Summary
A deep-level gold mine in South Africa operating at 2,800 m depth required a step-change improvement in stope ground control. Hypotenuse integrated microseismic monitoring, extensometers, and stress cells into an AI-driven hazard classification system that reduced rockfall incidents by 79% and enabled dynamic exclusion zone management.
Background & Context
Underground mining at extreme depths exposes workers to severe seismic hazard. Stress redistribution around excavations creates complex failure environments where conventional observation-based risk assessment is inadequate. The mine had experienced 14 rockfall incidents in the previous 12 months, two of which resulted in lost-time injuries.
Technical Problem Statement
At 2,800 m depth, the in-situ stress field exceeds 70 MPa. Seismic events triggered by distant stope blasts can destabilise hangingwall over areas of several thousand square metres within seconds. The challenge was to distinguish harmless routine seismicity from escalating damage sequences that precede major collapses.
Sensor Deployment
AI Model Architecture
A gradient-boosted seismic hazard classifier analyses rolling 4-hour windows of microseismic activity, extracting 23 features including b-value, energy index, apparent stress, and inter-event clustering. The model assigns each production block a hazard class (Low / Elevated / High / Imminent) that drives re-entry decisions after blasts.
Worker Alert & Evacuation Protocol
When a block is classified Imminent, an automatic PA announcement initiates re-entry lockout for the affected section. The mine captain receives a real-time map update on a ruggedised tablet. Re-entry is only permitted after two consecutive Low classifications separated by at least 4 hours, verified by a qualified person and the Hypotenuse dashboard.
Key Outcomes & Results
Rockfall Incidents
Reduced by 79% in the 18 months post-deployment
LTI Frequency Rate
Dropped from 2.3 to 0.4 per million man-hours
Re-entry Efficiency
Average re-entry delay reduced by 35% through evidence-based clearance
Seismic Events Detected
>12,000 events classified — 94% classification accuracy verified
79%
Rockfall reduction
94%
Classification accuracy
2.8km
Mine depth
24
Geophone array
Deployment Snapshot
Location
South Africa, Witwatersrand Basin
Mine Depth
2,800 m
Sensor Count
56 instruments
Deployment Phase
18 months