Time-series models assume timestamps mean “when the world happened.” On plants, historians, PLCs, edge gateways, and cloud feature stores often disagree by seconds to minutes—sometimes more after a network event. Join a soft sensor feature to a quality label on the wrong clock and you invent lead-lag that never existed. Compression traps own false features from storage. Tag aliases own wrong signals under familiar names. This brief owns clock integrity as an AI prerequisite.
Where skew is born
| Source | Typical failure | | --- | --- | | PLC free-running clocks | Drift until someone notices batch genealogy looks “early” | | Historian vs MES time zones | DST and UTC mix in joins | | Edge buffers stamped on upload | Event time replaced by ingest time | | NTP outage / wrong stratum | Silent skew across islands |
Models do not apologize. They fit the lie.
If two systems disagree on “now,” causality is optional.
Time distribution is plant infrastructure—not an IT courtesy.
Controls that keep sequence honest
- Declare a plant time authority (NTP/PTP) and monitor offset alarms like process alarms.
- Store event time and ingest time separately in feature pipelines.
- Reject training joins when source clock offsets exceed a frozen threshold.
- On model drift, ask “did time break?” before “retrain harder.”
- Document DST and timezone rules in the same MoC as tag dictionaries.
An anonymized packaging line that “proved” a causal AI story about oven setpoints vs scrap was actually joining MES scrap timestamps stamped at upload after a store-and-forward outage. The oven looked guilty. The clock was.
Adjacent fences
Historian compression ML owns storage artifacts. Tag alias collisions own identity lies. Override telemetry owns human force labels. SOP RAG deltas own document version skew. This page owns time skew as causal poison. Do not celebrate a foundation model on OT data and leave NTP as an unchecked checkbox.