Reducing unplanned downtime: what your inspection reports already say
What inspection reports already contain about the failures ahead, how to surface it, and why the signal existed long before the plant went down.
After an unplanned shutdown, the investigation tends to reach the same conclusion: the signal was there. It sat in an inspection report read eighteen months earlier, in a caveat raised by a contractor, in a measurement that had already started to drift. Nobody missed it through negligence. It was simply buried inside a document that nothing connected to the ones that followed.
The essentials
Most unplanned downtime on monitored equipment does not come from a failure to detect. It comes from a failure to connect the dots. The information was there, scattered across several reports, with no mechanism to bring it back to the surface. Reducing this kind of downtime therefore rarely starts with new sensors: it starts with using what has already been measured and written down.
The three families of unplanned downtime
They do not call for the same responses, and confusing them leads you to invest in the wrong place.
Random failures. A sudden rupture, a component defect, an external event. No amount of monitoring announces them in advance. The right responses are redundancy, a spares holding and a shorter time to restore service, not measurement.
Undetected progressive failures. The degradation was under way, but nothing was measuring it. The response here is to instrument or to inspect, in other words to add a monitoring policy. See preventive, predictive and condition-based maintenance.
Detected but unused progressive failures. The degradation was already in the reports, but it was never tracked over time or brought to a decision. This is the most common family on regulated equipment, and the only one that costs nothing to address: the data already exists.
It is this third family that repays the effort most, and it is the one that receives the least attention, precisely because you cannot fix it by placing an order.
Why the signal never resurfaces
An inspection report is read once, when it arrives, and then filed. That single reading sits at the heart of the problem, for three reasons.
First, an isolated finding looks harmless. A minor caveat, a reading slightly below the previous one, an observation with no clear qualification: nothing that warrants immediate action. These items only take on meaning as a series, and the series is never reconstructed. A wall thickness that has dropped by a fraction of a millimetre means little on its own; the same drop repeated across three campaigns describes a corrosion rate that is worth acting on.
Then, the reader changes. The 2022 report was read by one person, the 2025 report by another. The memory of the first does not travel with the file. What one engineer understood as a live concern becomes, for the next, just another archived PDF.
Finally, nothing triggers a second reading. No mechanism says "this finding has been open for three years" or "this reading is drifting away from the trend". Without a trigger, the information stays true and inert, technically available yet practically invisible.
What a historical review reveals
When you go back over several years of reports across a plant and reconcile them by equipment and by condition monitoring location (CML), three things come out every time.
Findings that were never closed out: caveats raised by a contractor, treated at the time as observations, where nobody can now say whether they were ever resolved. They appear in no tracking table because they were never extracted from the report in the first place.
Locations that are drifting: a handful of CMLs whose loss is accelerating relative to the average for the plant. Read one at a time, each reading stayed within tolerance. Read as a series, the change of slope is plain to see.
Equipment with no recent data: items whose last inspection is older than the plan allowed for, without that overrun having been flagged anywhere.
None of these three categories calls for new measurements. They only call for reading together what was written separately.
What actually brings unplanned downtime down
Reconcile the history by equipment
Move from filing by date and by contractor to filing by equipment. This is the single step that turns a stack of documents into a history. See centralise your inspection reports.
Make findings trackable
A finding needs a state: open, in progress, closed. As long as it remains a sentence inside a PDF, it cannot be counted, chased or presented at audit. A finding you cannot count is a finding you will forget.
Surface the exceptions, not the data
A dashboard that shows every measurement will not be looked at. A dashboard that shows the five items whose trend has worsened since the last campaign will be looked at. The value lies in the filter, not in completeness. The point of the exercise is to spend attention where it changes a decision.
Tie the anticipation to the shutdown windows
An estimated due date is only worth anything if it lands before the decision window: booking a contractor, ordering a long-lead part, adding the item to the scope of a turnaround. Getting three months of warning when the slot has to be reserved six months ahead achieves nothing.
Document the deferrals
When a job is put back, the decision must be recorded together with its justification. This is what separates an accepted risk from a forgotten one, and it is what protects the team on the day the failure arrives.
What does not work
- Adding sensors before using what you already haveinstrumenting a plant whose current readings are not being reconciled simply stacks one unread source on top of another.
- An exhaustive dashboardthe more indicators it carries, the less it gets consulted. A useful dashboard fits on five lines and flags exceptions.
- Promising a percentage reductionthe effect on unplanned downtime is measured over several years and depends on the life cycle of the equipment. Announcing a one-year figure discredits the whole effort at the first review.
- Treating random failures with monitoringthat is spending with no effect. The answer lies elsewhere: redundancy, spares, time to restore service.
- Relying on the team's memoryit works remarkably well, right up to the first departure. Anything not written down and reconciled leaves with the person.
How to measure progress honestly
The count of unplanned shutdowns is a poor indicator in the short term: it depends too much on chance and on equipment life cycles. Three intermediate indicators tell you sooner whether the effort is having an effect.
The number of open, unresolved findings, and their age. It should fall, and it is entirely within your control. If it is not moving, nothing downstream will move either.
The share of jobs decided in advance rather than in reaction. This is the real turning point: fewer jobs triggered by an event, more triggered by an anticipated due date. It is the clearest sign that the plant has shifted from firefighting to planning.
The lag between spotting a drift and acting on it. This is the one that measures the quality of the reconciliation, and it is the one that moves fastest once you stop reading reports in isolation.
All three indicators are built from the inspection history. This is the work an integrity tool should take on: pull together the existing reports, rebuild the history by equipment and by condition monitoring location, track the state of findings, and flag whatever is degrading faster than expected.
For the full approach, see risk-based maintenance, our Maintenance Intelligence page, and the page on preparing a turnaround for how this ties into the shutdown windows.
Can unplanned downtime be eliminated altogether?
No. Random failures exist and no monitoring announces them. The realistic goal is to make the third family disappear (the failures whose signal was already sitting in the documents) and to reduce the second through targeted monitoring.
Do we need to invest in sensors?
Not first. On most plants, using the measurements already taken during inspection campaigns delivers a gain before any hardware spend. Sensors earn their place afterwards, on the equipment where measurement frequency is the limiting factor.
How long before we see a result?
The drop in unresolved findings and the rise in anticipated jobs show up within a few months. The effect on the number of unplanned shutdowns takes several years to become statistically legible. Keeping the two horizons distinct avoids promises you cannot keep.
What do we do with the older reports?
Start with the last three years, which feed the current decisions. The older history can be brought in afterwards: it mainly serves to establish long trends, useful for remaining useful life but rarely for the arbitration of the quarter.
Written by Adama CamaraAI Consultant · Industry · view profile
Published on May 19, 2026
Software and data
Industrial AI: What It Really Is, and What It Changes on the Plant Floor
Industrial AI without the jargon: use cases by function and by sector, limits, costs, financing and a roadmap to get started in your plant.
Maintenance strategy
Maintenance planning: sorting the work when everything looks urgent
How to rank maintenance work when everything looks like a priority: objective criteria, a sorting method, and what you have to accept not doing.
Engineering and projects
AI in industrial production: where it really helps, where it over-promises
Where AI really helps in production: scheduling, process tuning, OEE, energy. Where it over-promises. The real prerequisite: data quality.
Maintenance strategy
Preventive maintenance plan: building it from real risk
How to build a preventive maintenance plan from equipment criticality and real condition, instead of inherited intervals that no one questions any more.