The most honest record of what goes wrong in a steel plant is usually written in a hurry, such as changed bearing, ran ok. Multiply that by years of work orders, handover notes and inspection comments, and the plant holds a large body of knowledge that no dashboard can read. Predictive analytics needs that history, but free text does not fit a model until someone gives it structure. This guide covers how to convert it and capture better records going forward. You can set up structured work order capture in Oxmaint to improve the source.
AI and Data · Steel Plant Maintenance Records
Steel Plant Unstructured Maintenance Data Management for Predictive Analytics
Convert free-text work orders, handover notes and inspection comments into structured failure history, so predictive analytics learns from what your technicians actually saw.
One technician note, read as data
"Stand 4 gearbox noisy again, topped up oil, bearing on input shaft hot, ran ok after. Check next stop."
AssetStand 4 gearbox
SymptomNoise, high bearing temperature
ComponentInput shaft bearing
ActionOil top-up, no repair
RepeatYes, "again"
Follow-upInspect at next stop
What Counts as Unstructured Maintenance Data
Structured fields such as asset ID, date, cost and status are easy to count. Everything else carries the reasoning.
Free-text work order notes
Problem descriptions, work performed and recommendations, often abbreviated and plant-specific.
Shift handover logs
Verbal observations written down: unusual sounds, temporary fixes and things to watch.
Inspection comments
Round sheet remarks and photos that note condition without a coded severity.
Incident and RCA reports
Documents with causes and corrective actions, filed as PDFs or presentations.
OEM manuals and service reports
Vendor findings and recommended intervals held outside the maintenance system.
Photos and attachments
Images of wear, leaks and damage with no tag describing what they show.
Why Predictive Models Struggle With Raw Maintenance Records
01
Inconsistent vocabulary
One failure is called a seized bearing, a locked roller or a dead idler depending on who wrote the note.
02
Symptoms recorded as causes
"Motor tripped" describes an effect. The model needs the failure mode behind it.
03
Missing links to assets
Notes mention equipment by nickname, so history cannot be joined to sensor data for the same machine.
04
Failures recorded without timing
If the failure date is the closing date, the model learns the wrong lead time between warning and fault.
05
Mixed languages and shorthand
Multi-site plants combine languages, local abbreviations and typing errors in one dataset.
The Conversion Pipeline: From Text to Training Data
A
Collect
Export work orders, notes, inspection records and reports with asset, date and author attached.
B
Clean
Remove duplicates and templates, fix encoding, and expand known abbreviations.
C
Extract
Pull asset, component, symptom, cause and action from the text with rules, language models or both.
D
Standardize
Map extracted terms to a controlled failure vocabulary and the asset hierarchy.
E
Validate
Have senior technicians review a sample, and measure how often the extraction agrees with them.
F
Publish
Write structured fields back or alongside, keeping the original text as evidence.
Fix the Source Before You Clean the Archive
Every free-text note you clean today is one you will have to clean again if the next one is written the same way. Add structured fields to the work order and the problem shrinks.
Designing a Failure Vocabulary Technicians Will Actually Use
A controlled list works only if it is short enough to choose from at a mill-side tablet. Use a hierarchy that narrows in a few taps.
| Level | Question it answers | Illustrative values |
|---|---|---|
| Equipment class | What kind of machine was it? | Gearbox, motor, pump, conveyor idler, hydraulic cylinder, crane hoist |
| Component | Which part failed or was affected? | Input bearing, seal, coupling, filter, contactor |
| Symptom | What was observed? | Noise, vibration, leak, overheating, trip, loss of pressure |
| Failure mode | How did it fail? | Wear, fatigue, contamination, misalignment, blockage, electrical fault |
| Likely cause | Why did it fail? | Lubrication, overload, installation, ageing, operation, unknown |
| Action taken | What did the technician do? | Replace, repair, adjust, clean, lubricate, inspect only, defer |
Capture Better Going Forward: Before and After
Typical closing entry
Structured closing entry
"Fixed pump, working now."
Asset: descaling pump 2. Component: mechanical seal. Symptom: leak. Failure mode: wear. Action: replaced. Parts: seal kit. Photo of the removed seal attached.
"Checked motor, ok."
Inspection result: no fault found. Measured value entered. Condition rating selected. Next check set.
"Temporary repair, redo at shutdown."
Action: temporary repair. Follow-up work order created and linked to the next planned outage.
Making the Technician's Job Easier, Not Longer
- Offer picklists for component, symptom and action, with a free-text box for extra detail.
- Allow voice-to-text and photos on mobile, so detail is captured at the machine.
- Pre-fill asset, location and recent history when a job is opened.
- Keep mandatory fields to the few that models truly need, and explain why they matter.
- Show technicians how their entries were used, such as a repeat failure spotted from their notes.
Data Quality Levels for Analytics Readiness
Level 1
Recorded
Work orders exist, but most meaning is in free text.
Level 2
Linked
Every record points to a valid asset and carries a true failure date.
Level 3
Classified
Failure mode and action are coded with a controlled vocabulary.
Level 4
Connected
History joins to sensor and production data on the same asset and timeline.
What Cleaned Text Makes Possible
Repeat failure detection
Find assets with the same symptom recorded again and again, even when wording differs.
Failure-mode analysis
Rank what actually fails on each equipment class to focus PM and spares.
Labeled history for models
Give predictive models real failure examples with dates, modes and the condition data before them.
Overlooked recommendations
Surface follow-up advice written in a note that never became a work order.
Faster troubleshooting
Let technicians search similar past faults and see what fixed them.
Better job plans
Update tasks, durations and parts lists from what crews actually did.
Risks to Manage When Using Language Models on Maintenance Text
Wrong extraction A model may label a note incorrectly. Sample and review results, and keep the original text beside every extracted field.
Hidden bias in history If a failure was never recorded, the model cannot learn it. Check which assets and shifts are under-reported.
Confidential content Reports may include names, contracts or safety details. Set access rules before text leaves the maintenance system.
One-time clean-up Without better capture, quality decays again. Treat vocabulary and field design as ongoing maintenance of the data.
How Oxmaint Supports Better Maintenance Data
Structured work orders Fields for problem, cause and action alongside notes, so records are searchable and countable.
Asset management Every record ties to a place in the hierarchy, which makes joining to other data possible.
Mobile inspections Checklists with readings, ratings and photos instead of loose comments.
Reports Repeat failures, backlog and completion quality visible to reliability and planning.
Extraction Methods Compared
There is no single method that fits every plant. Many teams start simple and add complexity only where results justify it.
| Method | How it works | Strength | Limit |
|---|---|---|---|
| Keyword and rule matching | Searches text for known words and patterns | Transparent and quick to start | Misses new wording and misspellings |
| Dictionary with fuzzy matching | Maps variants and typing errors to a standard term | Handles local shorthand once it is catalogued | Needs a maintained vocabulary list |
| Trained classifiers | Learns categories from labeled examples | Good at assigning failure mode or symptom | Needs enough labeled records per category |
| Language models | Reads a note and fills structured fields from instructions | Copes with varied phrasing and longer notes | Outputs must be checked, and sensitive text needs access controls |
| Human review | Senior technicians correct a sample | Gives trusted labels and builds confidence | Time-consuming, so use it where it matters most |
Steel Plant Vocabulary to Standardize
Rolling mill
Roll, chock, stand, gearbox, spindle, coupling, bridle. The same part may be named differently by mill and shift.
Continuous casting
Segment, mould, oscillator, tundish, ladle turret, roll bearing. Include both operator and maintenance names.
Blast furnace and stoves
Hot blast valve, tuyere, cooling plate, blower, stove burner. Capture the location as well as the part.
Utilities
Compressor, cooling pump, water treatment, fan, oxygen and gas lines. Note the service and the system.
Material handling
Conveyor belt, idler, pulley, skirt, chute, crusher. Position along the belt matters for history.
Cranes and lifting
Hoist, brake, rope, limit switch, runway. Distinguish the crane from the component.
Measuring Extraction Quality
| Measure | Meaning | How to use it |
|---|---|---|
| Agreement with reviewers | How often the extracted field matches what a senior technician would assign | Set a minimum before results feed any model or report |
| Coverage | Share of records where a field could be filled at all | Low coverage points to notes that are too thin to use |
| Unknown rate | Share assigned to an unknown or other category | A high rate shows the vocabulary needs new terms |
| Drift over time | Changes in agreement as wording and crews change | Recheck a sample every quarter |
Photos, PDFs and Other Attachments
- Attach each file to the work order and asset it belongs to, with the date of capture.
- Add a short tag saying what it shows, such as wear pattern, leak point or nameplate.
- Store RCA reports and vendor findings against the failed asset, not in a general folder.
- Extract the failure mode, cause and corrective action from reports into structured fields.
- Keep the original file as evidence and avoid overwriting it with the extracted summary.
A Review Loop That Keeps Quality From Decaying
Weekly Planners scan new work orders for missing or doubtful classifications and correct them while the job is fresh.
Monthly Reliability engineers review the unknown category and add new terms to the controlled vocabulary.
Quarterly Senior technicians check a random sample, measure agreement and decide whether rules need changes.
Yearly The team reviews the vocabulary against new equipment, changed processes and recurring failures.
Governing Text Data Safely
- Decide which fields may include names, contracts or safety details, and mask them before analysis.
- Limit access to raw incident reports to those who need them, and log who reads them.
- Record which version of the vocabulary and extraction method produced each structured field.
- Keep corrections made by reviewers, because they are the best training examples you will ever have.
Common Vague Entries and the Question Behind Each
What the note says
What a reviewer should ask
"Replaced as per usual."
Which part, how many, and why was it replaced? Was it a failure, wear-out or a scheduled change?
"Motor tripped, reset."
What caused the trip, and did it repeat? Was the load, temperature or supply voltage checked before the reset?
"Checked and cleaned."
What was found before cleaning, and what condition was recorded afterward? Was a reading or photo taken?
"Same problem as last week."
Which earlier job does this refer to? Should the earlier fix be linked and the cause reopened for review?
"Operator reported noise."
Where was the noise, when did it start and what was the running condition? Did the technician confirm it?
A First-Month Plan for Cleaner Maintenance Data
- Week one: choose the critical assets and list the ten most common failure modes on each equipment class.
- Week two: build the short pick lists for component, symptom, failure mode and action, with the people who do the work.
- Week three: add the pick lists to work order closing, brief the crews and fix the first complaints about speed.
- Week four: review a sample of closed jobs, correct misclassified ones and agree the next set of fields to add.
- Afterward: run the extraction on older records for the same critical assets, and compare results with the new coded entries.
Questions to Ask Before Choosing an Analytics Approach
- Which decision will the analysis support, and who will act on it?
- Do the failure events we care about occur often enough in our history to be learned from?
- Can the extracted fields be traced back to the original note and the person who wrote it?
- How will senior technicians correct mistakes, and will corrections feed back into the method?
- Where does the text go during processing, and who can see it?
- What will we do with the answer if the data turns out to be too thin?
Who Owns What in Maintenance Text Data
Technician
Writes the record
Chooses codes at closing and adds a clear note about what was found and done.
Planner
Checks quality
Reviews new closed jobs, corrects obvious gaps and links repeat issues.
Reliability
Owns the vocabulary
Maintains failure codes, reviews unknown entries and checks extraction results.
Data steward
Keeps the pipeline honest
Tracks coverage and agreement, and reports problems to the right owner.
Frequently Asked Questions
How much history do we need before predictive analytics is useful?
It depends on how often each failure mode occurs. Rare failures need several years or related assets, while frequent ones show patterns sooner.
Should we clean the whole archive first?
No. Start with critical assets and the most frequent failure modes, and fix capture at the source in parallel. Book a demo to plan a first scope.
Can photos and PDFs be used?
Yes, once they are tied to an asset and a date and tagged with what they show. Untagged files are hard to use in analysis.
Do we still need free text?
Yes. Structured fields hold the facts, while notes hold the nuance. The goal is both, linked to the same record.
What is the first change that improves data quality?
Make closing a work order require a failure mode and action from short lists. You can sign up and configure this in minutes.
Turn Years of Technician Notes Into Evidence
Capture structured failure history from today and prepare older records for analysis, so predictive maintenance is trained on your plant's own experience.
