Unified namespaces clean names. Time-series foundation models score anomalies across thousands of tags. Soft sensors estimate missing qualities. OT cyber keeps strangers out. What still wrecks all of them is quiet: frozen transmitters, flatlined values, unit mismatches, and “good” tags that stopped representing the process months ago.
Historian data-quality programs—bad-actor lists, stale/flat detection, and tag ownership—are returning as standing reliability work, not a one-time cleanup before an AI pilot.
The industrial point is trusted engineering units in time. A beautiful dashboard on poisoned tags is a faster way to make bad decisions.
When the model is fine and the tag is lying
A familiar failure mode: PdM or TSFM alerts spike; operations learns the transmitter impulse line is plugged or the tag was repurposed after a turnaround and never retargeted. Mature sites publish bad-actor rates (for example, percent of critical tags flat or out-of-range over a rolling window) and assign owners the same way they assign pump reliability.
Many “AI failures” start as instrumentation and configuration failures.
An anonymized polymers plant froze new analytics rollouts until a weekly data-quality board cleared critical tags: flatline detectors, stale heartbeat checks, and a rule that any tag used in a soft sensor must have an owner and a last-calibration reference. False anomaly tickets fell before the models were even retrained—because the inputs stopped lying.
What a tag DQ program actually owns
- Detection rules — Flat, stale, spike, and impossible engineering ranges.
- Ownership — Who fixes the field device versus the tag configuration?
- Analytics admission — Tags must pass DQ gates before they feed models or KPIs.
DQ is a meeting with disposition—not only a silent scoring job.
Mistakes that still waste AI budget
Buying another analytics layer on an ungoverned historian. Treating UNS rename projects as a substitute for fixing dead instruments. Expecting TSFMs to “learn around” chronic bad actors without a quarantine list.
This is not OT network zoning, not ISA-18.2 alarm floods, not ePTW isolation, and not MOC drawing control. Those protect other truths.
Signals before the next analytics CapEx
- Bad-actor trend on the critical-tag set—down and owned, not only detected.
- Whether soft sensors and TSFMs refuse inputs that fail DQ—not just warn.
- Calibration and configuration workflows that close the loop from DQ ticket to field fix.
Buyer checklist (short)
- Critical-tag list — Exists, sized, and reviewed on a cadence?
- Quarantine — Can a bad tag be blocked from consumers quickly?
- Owner map — Instrument vs historian vs application owner clear?
- Pilot rule — No new model on tags below a DQ threshold?
Namespaces organize meaning. Data quality decides whether that meaning is still true at 02:00 on a Sunday.