TwinEdge / Platform · Spotting troubleWorking out what is wrong
Spotting trouble early
Find the change that deserves attention.
Compare signals with their operating context to detect developing patterns that individual alarm limits can miss.
Why five methods and not one
| Agreement | What happens |
|---|---|
| One method flags it | Logged. Nobody is woken up. |
| Several agree | Raised for review, with what each one saw. |
| They agree and physics agrees | A job is drafted with the readings attached. |
§ 01
The real problem is crying wolf.
Detecting an anomaly is not difficult. Detecting one that a maintenance team will thank you for is very difficult, and the difference is almost entirely about how often you are wrong.
An operator who is woken at three in the morning twice for nothing will mute the third alert, and the third one is the bearing. Every industrial alerting system dies this way.
So agreement matters more than sensitivity. One method noticing something is information. Several methods and the physics agreeing is a reason to send somebody.
“It is not missing the fault. It is being wrong often enough that nobody reads it any more.”
The failure mode of every alerting system
§ 02
The five ways it watches.
They have different blind spots on purpose. What one misses, another usually catches.
- 01Odd combinations
Vibration is fine. Temperature is fine. Flow is fine. All three together are a combination this machine has never produced before. That is the one a fixed limit will never catch.Under 50 ms
- 02A learned picture of normal
Builds a picture of how this particular machine behaves, then measures how far today sits from it. Good at slow, subtle change.Under 80 ms
- 03Statistical drift
The long-established charts for spotting a gradual shift in a process. Unfashionable, cheap, and very hard to argue with.Under 5 ms
- 04Behaviour unlike anything seen
Draws a boundary around everything the machine has done while healthy and reports when it steps outside it.Under 60 ms
- 05Rhythm that changed
Looks at the frequency content rather than the value, which is where bearing and gear faults show up first.Under 90 ms
§ 03
What happens to a finding.
A finding that stops at a notification has not helped anybody. Here it carries what was seen, which sensors, how far from normal, what the physics says about it, and how sure the system is.
If it clears the bar you set, a job is drafted for a planner. If it does not, it is kept and counted, so the pattern is visible next month even though nobody was disturbed today.
Test it against a failure you already had.
Pick a machine that failed and still has its history on file. We run the readings forward from before the failure, and you see two things: how early it would have spoken up, and how many times it would have bothered you in the months when nothing was wrong.