This prototype does not use a trained model. It applies ordered heuristics to synchronized longwall and bunker telemetry plus observed or calibrated outby context, then turns the inferred state into a surface-facing recommendation. The live Why the model thinks that list in the action panel is generated from these same branches.
1. Signals the rules look at
Face activity
Shear position, shear motor amps, AFC head amps, crusher amps, and stage-loader amps determine whether the face is actively cutting and which direction the pass appears to be moving. The current prototype also watches for short level-offs near the ends of travel.
Outby flow
Outby belt amps, bunker motor amps, IP98, bunker infeed, and bunker outfeed determine whether material is still moving through the bunker and belt network. Belt/IP98 are directly observed only in the validated overlap window and estimated elsewhere.
Hold vs switch risk
Bunker occupancy, drain rate, shear variance, and seeded review hints determine whether the system should keep clearing, prepare for a switch, or route a segment for review.
3. Ordered state inference
Unknown / anomalous
If the cut is idle, the outby system is quiet, and bunker level is below about 8%, the app does not force a material-state call.
Hold / clearing
If face signals fall away but belts, bunker motor, or IP98 are still live, the app treats the system as clearing. If bunker-clear time stays above about 2.2 minutes, it remains a hold interval.
Transition imminent
If bunker-clear time drops below about 2.2 to 2.8 minutes, bunker occupancy is below about 14% while outby flow is still active, or cutting has not committed strongly enough to one direction, the app marks a pending switch window.
Clean pass
Reverse shear travel faster than roughly 3 position units per minute, plus active cutter, AFC, crusher, and stage-loader load, maps to the learned clean-pass direction. If the shear pauses in the high-end band first, the prototype can start the clean pass after that pause instead of forcing the boundary exactly at the peak.
Dirty pass
Forward shear travel faster than roughly 3 position units per minute, plus active cutting and heavier cutter load, maps to the learned dirty-pass return signature. If the shear pauses in the low-end band first, the prototype can start the dirty pass after that pause instead of forcing the boundary exactly at the valley.
These checks are evaluated top to bottom. The first matching envelope wins, then confidence is adjusted from signal strength, directionality, and any recent end-of-travel pause that matches the current geology baseline.
4. Current geology baseline
Dirty-pass start band
--
When shear position settles in this low-end band and then resumes upward travel, the prototype can treat that release as the dirty-pass boundary.
Clean-pass start band
--
When shear position settles in this high-end band and then reverses, the prototype can treat that release as the clean-pass boundary.
Pause detection and fallback
--
--
This is a lightweight stand-in for the future geology input function. The numbers can move as seam conditions and expected signal shapes change.
5. When the app asks for review
Confidence is too weak
Segments are flagged if average confidence drops below 56% or the state resolves to unknown.
Telemetry conflicts with the envelope
Examples include heavy clean-pass cutter load, dirty-pass direction with unexpectedly light AFC head load, or active material load while the belt-run signal is off.
Clearing looks materially slow
Hold and transition windows that average more than 8.8 minutes, or bunker-clear estimates above 10.5 minutes, are escalated for review.
6. Recommendation mapping
Needs review
Route to Andy confirm / correct review
Clean pass
Continue clean cut
Dirty pass
Protect dirty window
Hold / clearing
Hold and clear bunker
Transition imminent
Prepare for product switch
Unknown / ambiguous
Review before switching