A bearing fails over ninety days. Only one of these two plots notices.
Both come from the same 1490 rpm pump. The upper one is what the accelerometer measures, with the physical model drawn over it. The lower one is the difference between them. Drag the control and watch which of them changes.
- Measured RMS
- 0.460 —
- Residual kurtosis
- 2.9 —
The bearing is sound. Measurement and model agree; the residual is noise.
Fig. 1 — Kurtosis measures how impulsive a signal is rather than how large. A spalled race strikes sharply and periodically, so kurtosis climbs long before amplitude does. The values are computed from the plotted samples as you drag.
By day seventy the measured amplitude has moved about one percent — inside the band any usable threshold has to tolerate if it is not to cry wolf every shift. The residual's kurtosis has moved by several hundred percent. That gap is the whole proposition: the evidence was in the signal the entire time, buried under the machine's own healthy vibration, and subtracting a model that already knows what healthy looks like is what uncovers it.
Fig. 2 — Where it runs. Everything inside the dashed boundary sits on hardware bolted to the machine; the only thing that leaves the plant is the diagnosis. Ten milliseconds is not a marketing figure — it is the budget a control loop leaves you if the result is to be acted on rather than logged.
| Inference latency | < 10 ms | on-device, no network in the control loop |
|---|---|---|
| Model basis | Physics + residual | known structure is not relearned from scratch |
| Deployment | NPU / MCU | at the machine, operates network-down |
| Data egress | None | process data never leaves the plant |
Bring me a machine that costs you money when it stops.
Four weeks, one machine, a fixed fee, starting from the sensors already fitted. You keep the hardware and the model whether or not we carry on.
info@noether.se