Geo-Safe AI

Mine risk from multiple kinds of evidence

Geo-Safe AI combines geotechnical telemetry and aerial observations into spatially actionable mine-risk analysis. Microseismic activity, groundwater pressure, slope deformation and imagery describe different aspects of instability. A zone-level assessment joins those signals while preserving the evidence behind an alert.

The architecture separates acquisition, temporal prediction, geospatial interpretation and operator response. A digital twin exposes affected zones and contributing signals. Blast planning occupies its own branch, connecting peak-particle-velocity forecasting to parameter optimization instead of conflating every recommendation with the rockfall risk score.

Technology and role

Scroll horizontally to compare.

Architecture walkthrough

• Edge acquisition groups microseismic fracture signals, piezometer pressure, inclinometer deformation and extensometer movement with environmental context. Drone imagery, LiDAR elevation and ERT moisture observations add geospatial evidence.
• Ingestion normalizes the feeds and associates them with time windows and mine zones. Raster/image preprocessing prepares aerial observations for segmentation and spatial interpretation.
• Temporal models examine changing sensor signatures. U-Net segments visual cracks; the graph branch represents fracture geometry and relationships between neighboring zones.
• The fusion layer combines model outputs with threshold/rule checks. Corroboration across modalities and uncertainty handling shape the risk presented for each zone.
• Operator views expose the affected area and contributing signals. Acknowledgements, mitigation choices and overrides feed the decision record and subsequent review.

Core mechanisms

• Spatial and temporal alignment prevents unrelated image observations and sensor windows from contributing to the same zone assessment.
• Modality-specific models preserve the distinction between visible surface cracks, subsurface pressure and changing deformation signatures.
• GNN relationships capture dependencies across fracture geometry instead of treating every sensor location as independent.
• Cross-signal corroboration helps prioritize alerts when a single noisy channel crosses a threshold. Missing channels affect confidence rather than silently appearing as normal readings.
• The blast branch forecasts PPV from burden, spacing, rock-quality designation, charge per delay and distance. ICA searches the blast parameters separately from the forecast model.
• Acknowledgement and override paths keep alert delivery separate from an operator's mitigation decision.

Engineering decisions

Scroll horizontally to compare.

System deliverables

The SIH 2025 design specifies multimodal acquisition, geospatial preprocessing, model responsibilities, risk fusion and a mine digital-twin interface. It also defines a separate vibration-aware blast-planning path and a decision ledger. The resulting architecture makes the journey from raw measurements to an explainable, spatially located recommendation concrete.