Key takeaways
- Assess data readiness for a specific use case, not for the plant as a whole.
- Context such as asset hierarchy, units and tag descriptions saves more time than any modeling technique.
- Check completeness, time alignment, resolution, historian compression and label consistency before modeling.
- Weak outcome labels are the most common gap; decide explicitly whether to fix capture, hand-label or use methods that need fewer labels.
Why data readiness decides most projects
Models are rarely the bottleneck in manufacturing AI. The slow, expensive part is getting data that is complete, correctly time-stamped, labeled with what actually happened, and connected to the equipment, product and process it describes. A plant can have years of historian data and still not be ready, because nothing links a temperature trend to the lot it affected or the failure that followed.
Data readiness is not a yes-or-no state for the whole plant. It is specific to a use case. The data needed to predict bearing failures on a set of pumps is different from the data needed to predict tensile strength from furnace conditions. Assess readiness against the project you actually want to run.
Know where your data lives
Most plants have the same handful of sources, each with its own strengths and gaps.
| Source | What it usually holds | Common gaps |
|---|---|---|
| PLCs and controllers | Live process values, states, alarms, counters | Little history; tag names are cryptic; values may be scaled or in raw counts |
| SCADA and HMI | Supervisory view, alarm logs, operator actions | Alarm floods; operator actions not always logged with context |
| Process historian | Time-series history of tags | Compression settings that discard detail; missing metadata; no link to product |
| MES | Work orders, lots, genealogy, downtime reasons, production counts | Downtime reasons entered inconsistently or defaulted |
| Quality systems and LIMS | Inspection results, lab tests, nonconformances | Results recorded hours after production; sampling is sparse |
| CMMS or EAM | Work orders, failure codes, parts used | Failure codes left blank or set to “other”; work logged after the fact |
| ERP | Orders, materials, costs, suppliers | Coarse time resolution; costs allocated rather than measured |
Map which sources matter for your use case, who owns each one, and how you would get data out of it. That map is the first deliverable of any AI project.
Tags, naming and context
A historian tag called FIC_2041.PV means something to the controls engineer who configured it and nothing to anyone else. AI projects need context: which asset the tag belongs to, its engineering units, its normal range, whether it is a measurement or a setpoint, and what process step it sits in. Without that, data scientists spend weeks reverse-engineering the plant.
Practical steps that pay off across many projects:
- Adopt an asset hierarchy, typically following ISA-95 levels such as enterprise, site, area, line and work cell, and attach tags to it.
- Record units, ranges and descriptions in a place other systems can read, not in a spreadsheet on one engineer's laptop.
- Separate setpoints, process values and controller outputs clearly so a model does not learn from the wrong one.
- Note when instruments were replaced, recalibrated or re-ranged, since those events change the data.
Data quality checks to run
Before modeling, pull a sample covering at least a few months and run basic checks. These take days, not weeks, and they surface most of the problems that would otherwise appear halfway through a pilot.
- Completeness. What share of expected readings is missing? Are gaps random or tied to shutdowns, network outages or shift changes?
- Time alignment. Do clocks agree across systems? A quality result stamped when the lab entered it, not when the part was made, will mislead any model that tries to link cause and effect.
- Resolution. Is the sampling rate fast enough for the physics? Vibration analysis may need high-frequency waveforms; a historian storing one value per minute will not capture them.
- Compression. Historian deadband and compression settings can flatten the small changes that matter for early fault detection. Check them for the tags you care about.
- Flatlines and stuck values. A sensor that reports the same number for days is often failed, not stable.
- Labels. Are failure codes, defect codes and downtime reasons used consistently? Read a few hundred records by hand. You will learn more than from any summary.
Labels and outcomes are usually the real gap
Supervised models learn from examples of inputs paired with outcomes. In manufacturing, the inputs are often plentiful and the outcomes are thin. Maintenance logs say “replaced motor” without saying why it failed. Scrap is recorded by shift, not by lot. Inspection images are saved but never tagged with the defect type.
You have three options when labels are weak. You can improve how outcomes are captured going forward, which takes time but builds lasting value. You can label historical data by hand with subject-matter experts, which is slow but often worth it for a focused pilot. Or you can choose methods that need fewer labels, such as anomaly detection that learns normal behavior and flags departures. Each is legitimate; just be honest about which one you are choosing.
Access, architecture and ownership
Even clean data is useless if you cannot get to it safely. Confirm early how data will leave the OT network, whether through a historian replica in a demilitarized zone, an edge gateway publishing to a broker, or scheduled exports. Involve cybersecurity before anyone connects a laptop to a control network. Decide who owns the integrated dataset, who can change tag mappings, and how changes in the plant, such as a new line or a replaced sensor, will be reflected in the data. A project that depends on one engineer manually exporting files every week is not ready to scale, even if the pilot works.
A readiness scorecard
Rate each item as ready, partly ready or not ready for the specific use case. Anything not ready on the first three lines should be fixed before a pilot starts.
| Item | Question |
|---|---|
| Outcome data | Is the thing we want to predict or detect recorded reliably, with timestamps and asset or lot IDs? |
| Input data | Are the relevant process signals captured at a useful rate and resolution? |
| Linkage | Can inputs be joined to outcomes through a common key or reliable time alignment? |
| Context | Are tags mapped to assets, units and process steps? |
| History | Is there enough representative history, including enough examples of the events of interest? |
| Access | Is there an approved, secure path for data to reach the analytics environment? |
| Ownership | Is someone accountable for keeping the data correct as the plant changes? |
Related use cases
Operations & production
Machine monitoring & OEE analytics
Automatically capture machine states, downtime reasons and output to measure OEE accurately and target the biggest losses.
Maintenance & reliability
Predictive maintenance
Use sensor and maintenance history to spot developing equipment failures early and schedule repairs before they stop the line.
Quality
Quality prediction & virtual metrology
Predict product quality from live process data so problems are caught before parts are measured, tested or shipped.
Operations & production
Process anomaly detection
Models learn normal process behavior across many sensors and flag unusual patterns before they become quality, safety or equipment problems.