Bearing diagnostics experiment
A second opinion on a bearing alarm.
A motor speeds up and its vibration alarm goes off. Does the bearing need an inspection, or did the operating conditions change? VesselSeed is exploring whether an analyst can answer that question using a recording they already have and the shaft speed at the time.
Our first test asked whether the timing of repeated vibrations adds anything to a conventional speed-adjusted analysis. It flagged 46 healthy recordings with phase on and 46 with phase off. That result gives us no reason to recommend switching tools.
Measured on real bearings
Switching phase on made no difference to the false-alarm count.
We compared 96 recordings from three bearings. Each method caught 44 of 48 damaged recordings (91.7%). The bars count how many of the 48 healthy recordings also triggered an alarm.
On this rig, at this detection rate, every method flagged almost every healthy recording. The one-record difference from the standalone baseline persisted with phase switched off.
What was measured?
96 recordings from three physical NU206 bearings, before and after artificial inner-ring damage. Each method received the same first two seconds of three-axis acceleration sampled at 20 kHz and the recorded shaft RPM. The subset used four speed settings, the lowest and highest load settings, one mounting position and two runs. A physical bearing contributed repeated observations.
For each test recording, model fitting excluded its entire bearing and its target speed. The baseline measures envelope power around the first three harmonics of the expected inner-race frequency. A second fitted model adds the published multi-harmonic RIC phase-alignment score (equations 4 and 5), applied to phase estimates across eight intervals. The ablation sets that feature to its training mean with the fitted weights held fixed.
The matched-sensitivity comparison selected thresholds retrospectively using out-of-fold scores. In a separate check with thresholds set using training data alone, the baseline detected 36/48 with 36 false alarms, phase on detected 36/48 with 34 false alarms, and phase off detected 27/48 with 36 false alarms. The matched experiment tests one implementation of a published component. Industrial performance and a full RIC pipeline remain untested.
Download per-recording scores · Reproduce the comparison
Data from Christopher Schnur and colleagues, Saarland University, used under CC BY 4.0. The scores are derived from excerpts. Recordings · Rig and fault frequency · Published phase equations
Synthetic vibration
The simulator uses a seeded noise signal and a damped pulse at 7.76 crossings per shaft revolution. Turn the impulse to zero to see shaft vibration and noise alone. The controls affect only the illustration.
The RIC research behind the test
Resonance Intelligence Core (RIC) describes a way to measure how consistently the phases of a signal line up. For this experiment we used equations 4 and 5 from Data Processing via Prime Coherence, published U.S. patent application US 2026/0220223 A1. We measured alignment across short intervals around the frequencies where an inner-ring defect should repeat as the shaft turns.
VesselSeed is the name of this experiment. The benchmark evaluates that RIC component on bearing recordings. Read the published application and equations.
Work on the next test
If you review vibration alarms or build monitoring tools, I’d like to hear about a case where a change in speed or load made an alarm hard to interpret. What did the inspection eventually show? An anonymized case description is enough to start.
I’m especially interested in a stronger conventional baseline or a blind comparison on historical alarms with known outcomes. Email Devin at devin.bostick@gmail.com.