The alarm is not the first thing the machine tells you.
A threshold tells you a reading crossed a limit. It does not tell you how long it has been heading there, how fast, or whether the same shape has been seen on this pump class before. Most alarms arrive with none of that attached, which is part of why so many of them get acknowledged and left.
One reading says little. Two moving together say a great deal.
Vibration inside its limit is unremarkable. Vibration climbing while bearing temperature climbs, at steady load, is a mode developing.
Every individual reading on this pump stayed inside its alarm limit for the whole fortnight, so a threshold alarm would have said nothing. What matters is the shape: a steady departure from the band the machine normally holds, in two variables at once, while load stayed flat.
That pattern matches early bearing wear recorded on the same pump class elsewhere on site. Naming it now buys the P-F interval: the weeks between a fault you can detect and a failure that takes the machine off you. That is long enough to put a bearing swap on a shutdown you already have booked.
A recommendation, not another unexplained alert.
An alert that says a number moved leaves the engineer to do all the work. Muono names the asset, the failure mode it believes is developing, the evidence behind that belief and the window in which intervening is still cheap. Then it tells the engineer responsible before the next reporting cycle, not after it.
How a developing failure reads.
A feed pump on a two-week trend. Each line is something the model can point at, so the reliability engineer can argue with it instead of taking a score on trust.
- The signal
- Bearing vibration rising steadily, still inside its alarm limit on every individual reading.
- The correlation
- Rise tracks a climb in bearing temperature and holds through a period of steady load, so load does not explain it.
- The pattern
- Sequence matches the early stage of bearing wear recorded on the same pump class elsewhere on the site.
- The conclusion
- Probable bearing degradation rather than an instrument fault. The two readings would not move together if it were.
- The window
- Weeks rather than days, while the repair is still a planned swap instead of an unplanned outage.
- Who is told
- The reliability engineer who owns the pump, with the trend and the open work attached.
Your questions, answered
- Is this just threshold alarms with a nicer interface?
- No. A threshold fires when a single reading crosses a line. Prediction looks at how conditions develop over time and how variables behave together, so it can flag a machine whose every individual reading is still in range.
- How do we know it is not guessing?
- Because every recommendation shows the readings and the pattern behind it. If the evidence does not convince the engineer who owns the asset, it should not convince anyone. They can see it rather than take a score on trust.
- What happens to alerts nobody acts on?
- They stay on the asset's record with what was known at the time. When the same mode develops again the history is there, which is how the model and your team both get better at the call.
- Does it need years of failure history to start?
- It needs enough signal to establish how the asset normally behaves. Known failure modes for a class of equipment carry over, and the site-specific patterns build as it runs.