Flagged as it drifts. Not after it breaks.
Every machine has its own normal. Misamtek learns that normal, then flags any reading that leaves it - the moment the drift begins, not when the alarm limit is finally crossed.
Normal is learned, per machine
A single hard limit is a blunt instrument. A machine can operate perfectly at one temperature under one load and drift dangerously at another. The engine learns a duty-cycle-aware envelope for each channel on each machine, taking into account time of day, load state, ambient conditions and cycle position.
Once the envelope is established, drift outside it is flagged the moment it appears, along with the magnitude and shape of the deviation.
Not just amplitude - shape
Two anomalies with the same amplitude can mean two entirely different failure modes. The engine scores amplitude, spectral shift, symmetry, harmonic content and time-of-cycle position together. That composite signature is what informs the diagnostic suggestion downstream.
Worked example
A 55 kW induction motor driving a centrifugal pump. Baseline vibration on the drive-end bearing sits at 2.1 mm/s RMS across a typical duty cycle. Over the course of a week the average drifts to 2.6 mm/s while staying well below the 4.5 mm/s alarm limit. Meanwhile a peak appears at 1x running speed in the FFT that was not present before.
A conventional hard-limit alarm would not fire. The engine flags this at day two, calls out the specific spectral peak as consistent with early-stage bearing race spalling, and recommends inspection within a specific window. Maintenance replaces the bearing on a planned stop rather than during a failure.
False positives
Envelope learning includes explicit handling for regime shifts - a scheduled load change, a shift start, a duty-cycle transition - so a new steady state is not repeatedly flagged as an anomaly. The engine also weights persistence: a single one-cycle outlier is treated differently from a sustained drift.
Pick an equipment type. Watch its sensor pulse move.
This is not a mockup. The waveforms below are being generated the same way we score real streams - learned envelope, drift detection, diagnosed callout. Pick a machine and watch an anomaly emerge.