TwinEdge / Platform · Building modelsAcross every site
Building models
Build models around the equipment.
Use prepared operating data to train, compare, and manage models. Connect model performance to the assets and conditions they monitor.
How candidates are compared
| Measure | Why it is there |
|---|---|
| How often it is right | The obvious one, and the least interesting on its own |
| How often it cries wolf | The number that decides whether anyone keeps reading it |
| Cost to run | It has to run on a small box at a site, continuously |
| What it used | Ranked inputs, so an engineer can sanity-check it |
§ 01
What it does for you.
- 01Work out what to feed it
Derives the quantities that usually matter — vibration levels, temperature differences, frequency peaks, power ratios — from your readings, and discards the ones that carry nothing.
- 02Try several approaches, not one
Different methods suit different failures. Several are trained on the same data so the comparison is fair rather than a defence of whichever one was tried first.
- 03Compare on more than accuracy
How often it is right, how often it cries wolf, and how much it costs to run on a small box at a site. A model that is 1% better and forty times slower is not better.
- 04Show what it paid attention to
Which inputs drove the answer, ranked. This is what lets an engineer say “that is reasonable” or “it has learned that our night shift runs differently, which is not a fault”.
- 05Ship it carefully
The chosen model is compressed for the edge and rolled out to a few sites alongside the current one before it replaces anything.
§ 02
What it deliberately will not decide.
It will not decide whether a model is sensible. It can tell you that one is 96% accurate on historical data; it cannot tell you that the 4% it gets wrong are all the failures that actually matter to you.
Nor will it decide what counts as a failure. If your history records a failure on the day the work order was raised rather than the day the machine went, everything trained on it inherits that, and only somebody who knows the plant will notice.
So the ranked inputs are not decoration. They are the point at which a person who knows the equipment can catch a model that has learned the wrong thing.
“The model is excellent at predicting when somebody typed something into the maintenance system.”
The most common way these go wrong
Build one on your own failures.
Bring history for one family of machines with a few recorded failures.1 We build candidates on it, show you what each one learned, and let your engineers decide whether any of them deserve to run.
- 1How much history is enough depends on how often the machines fail and how well those failures were recorded. A family of similar machines with two years of readings and a handful of genuine failures is usually workable; a single machine instrumented last month is not.