ai · time series
Forecasting + typed decisions
Zero-shot foundation models forecast the sensors; a small decision model judges how serious each alarm is.
An experiment on open data: can an industrial process warn you before things go wrong, without training a model on it first?
- Forecast every sensor a few hours ahead with a zero-shot time-series foundation model. No training on the plant’s data.
- Raise alarms with plain, auditable rules applied to the forecasts, not just to the measurements.
- Judge the severity of each alarm with a small typed decision model that reads the context and answers with a level.
The data comes from a published physics model of the process, with scripted incidents. On the replay day, a critical incident was flagged 79 minutes before it happened.
