Predictive maintenance work order: bu ne anlama geliyor
Predictive maintenance has the clearest business case in industrial AI and the worst record of realising it. The models mostly work. The pilots mostly succeed. The rollouts mostly disappoint.
The failure is almost never in the prediction. It is in the eleven days between the prediction and the intervention.
What happens in those days
The alert reaches a dashboard. Somebody sees it, or does not. If they do, they raise a work order. The work order joins a queue prioritised by other criteria. Parts availability is checked, often manually. A maintenance window is found — but the line is running and stopping it needs production’s agreement.
Each step is reasonable. Together they consume the warning.
The fix is unglamorous
Connect the prediction to the work order system, with the parts check and the maintenance window as part of the same decision. Let the agent hold the whole chain: predicted failure, required parts, available windows, production impact of each.
The model was never the constraint. The handover was.
Why the model is not where this fails
Vibration analysis on a pump is close to a solved problem. A model that flags a developing bearing fault two weeks out is available commercially and works.
The failure is downstream. The alert reaches a maintenance planner who already has a backlog, a shutdown window three weeks away, and no way to tell whether this alert is more urgent than the four from last month that turned out to be nothing.
So the alert joins the backlog. The bearing fails. The post-mortem concludes the model needs tuning.
What the alert is missing
Not confidence. Every model emits a probability and every planner has learned to discount it, because a probability without consequence is not decision-grade information.
What is missing is the connection to the schedule. An alert that says “bearing degrading, 78% confidence” is a fact. An alert that says “this pump feeds line 3, which has committed output on Thursday, and the next planned window is in nineteen days” is a decision.
The second requires the maintenance system to know what production depends on the asset. In most plants that link exists only in the heads of two people.
The work order is the real interface
A predictive maintenance system that does not create work orders will be evaluated on whether its alerts were right, which is unanswerable — an alert that led to an intervention cannot be tested, and one that did not led to a failure nobody attributes to the alert.
A system that creates the work order can be evaluated on outcome: did the intervention happen, did the failure occur, what did the window cost. Those are recorded, and they accumulate into evidence rather than opinion.
The sequence that works
Connect the asset register to the production plan before deploying any model. This is unglamorous data work and it is the step most projects skip.
Then run the model in advisory mode, generating draft work orders a planner accepts or rejects. Rejections are the signal — an alert rejected because the asset has a standby unit is telling you the model is missing a fact that exists in the plant.
Only then automate the work order creation, and only for asset classes where the record supports it.
What good looks like after a year
Not zero unplanned failures; that target drives over-maintenance and costs more than the failures did. The measure is the share of interventions that happened in a planned window rather than an unplanned one, against total maintenance hours.
A plant that moves from 60% planned to 85% planned without increasing hours has captured most of the available value, and it did so through the work order rather than through the model.