanaakhorol, M, Boldbaatar, A, Nasantogtokh, K, Ganbaatar, U, Soyolbaatar, B & Battulga, B 2026, 'Artificial-intelligence-driven fragmentation assessment for safer block caving operations', in A van As, D Cumming-Potvin & J Wesseloo (eds), Caving 2026: Proceedings of the Sixth International Conference on Block and Sublevel Caving, Australian Centre for Geomechanics, Perth, pp. 1-12, https://doi.org/10.36487/ACG_repo/2645_30



Abstract:
a Oyu Tolgoi LLC, Mongolia In cave mining, drawpoint rock fragmentation influences operational safety, particularly with respect to inrush risk in block and panel caving operations. While fragmentation is linked to overall caving behaviour, the size distribution observed at drawpoints is often modified by material flow and progressive breakage within the caved column. Despite this, drawpoint fragmentation remains a key control on operational performance, influencing inrush hazards (both dry and wet), dilution entry, productivity and secondary blasting requirements. Adverse rock fragmentation – especially excessive fines and poorly distributed fragment sizes – increases the potential for inrush events, equipment damage and personnel exposure to hazardous conditions. Reliable assessment of drawpoint fragmentation is therefore essential for effective inrush risk mitigation and sustained production performance. Despite its importance, fragmentation assessment in underground environments remains challenging. Visual inspection is widely used but is subjective and inconsistent, while existing commercial tools are unsuitable due to safety constraints and reliance on physical scale references. AI-based approaches offer potential but are limited by extensive labelling requirements. While foundation models can reduce this dependency, they demand substantial computational resources. To address these challenges, we propose a computationally efficient and accurate fragmentation analysis system comprising three components: an offline image collection application, a rock detection model and a web application. The rock detection model is trained using Mask R-CNN and fine-tuned on 872 labelled images to accurately identify and categorise rock fragments. The model achieves 96% precision and processes images in 20–50 seconds, a 720-fold speed improvement compared to foundation models. Proof-of-concept trials across more than 200 drawpoints in Panel 0 at Oyu Tolgoi demonstrate the feasibility of this low computational effort, AI-driven workflow for assisting geotechnical engineers with rock size distribution analysis, drawpoint monitoring and operational decision-making in underground block caving operations.

Keywords: rock fragmentation, AI, computer vision, rock detection, rock segmentation

References: