BCMMINING TECHNOLOGY
Home / Platforms / Operations Advisor
Threshold alerting needs somebody to know the number in advance. Most of what quietly costs an operation money has no number: a haul route that got slower, a drill that dropped off its own average, a face that stopped behaving like itself. This compares every unit of work against its own history and writes up the departures, with the effect size, the sample it is standing on, and how confident it is.
In practice
Two findings on the same haul route, at different points in the shift. Nobody configured a limit for either. The route was measured against its own median across several hundred prior cycles, and the wording, colour and confidence all move with the size of the departure.
What to notice: a second signal has to agree before either card appears, queue and delay time on the same route, which is what separates a genuinely slower haul from ordinary variation. The sample count is on the face of the card so a supervisor can weigh it in seconds. And the suggestion names what would make it a false alarm, a recent routing or dispatch change. An advisory that tells you how it could be wrong is one people will act on.
And that they are graded: five percent is amber, headed Attention, at 85 percent confidence. Fourteen and a half percent is red, headed Action, at 90 percent. Small drift gets noted rather than escalated, so a channel stays triageable by scrolling instead of reading.
Screenshots from a production system. Destination identifiers have been altered; figures, baselines and sample counts are unchanged.
The baseline
The advisor cannot report anything until it knows what your operation's normal is, and a baseline is not something that can be shipped in a box. It has to be accumulated from your own completed work. Several hundred cycles on a route before a comparison is worth making, which a busy haul reaches in days and a road to a rarely used dump takes considerably longer. Through that period the service is running and building history, and it says nothing.
This is worth being direct about, because it is the honest test of any advisory system sold into mining. Anything that starts making confident claims about your operation in its first week is not comparing your operation to itself. It is comparing you to a model built somewhere else, on somebody else's pit, somebody else's fleet, somebody else's haul profile. The waiting is what makes the output about you.
The service reads completed work and builds a per-unit history. Nothing is published and nobody is notified. If your warehouse already holds cycle history, this phase is largely already done and the runway shortens accordingly.
Findings are generated but held back for review rather than posted to a channel. You see what it would have said, and the sensitivity gets set against what your supervisors actually found when they went and looked.
Advisories start arriving in the channel, graded. The baseline keeps moving from here, so the reference is always recent history rather than a number frozen at commissioning.
Because it is recomputed rather than fixed, the comparison stays valid as the pit deepens, routes change and the fleet turns over. A threshold set at commissioning is wrong within a quarter and nobody notices.
Mechanism
Scope
The cards above happen to read a Cat Minestar dataset, but the method needs only a history of completed work with a start, an end, a duration, and somewhere the time went. Every fleet management system produces that. Once those records are in your own warehouse, the advisor reads the warehouse rather than any vendor's schema, so it survives a fleet system change instead of being written off with it.
Drill patterns against their own metres per hour. Crusher throughput by feed source. Loading tool performance on a given face. Plant recovery by ore type. Pump runtimes, tyre life by route, fuel burn by machine and grade. Wherever an operation has a normal it never wrote down, there is something here to compare against.
The advisor is a consumer of the governed warehouse rather than a separate stack. If the shift calendar, the rehandle logic and the coordinate transforms are already settled there, the advisory inherits definitions everyone already agrees on.
Why it pays
A route five percent slower breaches no threshold, appears in no exception report, and costs money every shift until somebody happens to notice. Multiply it across the routes an operation runs and it is the largest category of loss nobody has a name for.
Even where a limit exists, it was set once, by somebody who has usually moved on, for a pit that no longer looks like this. A learned baseline does not need revisiting because it revises itself.
Industry benchmarks and vendor models describe an average operation. Your haul profile, your grades, your road standards and your fleet mix are not average. Comparing an operation to itself is the only comparison it cannot argue with.
In development
A machine learning model is being built to run next to the statistical layer described above. Alongside is the important word, and it is a deliberate choice rather than a transitional one.
Everything on this page so far is explainable. Each advisory states its own arithmetic: this median, against that baseline, across this many cycles, corroborated by that second measure. A supervisor who disagrees can disagree on the numbers, and be right sometimes. A model cannot offer that, and an advisory nobody can audit gets filed as an opinion no matter how good it is. So the transparent layer stays the part that talks to people.
What a model adds is the work statistics cannot do: patterns across many variables at once rather than one measure against its own history, earlier warning from conditions that precede a problem instead of accompanying it, and learning over time which findings actually proved worth acting on when somebody went and looked. That last one matters most, because precision is what decides whether a channel keeps getting read.
Delivery
Next step
If you can answer that, you did not need this. The interesting version of the question is how you would know, and that is usually the whole conversation. Tell us what history your systems keep and the runway becomes obvious quickly.