Research · Systems & Control
Industrial Autonomy & Control
Learned models inside a control loop, where the consequence of being wrong is physical.
Why this is hard
A control room is the hardest place to put a model. It cannot pause, it cannot ask the operator to retry, and every action has to be defensible afterwards. Our work is on what learning can safely contribute here: seeing what a fixed alarm limit cannot, inferring a measurement that arrives too late to be useful, and doing both without taking authority away from the person on shift.
Questions we are working on
- What can a multivariate model detect that no single-tag alarm threshold ever will?
- How far ahead can a soft sensor infer a laboratory result, and how should it recalibrate when the real one lands?
- How should a model surface itself to an operator mid-upset without adding to the noise?
- Where is the boundary between advisory and actuating, and what evidence justifies moving it?
What comes out of it
- Multivariate anomaly detection over process telemetry, learning normal per unit and per duty
- Soft sensors for measurements that otherwise arrive hours late from a laboratory
- Remaining useful life estimation with intervals wide enough to be honest and narrow enough to plan against
- Operator-facing presentation designed for the moment of an upset, not for a report
Where it ships
How it connects
Industrial autonomy in the practice.
This field and the others in its group. None of them is pursued in isolation.
Industrial Autonomy & Control
Learned models inside a control loop, where the consequence of being wrong is physical.
Data Systems at Scale
Keeping a number correct while the system underneath it is concurrent, partitioned and always moving.
Read the fieldBring us a problem in industrial autonomy.
Joint research, a proof of concept against your own data, or an honest read on whether the thing you want is possible yet. An engineer who works in this field replies.