Soft sensors estimate a missing quality. Predictive maintenance watches a known failure mode. Time-series foundation models (TSFMs) aim wider: pretrain on diverse industrial and sensor corpora, then adapt to a plant’s historian so anomaly scores and short-horizon forecasts appear across thousands of tags without a custom model per asset. Vendors and integrators are wiring these engines to PI, Ignition, and cloud historians as shared services—not one-off notebooks.
The industrial point is transfer learning under messy OT data. A foundation model that ignores units, sampling rates, and bad-value flags becomes a confident false-alarm machine.
What TSFMs actually change
- Cold-start coverage — Useful baselines on assets that never got a dedicated PdM project.
- Cross-tag context — Joint patterns across related measurements, not single-channel thresholds.
- Forecast assist — Short horizons for energy, load, and buffer planning when physics models are thin.
What still goes wrong
Domain shift between industries, unlabeled regime changes, and alarm floods when scores lack suppress logic. Models that cannot explain which channels drove an alert fail change control. This is not closed-loop soft sensors, not industrial RAG over manuals, and not PLC copilots drafting logic.
What to watch next
- Which platforms ship plant-local fine-tunes that never leave the OT DMZ.
- Whether alerts carry channel attributions operators will trust on shift.
- Measured nuisance-alarm reduction versus classic SPC and rules packs.
Historians store the past. Foundation models try to read it at fleet scale—without inventing a new soft sensor for every CV.