Key takeaways
- Synchronize clocks across PLCs, historians, MES and edge devices, and record which timestamp each value carries.
- Historian compression and deadbands are tuned for trending, not for machine learning. Know your settings before training a model.
- Sensor drift and calibration gaps look like process changes. Link calibration records to the data.
- Without batch, product and shift context, most production data cannot be used for AI.
- Every data domain needs a steward who owns meaning, quality rules and fixes. Quality without ownership decays.
Ask a data scientist what went wrong with an industrial AI pilot and you will often hear the same story. The model looked good in the notebook. It fell apart on the line. When someone finally dug in, the cause was a clock that was four minutes out, a pressure transmitter that had drifted for a year, a historian deadband that flattened the signal, or a dataset that did not record which product was running.
These are not exotic problems. They are the normal state of plant data, because plant data was collected to run and trend the process, not to train models. This page covers the quality issues that matter most for industrial analytics and AI, how to spot them, and who should own fixing them.
Does this apply to you?
If you plan to use historian, MES, lab or camera data for any model, dashboard or AI tool, yes. The NIST AI Risk Management Framework treats data quality and documentation as part of managing AI risk, and in a plant that risk shows up as wrong recommendations on real equipment.
The big six industrial data quality problems
| Problem | What it looks like | Typical cause | Owner |
|---|---|---|---|
| Timestamp misalignment | Effects appear before causes | Unsynchronized clocks, mixed time zones, source vs receipt time | OT infrastructure |
| Sensor drift and calibration | Slow trend in a "stable" value; step after calibration | Wear, fouling, no calibration link | Instrument / maintenance |
| Compression and deadband | Flat lines, missing small changes, sparse data | Historian settings tuned for trending | OT engineer / historian admin |
| Missing context | Cannot tell product, batch, shift or state | Context lives in MES, paper or heads | Process / production owner |
| Bad or missing values | Gaps, frozen values, out-of-range readings, bad quality flags ignored | Comms faults, sensor failure, substitution | OT engineer |
| Inconsistent labels | Same defect labeled differently by shift | No defect catalog, no agreement checks | Quality |
1. Timestamp alignment
Plant data comes from many clocks: PLCs, HMIs, historian servers, MES, lab systems, edge gateways and cameras. If they disagree, correlations break.
What to do:
- Synchronize clocks to a common time source, typically via NTP (or PTP where you need sub-millisecond precision), within your zone design.
- Store timestamps in UTC and convert only for display. Daylight saving changes create duplicate or missing hours otherwise.
- Record which timestamp a value carries: source (device), server (historian), or receipt (cloud). OPC UA, for example, distinguishes source and server timestamps. Keep both where available.
- Note known latencies, such as lab results that arrive hours after sampling, and align on sample time, not entry time.
2. Sensor drift and calibration
A drifting sensor produces data that looks like a real process change. A model will happily learn it.
What to do:
- Link calibration records (date, as-found, as-left) to tags, so analysts can see when a sensor was verified.
- Use redundant measurements, mass or energy balances, or lab results to cross-check critical sensors.
- Flag data from sensors that are overdue for calibration.
- Treat recalibration as an event in the dataset, because the step change it causes can confuse models.
3. Historian compression and deadbands
Historians save storage by recording only significant changes. Exception deadbands discard values that do not move enough; compression algorithms such as swinging-door keep only points that deviate from a predicted line.
These settings are tuned for operators looking at trends. For AI they can:
- Remove the small, early changes that precede faults.
- Shift the apparent timing of events.
- Make interpolated values look like measured values.
What to do:
- Export compression and exception settings per tag before training anything.
- For high-value tags in a model, consider tighter settings or a separate uncompressed stream at the edge.
- Mark interpolated values clearly in any dataset.
Watch out: Changing historian compression settings changes what you record from that point on. It is a configuration change to an OT system. Review and schedule it like one, and document the date so analysts know the data changed character.
4. Missing context
A temperature trend without knowing which product, batch, recipe, shift or machine state it belongs to is hard to use. Most production AI needs:
- Product or SKU and recipe.
- Batch, lot or order number.
- Machine state (running, idle, changeover, fault, planned stop).
- Shift and crew.
- Quality outcome, linked back to the production run.
What to do: capture context at the source or the edge, using a consistent ISA-95 based asset model, and publish it alongside process data. Reconstructing it later by joining on timestamps is fragile.
5. Bad, missing and substituted values
- Respect quality flags from OPC UA servers and historians. "Bad" or "uncertain" values should not be silently treated as good.
- Detect frozen values (identical readings for longer than physically plausible).
- Apply range checks based on engineering limits.
- Document how gaps are handled: left as gaps, interpolated, or forward-filled.
6. Labeling for vision QC
Vision-based quality inspection lives or dies on labels.
- Build a defect catalog with images, definitions and borderline examples, owned by quality.
- Measure agreement: have a sample labeled by two inspectors and resolve disagreements against the catalog.
- Record context with each image: product, line, camera, lighting setup, date.
- Version the catalog and dataset together. When the definition of a defect changes, the old labels may be wrong.
- Keep the customer or regulatory definition of a defect in view. A model that matches your inspectors is only useful if your inspectors match the specification.
Plant-floor reality: Disagreement between day and night shift on what counts as a defect is common. A vision model will learn whichever shift labeled more images. Fix the definition first.
Ownership and stewardship
Quality rules without owners decay within months. A simple model that fits a lean team:
| Role | Responsibility | Typical person |
|---|---|---|
| Data owner | Accountable for meaning and fitness for use in a domain (e.g. Line 3 process data, vision defects) | Area or process owner, quality manager |
| Data steward | Maintains definitions, quality rules, tag mapping, defect catalog; triages issues | OT engineer, quality engineer |
| Platform custodian | Runs historian, broker, data platform; backups, access, retention | OT infrastructure or IT |
| Data consumer | Uses data for analytics or AI; reports issues | Data scientist, vendor, analyst |
Fold these into your AI roles and RACI so that every AI use case knows whom to call when data looks wrong.
Quality as part of AI risk
For each AI use case in the AI use case register, record:
- Which data domains it depends on and who owns them.
- Known quality issues and how they are handled.
- What happens to the model's output when inputs are bad or missing.
Then include data quality in the AI risk assessment. A model that degrades gracefully on bad data is lower risk than one that confidently outputs nonsense.
Common mistakes
- Treating data quality as a one-time cleansing project.
- Fixing data in the data platform but not at the source, so the problem returns.
- Ignoring historian quality flags.
- Training on compressed data without checking settings.
- Letting the AI vendor define your defect catalog.
- Assigning stewardship to IT alone, without process knowledge.
Monday-morning checklist
- Pick one AI use case and list every tag and data source it uses.
- Check clock sync for those sources and confirm timestamps are in UTC.
- Export historian compression and deadband settings for the tags.
- Check calibration dates for the critical sensors.
- Confirm product, batch, state and shift context are available and aligned.
- Name a data owner and steward for the domain.
- Log findings in the industrial data readiness scorecard.
Effort: a quality review for one use case is typically a few days of OT and quality engineering time. Establishing stewardship roles is mostly a management decision and a standing agenda item.
What good looks like
Clocks agree. Timestamps are UTC with known sources. Compression settings are documented and suitable for each use. Calibration is linked to tags. Context is captured at the source. Vision labels follow a versioned catalog with measured agreement. And every domain has a named owner who fixes problems at the source. For the bigger picture, run the industrial data readiness checklist, and see how predictive maintenance depends on these basics.
Sources and further reading
- ISA-95 Standard (ISA)
- NIST AI Risk Management Framework
- OPC UA Online Reference: published specifications, OPC 10000 / IEC 62541 (OPC Foundation)
- NIST SP 800-82 Rev. 3, Guide to Operational Technology (OT) Security
This guide is general information for manufacturers, not legal advice. Confirm regulatory obligations with qualified counsel.
FAQ
Frequently asked questions
Why does historian compression matter for AI?
Historians often store a new value only when it moves outside a deadband or deviates from a predicted line, such as swinging-door compression. That keeps storage small and trends readable, but it removes small variations and changes the apparent timing of events. Models trained on compressed data may miss early signs of faults. Check settings per tag and use uncompressed data where the use case needs it.
How do I detect sensor drift?
Compare sensors against redundant measurements, lab results or calibration checks; watch for slow trends in a value that should be stable during steady state; and link calibration records so you can see when a sensor was last verified. A step change right after calibration often reveals how much drift had built up.
Who should own industrial data quality?
Ownership should sit with the people who understand the process, not only IT. A common model is a data owner (process or area owner) accountable for meaning and fitness for use, a data steward (often an OT or quality engineer) who maintains definitions and quality rules, and a platform custodian (IT or OT infrastructure) who runs the systems.
How should we label images for vision quality inspection?
Write a defect catalog with examples and borderline cases, train labelers on it, and have a sample labeled by two people to measure agreement. Version the catalog and the labeled dataset together. Re-label when the defect definition, product or camera setup changes. Disagreement between inspectors is a quality signal, not just noise.
Related guides
Data Foundations
Moving OT data to a cloud historian: architecture, security and governance
Move plant data to the cloud through a one-way, outbound-only path from a DMZ, add context before it leaves the site, and treat every new data flow as a governed conduit with a named owner.
Guide · 7 min read
Data Foundations
Predictive maintenance programs: data, validation and governance
Predictive maintenance works when you have trustworthy failure history, the right sensors on the right assets, a validated model, and a named person who acts on each alert.
Guide · 7 min read
Data Foundations
Industrial data readiness for AI: a 40-point checklist with scoring
Before you fund an industrial AI use case, score your data against 40 checks in eight dimensions.
Guide · 6 min read
AI Governance
AI governance roles and RACI for mid-market manufacturers
Mid-market manufacturers need a small AI governance group drawn from roles they already have, a clear RACI across the AI lifecycle, and a firm split between what plant managers own (use…
Guide · 7 min read