Why mine safety monitoring produces false alarms
At most mines, mine safety monitoring is built around control posts: a motion sensor in the drift, a tag reader, a camera at the junction. The trigger threshold is the same for everyone — the system registers "movement in the zone" and issues an alarm.
Such a loop cannot distinguish whether a person is walking along the working, an LHD machine is moving, or rock is falling behind the support.
This is where dozens of false alarms per shift come from. Dust, steam from water, vibration from drilling rigs, a swinging cable in the frame — each factor crosses the threshold as an "event," although there is no event.
The requirements of Safety Rules for Mining Operations are formally covered, but not in fact: personnel monitoring in the workings turns into a stream of identical marks that no one reviews.
The cost is not in the number of alarms, but in the dispatcher's reaction. After a hundred empty signals, attention dulls: the operator confirms them without looking, and some zones drop out of review. Blind spots appear not where there are no sensors, but where false data continuously floods them.
We explain how this is addressed: object classification on the analytics side, tying an alarm to a specific event, and priorities under which a signal about a collapse or a person falling does not compete for attention with vehicle passage.
What changes after implementation: effect in measurable values
We calculate the effect by shift parameters: incident detection time, share of false alarms, coverage of control zones, and the failure class that stops going unnoticed. We set alarm thresholds according to the critical risk control methodology that ICMM relies on.
Detection time and shift response speed
Previously, a signal about a person in a hazardous zone or equipment at the side arrived via report — tens of minutes, longer on the night shift.
Video analytics issues an alarm within the first seconds, tied to a zone and timestamp, so the time from event to start of action becomes measurable and the same across all areas.
Share of false alarms and load on the dispatcher
Threshold sensors trigger on dust, headlights, and vibration — the dispatcher gets used to ignoring the stream.
A classifier on the video stream separates event from interference: the share of false alarms drops several times, and the load — from hundreds of signals per shift to dozens requiring a decision.
Coverage of control zones
Pit walls, dumps, warehouses, and temporary ramps rely on rounds: there is no wired communication or stationary sensors there. Portable control points cover these areas, and coverage grows to the full work perimeter — including zones that appear when the front moves.
Eliminated failure class
The failure class "no one noticed" goes away: channel break, dusty lens, disconnected sensor. Equipment self-check and signal liveness monitoring turn this from an accident into a maintenance request, and the shift sees which control points are currently not working.
Implementation formats: stationary loop, mobile posts, hybrid
The implementation format is determined by the facility geometry, access to power and communication, and which zones are kept under constant observation and which only during work. At an operating mine and at an open pit, the same set of models is deployed differently.
| Option | When suitable | Limitations |
|---|---|---|
| Stationary control loop | Permanent high-risk zones: shafts, inclines, unloading, explosives storage | Requires power and communication lines, installation in windows between cycles |
| Mobile posts | Mobile front, temporary areas, exploration | No permanent power, limited autonomy |
| Hybrid scheme | Operating facility with stable core and mobile zones | Two loops, different maintenance regulations |
Stationary loop at an operating facility
The loop is mounted on existing supports, gantries, and cabinets; power and communication are taken from already laid routes, so production does not need to be stopped — work proceeds in technological windows.
Permanently covered high-risk zones produce fewer failures: connectors are not re-plugged, calibration is not disturbed, the camera stream is predictable.
Hybrid scheme with portable posts
A core of stationary points covers key zones, while mobile posts are moved following the mining front.
There is one compatibility requirement: a model trained on stationary streams must work on a portable post without retraining — otherwise each movement will require re-validation on site.
Alarm rules and the event collection layer are common, so the dispatcher sees a single picture regardless of where the signal came from.
Implementation stages: from route audit to handover to the shift
Work proceeds sequentially, and each stage relies on data from the previous one. We start not with installation, but with analyzing how equipment and people actually move along levels and benches: dump truck routes, unloading zones, shift change points between areas.
Route audit and post placement scheme
- Route audit — we walk the facility together with the mining foreman and occupational safety service: blind spots, flow intersections, areas with limited visibility.
- Post placement scheme — we calculate how many observation points the route covers without overlaps, determine installation locations, power, and communication channel.
- Scheme approval — we mark responsibility zones by shifts so that the signal goes to a specific person, not "to the dispatcher's office."
Scenario setup and on-site testing
- Scenario setup — for each area we define rules: presence in a hazardous zone, absence of a high-visibility vest, driving in the oncoming lane, abandoned vehicle.
- On-site testing — we run scenarios during a real shift, including night, dust, and precipitation; we adjust trigger thresholds to remove false messages.
- Shift training — dispatchers and foremen go over the response to each event type and the escalation procedure at their workplaces.
- Handover to operation — we record settings, provide regulations and change log, and remain for support after launch.
How video analytics differs from acoustic monitoring
The methods catch different event classes, so the question "which is better" is incorrect: at a facility they solve different tasks and more often work in tandem — one provides early detection, the second confirmation and localization.
Video analytics recognizes people and equipment in the frame, calculates trajectories, and records crossing of hazardous zone boundaries.
In dusty environments and fog, contrast drops, and in darkness without light the method goes blind: illumination helps at short distances, thermal channel — at medium distances, but with dense suspension it loses detail.
Acoustic monitoring works where the camera is useless. Microphone arrays with dust protection compare the spectrum with a reference profile: impact, metal scraping, rock collapse, abnormal engine operation are audible at zero visibility.
There is one limitation — background equipment noise, so thresholds are calibrated for a specific open pit.
| Method | What it detects | What it misses | Dust, fog, darkness |
|---|---|---|---|
| Video analytics | people and equipment in the zone, boundary crossing | events outside the frame and behind obstacles | contrast drops; illumination or thermal channel needed |
| Acoustic monitoring | impact, scraping, rock collapse, abnormal equipment operation | quiet events, visual assessment of the situation | works regardless of light and visibility, limited by background |
The choice of monitoring method is dictated by the facility: in open areas video covers the perimeter, acoustics — pit walls and faces; in unlit workings acoustics becomes the main loop.
The coincidence of two signals in time and place gives a confirmed event and removes most false alarms.
What we hand over: configurations, schemes, shift regulations
We hand over the system not as a black box, but together with settings that allow it to be developed in-house. After implementation, the customer retains a set of artifacts — from scenario configurations to shift regulations.
Placement schemes and control zones
The scheme records which sections of the shaft, drift, and open pit fall into camera view, where sensors are located, and how control zones overlap. It shows which blind areas remain and what needs to be additionally installed to cover them.
- Scenario configurations — a separate set of rules for each event class: person fall, entry into hazardous zone, reverse movement of equipment, presence under load. The file defines trigger thresholds, confirmation time, and notification priority.
- Placement schemes — binding of cameras and sensors to the mine plan with marking of control zones and overlaps.
- Recording access rules — user roles, retention periods for fragments, procedure for export for review.
- Response regulations — to whom and within what time the notification goes, what is recorded in the log, when the shift dispatcher joins the review.
- System documentation — solution passport, node composition, failure points, model update procedure.
- Quality report — for each scenario, the share of false alarms and misses is visible, so thresholds can be adjusted without us.
Documentation and shift regulations
Documentation is written in the language of the shift, not the developer: where to view the alarm, how to clear it, what to do when signal from the area is lost.
Regulations are broken down by roles — dispatcher, mining foreman, occupational safety service — and are updated together with scenario configurations.
Case study: open pit with dust and fog — indicators before and after
What it was like before implementation
The open pit operated around the clock, and monitoring of hazardous zones was carried out by a shift of two operators using a general overview.
The bench wall, unloading zone, and haul road were captured unevenly: part of the sectors were blocked by stockpiles and terrain, at night the picture was lost, and in fog and dust after blasting, contrast dropped so much that the silhouette of a person and a dump truck merged with the background.
A violation was noticed on average in 4-6 minutes — during this time equipment could pass through the area where people were.
Out of several dozen alarms per shift, more than half were caused by dust, fog, and headlights, so the signal began to be perceived as noise: reaction to a real event was blurred.
What changed after
Camera angles were rearranged so that each hazardous zone could be read from at least two directions, and thermal imaging modules FLIR were installed in areas with constant dust.
The model was trained on local footage, including night and foggy shifts, and it identifies people and equipment by contour and position in the zone, rather than by pixel brightness.
Time from entering the zone to the operator's command was reduced to 8-12 seconds. The false alarm rate fell from about 65% to 12%, coverage of hazardous zones increased from 11 to 17 out of 18 areas, including those previously blocked by terrain.
The main result of implementation is not the numbers themselves, but the restored trust in the signal: each alarm is investigated as an event, rather than dismissed with the wording "dust."
What hinders stable monitoring: communication, light, false alarms
Monitoring stability rests on three things: the communication channel with the underground area, conditions in the working itself, and a correctly selected trigger threshold. Below are typical doubts that come to us before starting, and how each is addressed engineering-wise.
What happens if communication with the underground area is lost?
The node in the area continues to measure and stores data in a local buffer, while a clear "no channel" indication goes to the dispatcher's office — the operator sees not a blank screen, but a specific state.
After restoration, the buffer is uploaded in the original order, so the shift report remains complete and suitable for review.
How does the solution work in low light and dust at the face?
Dust scatters light and reduces contrast, so an alarm cannot be based on a single image. We confirm the event with data from other sensors, and the illumination and recording mode switches according to the actual light level in the working.
In the dustiest areas, we plan more frequent maintenance in advance — this is cheaper than investigating a missed event.
How is the trigger threshold adjusted so that alarms are not false?
We select the threshold not by eye, but based on accumulated statistics for the area: we look at the spread of values across shifts and set the boundary above everyday noise.
Plus hysteresis and confirmation by two independent signs — a short spike does not raise an alarm, while a real deviation is recorded. We keep the configuration versioned: if a new setting caused a spike in false alarms, we return to the previous one without stopping monitoring.
If you still have questions about the channel, lighting, or thresholds for your workings — describe the area, and we will come back with an assessment and inspection plan.Let's discuss the task: routes, regulations, post composition
The timeline and scope of work on the safety monitoring system are determined not by technology, but by the facility configuration: number of zones, nature of processes, shift organization.
The more accurate the input data, the fewer assumptions in the assessment and the fewer clarifications that surface during commissioning.
- Facility scheme: plans of levels, benches, haul ramps, shafts, and ventilation workings — the basis for the control zone map.
- Shift and shutdown mode: shift change schedule, blasting times, planned downtime — set the schedule for rule triggering.
- Problem areas: sections with frequent violations and equipment failures — priority for cameras, sensors, and notification logic.
- Occupational safety requirements and internal regulations: what is considered a violation, who responds and within what time, what goes into reporting.
- Alarm output points: dispatcher's office, security console, mobile devices of responsible persons — where, in what form, and how the signal is confirmed.
- Composition of posts and patrol routes: where personnel are located, how equipment moves, which flows of people and machines need to be distinguished.
- Access and installation windows: access control, on-site escort, time when work can be carried out without stopping production.
- Existing infrastructure: installed equipment, available communication channels, power limitations.
Based on this data, we assemble a work plan: stages, delivery scope, acceptance points. Describe the task and attach what you already have — we will come back with a timeline assessment and a stage-by-stage plan, with specific deadlines and responsible persons.
