Part 6 of 7 in our Edge to Enterprise in practice series. In Sense, Decide, Act we shared an honest result: a value that froze inside its normal range looked normal to our first model. Retrained on the machine's own data, the CATIMi Advisor now catches it; this post is about how it learned.
A frozen value is a quiet fault
A sensor that stops changing can be the most misleading reading on a line. A stuck transmitter reports a perfectly reasonable number - inside its range, no alarm - while the process it is meant to measure drifts away. A threshold alarm cannot see it, because nothing crossed a threshold.
Our first model learned where the machine's value normally sat. A value that froze in the middle of that band looked, to the model, like a good day.
What retraining taught it
The answer was in the machine's own behaviour. A healthy reading is rarely perfectly still: it wanders a little, in a pattern of its own. We retrained the Advisor on recordings of the machine's normal running so that it learned that pattern as well as the level. To the retrained model, a value that suddenly stops wandering is as unusual as one that jumps.
On a simulated machine, not a customer plant, the retrained model raised its alarm on the frozen value, both on a recorded run and live, and still caught the faults the first model caught: refused replies, and a jump to an unusual value.
The lesson is broader than one fault. Most of the work in an anomaly model is not the model; it is the record of what normal looks like. So that is where we put the tooling.
Normal is gathered, not guessed
- On the controller, on a schedule. Recording runs on the controller itself: a one-off period that stops by itself, or a recurring window across shifts, nights and weekends, with nobody at the machine. A scheduled recording never cuts into one already running.
- Labelled at the source. Each recording is marked when it is made: good, for normal running the model learns from, or bad, for a known fault the model must prove it catches.
- Only real readings. Seconds when the controller had no good reading are left out of training, so a lost link costs recording time and nothing else.
- Sealed. Each finished recording is sealed, and one that has changed or gone missing is never trained on.
Good trains, bad proves
The labels carry through to training. Good recordings teach the model what normal looks like. Bad recordings are held back as the test: the model must raise its alarm on the faults they contain, and that is measured, not assumed. A recording labelled as a fault can never be used to teach normal; the trainer stops rather than learn a fault as normal.
Each result is reported in plain words: how much normal running the model learned from, how often it would have raised a false alarm on running it had never seen, which faults it caught, and how closely the compact version that runs on the AI accelerator follows the full model.
A new model goes on only when it is healthy
Putting a new model on the controller is a checked step. It waits while a recording is running. Once the model is in place, the Advisor must name it and pass its check on the AI accelerator within a set time; if it does not, the model it was replacing stays. The old model is removed only after the new one is healthy.
While the controller has no good data to score, the Advisor publishes no score at all, rather than a confident score of nothing. And the rule from Sense, Decide, Act holds throughout: the model advises; the controller decides.
What it means for you
- For executives: an AI result you can audit - what it learned from, what it was tested against, and what it caught.
- For project managers: gathering normal running is a schedule, not a site visit, and the effort goes into your plant's own data, where the value is.
- For engineers: labels at the source, sealed recordings, faults held out for the test, and a model that goes on only when it is healthy.
Your machine, your normal
If you have a machine whose quiet faults cost you more than its loud ones, we would like to hear about it. The CATIMi Advisor is now in field trials on CATIM Nimbus. Get in touch.
About CATIM. CATIM - Creating Automation Technology In Motion - empowers people to engineer automation solutions that solve hard technical problems. CATIM Nimbus, the CODESYS edge controller with advisory AI, is the next generation of the Nimbus platform it has built since 2019.