Title: Advancing from 3D Perception to Safety Reasoning: A Graph-Driven Approach for Real-Time Mine Surveillance

Abstract:

The environment of underground coal mining presents significant safety challenges, as personnel and heavy machinery must navigate confined, dimly lit, and shared spaces. In these settings, risks such as structural instability, equipment proximity breaches, and obscured blind spots are often hard to predict. Traditional monitoring tools, such as fixed cameras and rule-based proximity warnings, are limited in their ability to recognize complex or evolving dangers because they lack the capacity for 3D scene comprehension and contextual memory.

This study introduces a continuous monitoring framework designed to transform colorized 3D point clouds into structured, traceable outputs for safety reasoning. The system integrates several key technologies: 3D semantic perception, anomaly detection based on uncertainty, rule-based hazard verification, on-device Large Language Model (LLM) reasoning, and GraphRAG-based memory analysis. This combination allows for the identification of immediate threats while also interpreting long-term safety trends. Scene and temporal graphs function as the explicit knowledge backbone, connecting perception outputs across various reasoning phases.

To address the shortage of labeled data in underground environments, the researchers synthesized diverse hazard scenarios by merging real roadway scans, controlled object placements, and high-fidelity longwall simulations. Additionally, self-supervised pretraining was employed to enhance segmentation performance despite limited annotations. The resulting perception model delivered 92.7% accuracy at a speed of 30 frames per second while maintaining low memory consumption.

In testing across 115 hazard scenarios, rule-based checks alone covered 57% of cases. This coverage expanded to 76% when contextual LLM reasoning was applied, and reached 93% when memory-based reasoning utilizing historical records was included. Qualitative findings indicate that anomaly signals derived from uncertainty help interpret out-of-distribution hazards that fall outside predefined categories. Ultimately, the integration of graph-based knowledge representation with 3D perception and layered safety reasoning offers a robust foundation for intelligent decision support in underground mine monitoring.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC