Industrial AI is attractive because it promises to find patterns that human teams might miss. But predictive models cannot recover facts that were never recorded, distinguish two machines with the same name, or validate whether a manually written figure is correct. The first AI investment in many factories should be in trustworthy operational data, not a model subscription.
A factory does not need perfect automation to become data-ready. It needs consistent identities, defined measurements, sufficient event history, documented uncertainty and a clear operational question.
1. Define the decision before choosing the model
Different questions need different evidence.
Scroll horizontally if needed to read all columns.
| Operational question | Minimum useful evidence | Missing-data risk |
|---|---|---|
| Why did output fall last week? | Production by asset, time window, product, shift, quality outcomes, relevant events | Incorrect comparisons or false causes |
| Which assets fail repeatedly? | Stable asset IDs, incident dates, failure classifications, maintenance history | Repeat incidents may be counted as separate failures |
| Can we forecast material demand? | Historical demand, lead times, inventory movements, seasonality, exceptions | Stockouts and changes in product mix are hidden |
| Which quality conditions precede rejects? | Traceable quality results, batch and process context, times, equipment state where available | Spurious correlation or label leakage |
| Can we predict a specific failure? | Enough actual failure examples, relevant sensor or inspection history, maintenance interventions | Model learns maintenance policy rather than physical risk |
If no clear action would follow a model’s prediction, the project is probably premature. Define who would act, how quickly, what information they need and the cost of false alarms or missed events.
2. Six foundational data groups
Asset and process identity
Maintain authoritative site, area, machine, component, product and production-order identifiers. Record when an asset was renamed, moved or replaced. A row marked “Machine 2” is insufficient when several departments use that label.
Production and scheduling context
Capture the shift definition, product run, planned schedule, good output, rejected output and explicit unit of measurement. Without context, comparing 5,000 pieces of one product to 5,000 kilograms of another is meaningless.
Time-based incidents and changes
Record event timestamps, time zone, changes in machine availability, set-up events, material shortages and observed symptoms. Distinguish when an event happened from when a person entered the record.
Quality results
Retain inspection outcomes, specifications, units, sampling method, batch and relevant corrective action. A pass/fail label without the inspection rule or revision can mislead a model trained on historical data.
Maintenance and interventions
Record asset, work type, actual action, parts replaced, completion, and whether the record describes a failure, inspection or scheduled service. A completed job card is evidence of work, not necessarily evidence of machine unavailability.
Source provenance and revisions
For each record, preserve where it came from, how it was imported, validation status, corrections and revision history. This supports audits and prevents silently changing the training data after a model has been evaluated.
3. The difference between a data gap and an unknown answer
Consider a plant that reports 100 hours of scheduled time and 15 hours of maintenance jobs. It may be tempting to assume availability was 85%. That is not supported: jobs can be performed while machines are running, and unrelated downtime might not have a job card. A trustworthy system should show “insufficient downtime evidence” rather than fabricate a confident KPI.
The same discipline applies to AI. Missing values may be informative, random, systematically omitted or caused by a change in recording procedure. Treating them all as zeros can produce false trends.
4. Measure data quality using questions you can test
For an initial assessment, quantify:
- Completeness: what proportion of required fields is populated?
- Validity: do values satisfy units, ranges, formats and business rules?
- Identity resolution: can records be reliably linked to assets and events?
- Timeliness: what is the delay between occurrence and capture?
- Consistency: do different systems use compatible definitions?
- Traceability: can a displayed number be traced to original evidence?
- Continuity: does enough relevant history survive changes in people and software?
A high percentage of filled-in cells does not guarantee accuracy. Sample records against the originals and document the resulting error rate.
5. Establish a baseline without AI
Start with descriptive analysis that answers real questions: output by line and shift, repeated quality issues, record completion rate, machine-specific incident history and unresolved data gaps. Then ask whether simple thresholds, rules or statistical comparisons solve the problem adequately.
A model is justified only when a more complex prediction materially outperforms a simpler, reliable method and can be safely deployed. Compare against a baseline using an appropriate time-separated evaluation dataset. Avoid training on information that would not have been available at the moment a real decision must be made.
6. Pilot AI with an explicit risk contract
For a proposed model, define intended use, excluded decisions, data provenance, human responsibilities, evaluation design, monitoring, fallback behavior and criteria for stopping use. A failure-alerting model may be advisory; it should not silently alter machine settings. Controls affecting physical processes require separate safety analysis, authorization and controls engineering.
NIST’s AI Risk Management Framework emphasizes managing AI risks in the context of organizational decisions, while NIST’s OT guidance recognizes the special safety, reliability and performance needs of industrial systems.
Factory AI-readiness checklist
- We can name one operational decision AI could improve.
- Each relevant event has a dependable time and asset identity.
- Measurement units and product/shift definitions are explicit.
- Original records and correction histories remain available.
- Missing and uncertain values are distinguishable from zeros.
- There are enough representative examples to evaluate the proposed method.
- A simple non-AI baseline has been measured.
- We know the consequences of false positives and false negatives.
- Users know when not to trust the proposed output.
- Deployment and fallback work in the plant’s actual connectivity conditions.
Frequently asked questions
Do we need years of data? Not universally. The required volume depends on the outcome, variability, sampling rate, failure frequency and evaluation method. Rare failures may need more evidence than is available.
Can historical paper data help? Yes, when source quality, identity and time context are sufficiently recoverable. Uncertain linkages must remain labeled.
Is a real-time dashboard the same as AI readiness? No. Displaying values is useful but does not establish data quality or predictive validity.
Should a small factory start with predictive maintenance? Only if the required failure signals and intervention economics exist. Better incident records or simple rules may create more immediate value.
Related reading
Begin with Digitizing Paper Records, then review Industrial Performance Intelligence and Job Cards, Breakdowns and Downtime.
