Manufacturing
Production downtime logs: why AI teams value reason codes and fixes
By SourceX Editorial · Updated
Short answer
Downtime data is valuable to AI teams when each stop links an asset, a reason code, a duration and the fix that got the line running again. Reason codes show where time was lost; the fix and the technician's notes show how experienced people diagnose and repair. Logs tied to maintenance work orders rank highest.
Key takeaways
- An OEE dashboard summarizes losses; the event-level stop log with fixes is the record AI developers can learn from.
- Free-text repair notes linked to a stop event carry more signal than the reason code alone.
- Reason code lists change over time, so keep every version and a mapping between them.
- Operator and technician names, customer names and customer-owned tooling designs are removed or excluded before licensing.
What makes a downtime log more than an OEE report?#
A downtime log is the event-level record behind an OEE number: each stop, the asset, the time it started and ended, the reason selected and what was done about it. An OEE report rolls those events into availability, performance and quality scores, which helps manage a plant and teaches a model almost nothing.
The events live in a few places. MES and OEE systems capture machine state from PLC signals, operators pick reason codes on an HMI or tablet, and maintenance work orders sit in a CMMS such as Fiix, Limble, MaintainX, UpKeep, eMaint or IBM Maximo. Shift handover notes and andon calls fill in the rest.
For AI purposes the event log matters because it keeps sequence. A model learning to diagnose needs to see that a jam on the infeed followed a material change and preceded a sensor replacement; a weekly availability figure erases that order.
Which downtime fields matter, and what can AI learn from each?#
The downtime fields that matter most are the fix and the linked work order, because they record how a stop was actually resolved. Timestamps and asset IDs establish the event and reason codes classify it, but only the repair detail shows the expertise used to get the line running.
The linked work order is the field most plants lack. Where the MES and CMMS were bought separately, stops and repairs often share nothing but an asset name and a rough time, and matching them becomes the main preparation task. Plants that already require a work order number on any stop above a set length start well ahead.
| Downtime field | Example | AI task it can support |
|---|---|---|
| Asset and component | Press line hydraulic unit, infeed conveyor | Linking failure modes to equipment types |
| Reason code | Material starved, jam, tool change, electrical fault | Classifying stops from operator descriptions |
| Start, end and duration | Event timestamps from the PLC or MES | Predicting how long a stop will last |
| Operator comment | Sensor misread, cleaned lens, restarted | Learning how people describe symptoms in plain language |
| Fix and parts used | Replaced proximity sensor, reset encoder | Recommending the next diagnostic step |
| Linked work order | CMMS order with labor, cause and follow-up | Connecting production loss to maintenance action |
| Product and shift context | Product family running, shift, crew role | Spotting patterns tied to products or changeovers |
Why the fix matters more than the code#
The fix matters more than the reason code because codes are coarse and chosen under pressure. An operator with a stopped line picks the first plausible entry from a dropdown, and catch-all codes such as other or unknown tend to fill up quickly.
The repair narrative carries the expertise. A note saying the stop looked like a jam, the technician checked the photo eye, found it knocked out of alignment when a guard was removed, realigned it and added a check to the guard procedure is a full diagnostic chain: symptom, checks, cause, repair and prevention. Developers building maintenance assistants and troubleshooting agents need exactly that sequence, written by the people who do the work.
This is why a smaller log with consistent repair notes can outrank a much larger log of codes alone. Human-generated signal and domain expertise count for more than raw volume.
Common gaps, and how to describe them honestly#
Every downtime log has gaps, and describing them up front makes a package easier to scope and trust. Buyers expect imperfection; what slows a deal is a gap discovered late.
Each gap has a practical answer. Keep every version of the code list with a mapping table, record the threshold below which micro-stops are ignored, and match events to work orders on asset and time window where no shared key exists. Documented gaps become metadata; hidden gaps become disputes.
- Micro-stops below the MES threshold are not recorded at all.
- Reason code lists were revised during a reliability program, so old and new codes differ.
- Entries were backfilled at shift end with rounded times.
- Stop events and CMMS work orders share no common key beyond asset and time.
- Assets were renamed or renumbered after a retrofit or line move.
- Some lines were never connected to the MES and log stops on paper.
Illustrative: a stamping plant links stops to repairs#
Illustrative: a fictional metal stamping company runs several press lines for automotive and appliance customers. Its MES captures every stop from PLC signals, operators choose a reason on the press HMI, and maintenance logs work orders in a cloud CMMS. For years the two systems were reviewed separately.
Before a CMMS upgrade, the COO had the team export both histories and match stops to work orders on asset ID and time window. The reason code list had changed after a reliability program, so they kept both versions and a mapping table. Operator badge numbers were replaced with role labels, and customer names on work orders became codes.
The rights review excluded die drawings owned by customers, which had been attached to some work orders. What remained was a package of stop events, linked repairs and technician notes, with the code mapping and micro-stop threshold documented. The COO approved a prepared sample before anything left the plant.
What gets removed before downtime data is licensed?#
Downtime data is licensed only after people, customers and protected designs are taken out. Repair notes are written quickly and often include names, phone numbers and opinions about colleagues, so free text gets the closest review.
Safety records deserve their own decision. A stop caused by an injury belongs to a different process with different obligations: most US manufacturers also record work-related injuries on the OSHA 300 log and 301 incident reports, which name employees and must be kept for five years after the year they cover. The simplest approach is to leave injury-related stops out of the package entirely.
| Item | Typical treatment |
|---|---|
| Operator and technician names, badge IDs | Replaced with role labels or stable pseudonymous IDs |
| Customer names on work orders and job numbers | Replaced with codes |
| Customer-owned tooling and part drawings | Excluded |
| Free-text notes | Screened for names, contact details and personal remarks |
| Injury and safety incident details | Excluded or handled in a separate review |
| Plant addresses and network details in asset records | Generalized or removed |
How SourceX approaches downtime records#
SourceX handles downtime records within the SourceX five-step transaction of Supply, Rights, Preparation, Approval and Delivery, and nothing is shared during the initial assessment. Supply establishes from metadata which MES, OEE and CMMS tools hold the history, how many years survive and whether stops carry work order numbers. Rights checks customer-owned tooling and any equipment vendor terms that touch machine data. Preparation covers the matching, masking and free-text screening described above.
Through the SourceX Enterprise Data Value Framework, technician notes add human-generated signal and domain expertise, linked work orders add data cleanliness, and sensor-only streams from common machines are more reproducible, which reduces value. The SourceX Evidence Packet for a downtime package records provenance, including the code list versions and micro-stop threshold, alongside licensing rights, permitted use, the privacy record and release authorization.
Frequently asked questions
Is raw PLC or sensor data enough without reason codes?
Raw machine signals show that a stop happened and how long it lasted, but not why or how it was fixed. On their own they are closer to telemetry than to a decision record. Paired with reason codes, operator comments and linked work orders, the same signals become far more useful.
Our operators overuse the other code. Is the log still worth anything?
Often yes, if the comments or linked work orders explain what happened. Catch-all codes are common, and free text frequently carries the real cause. Describe the pattern openly in the inventory; a buyer can work with known noise more easily than with surprises found during preparation.
How is downtime data different from maintenance records?
Downtime data starts from the production side: when the line stopped and what it cost in output. Maintenance records start from the asset side: work orders, preventive tasks, parts and labor. The most useful packages connect the two so each stop links to the repair that ended it.
Can downtime logs reveal our plant's performance to outsiders?
They can if left raw, which is why preparation matters. Plant and customer identities are removed, asset names can be generalized, and the license limits permitted use and onward sharing. Your team checks a prepared extract for anything that points back to the plant before signing the release.
Do we need to clean up reason codes before a fit check?
No. The fit check uses metadata such as systems, years of history and whether stops link to work orders. Cleanup decisions come later and depend on the scope a buyer is interested in. Keeping old code lists and noting changes is more useful than rewriting history.
Sources
- 29 CFR 1904.33 requires employers to save the OSHA 300 Log, the privacy case list (if one exists), the annual summary, and the OSHA 301 Incident Report forms for five years following the end of the calendar year that the records cover. Source
Related resources
- QuestionDo AI labs buy code?
- InsightCan you license CAD and engineering drawings to AI companies?
- InsightCan you license code reviews and pull requests to AI companies?
- InsightCode snapshot vs full git history: what to include in a code license
- SolutionEnterprise data: the records of how organizations actually work
- SolutionProprietary data: information only your company has
See if your company qualifies
A short company assessment. No data uploads are needed.