VesselSeed

Archived VerifyRun experiment

Does this bearing alarm warrant an inspection?

If you’re responsible for keeping a plant running, an alarm after a pump changes speed leaves you with a hard call. Take it offline and you might stop a healthy machine. Wait and a worn bearing could fail on shift.

We built VesselSeed with VerifyRun to test a question for reliability engineers. Given a vibration recording and shaft speed, could the timing of repeated impacts spare an unnecessary inspection? We parked the project after testing it on recordings from two labs.

During motor startup, VesselSeed flagged every healthy recording. The usual speed-adjusted check caught all six damaged recordings and flagged all four healthy ones. Adding vibration timing changed zero decisions. On this evidence, keep your existing inspection process.

Two motor startups6206 bearing · 200 W
Measured shaft speed rises through both two-second windows, roughly from 670 to 1,280 RPM in the healthy run and 710 to 1,310 RPM in the damaged run.
Same bearing size and load. The shaft sped up in both runs. All three methods called for inspection on both recordings. Only one had inner-ring damage.

Measured shaft speed from Thuan and Hong’s Hanoi data, adapted under CC BY 4.0.

Could timing rule out an unnecessary inspection during startup?

We took two seconds from each motor startup with a measured rise in shaft speed. The conventional check followed the vibration frequencies expected from a damaged inner ring as the shaft accelerated. VesselSeed added the timing of those vibrations. We ran a third check with the same model and the timing switched off.

Hanoi motor startups · one call per recording
MethodDamaged recordings caughtHealthy recordings flagged
Speed-adjusted baseline6 / 64 / 4
VesselSeed, timing on6 / 64 / 4
Same model, timing off6 / 64 / 4

Zero false alarms avoided. Every method flagged every recording with the original alarm settings. The calls also stayed the same when we compared the methods at equal fault detection. With six damaged recordings, reaching the 90% target required catching all six.

Which recordings counted, and how was the test frozen?

We checked all 30 healthy or isolated inner-ring files in the Hanoi archive. Ten contained both startup vibration and a corresponding RPM trace. The other 20 lacked the pair required to track speed. Each method received the same two-second window selected from RPM alone. Matching healthy and damaged recordings by bearing size and load left four of each. All eight still triggered every alarm.

The rules, code, fitted weights and decision thresholds were saved before we scored startup vibration. Training used Saarland data. We had previously studied steady-speed signals from the same Hanoi rig. The ten recordings share equipment and include repeated loads. This tests real laboratory acceleration. Performance on customer machinery with inspection-confirmed faults remains unknown.

Freeze revision af918d25. Original recordings · Authors’ description.

Compare the steady-speed results · Read the project decision

Why we tested bearing alarms

What made the idea practical?

Schaeffler and its service partner describe one manufacturer going from 50 sensors in one plant to more than 5,000 across several European plants by September 2023. Maintenance staff confirmed bearing damage on an exhaust motor after an alert. Existing monitoring systems made it possible to propose a second review of past alarms using recordings plants already collect. Deployment case study.

Who might have paid for a second opinion?

Grand View Research estimates the global vibration monitoring market at $1.9 billion in 2026. That includes equipment and services. We considered plant reliability teams and monitoring firms that review variable-speed pump and motor alarms and can check what inspection eventually found. Market estimate and scope.

Our original U.S. base case was $6.21 million a year for reviews at manufacturing and wastewater plants. How many sites can supply the data, how often they need a review and what they will pay per case are assumptions.

What did we compare with existing tools?

SKF’s order tracking lines up vibration with shaft rotation so a change of speed does not smear a fault across frequencies. Its enveloping tools help analysts find bearing impacts. Those are established ways to investigate an alarm. SKF order tracking · How SKF detects repetitive impacts.

Two recordings can both have energy at the expected defect frequency. Our question was whether the rhythm of those impacts can distinguish damage after accounting for speed. Across our first experiment and the independent dataset, adding phase information saved no false alarms at the same fault-detection rate. We would keep the existing workflow on this evidence.

How large is the U.S. market for a second opinion?

Start with two U.S. customer groups that operate pumps and motors. The 2022 Economic Census counted 286,626 manufacturing establishments. The EPA counted 17,544 publicly owned wastewater treatment works as of January 2022. We count sites once, including those served by monitoring contractors. A contractor serving ten eligible plants represents ten sites in this calculation. U.S. Census Bureau · EPA survey.

The base case assumes 3% of manufacturing sites and 10% of wastewater sites have relevant variable-speed equipment, exportable waveforms and speed records, and someone who can obtain inspection outcomes. That gives about 10,353 eligible sites. Assume four difficult cases per site per year at $150 each. The result is $6.21 million annually, about $5.16 million from manufacturing and $1.05 million from wastewater.

Annual U.S. serviceable spend if each eligible site bought the assumed reviews
AssumptionLowBaseHigh
Eligible manufacturing sites1%3%7%
Eligible wastewater sites3%10%20%
Reviews per eligible site / year248
Price per review$75$150$250
Annual spend$0.51M$6.21M$47.15M

The base case equals about 0.33% of the $1.9 billion global category as a scale comparison. The calculation uses the two U.S. site counts, with no extrapolation to other countries or industries. Halving either the case volume or the fee halves the result. Ten paying sites at the base price and volume would produce $6,000 a year. As a side project, eight analyst hours a week for 48 weeks at two hours per review would cap delivery at 192 cases or $28,800 a year before expenses, customer acquisition or idle time. These are capacity scenarios. We parked the project before validating the site filters, price or purchase frequency with buyers. Download the calculation inputs.

On a second rig, the full grid added no correct calls.

We froze a bearing-specific implementation of the published grid method, calibrated it on 96 Saarland recordings and tested it on an independent dataset from Hanoi University of Science and Technology. The test used all 30 steady-speed healthy and isolated inner-race recordings, covering five bearing sizes and three motor loads. Each file counted once. HUST data, version 3 · Experiment and bearing dimensions.

Retrospective comparison at 14 of 15 damaged recordings detected
MethodFaults detectedHealthy recordings flagged
Speed-adjusted baseline14 / 153 / 15
Baseline plus full grid14 / 153 / 15
Same model, measured phase removed14 / 153 / 15
Full grid alone14 / 1514 / 15

All three baseline-based methods flagged the same three healthy 6205 recordings and missed the same 6207 fault at 200 W. Used alone, the grid flagged eleven more healthy recordings. Load settings repeat across bearings, so 30 recordings do not mean 30 independent machines.

What would have happened with the original alarm settings?

The table above uses test labels to match detection rates. When we kept the alarm thresholds set on Saarland, the baseline, combined model and phase-removed control each flagged all 30 Hanoi recordings. They caught all 15 faults and also alarmed on all 15 healthy recordings. The grid alone caught 14 faults and flagged all 15 healthy recordings. A plant could not have obtained the table’s lower false-alarm rate simply by installing those settings.

What “full grid” means in this test

Both methods received the same two seconds of one acceleration channel, shaft speed and bearing geometry. The baseline measured envelope energy near three speed-adjusted inner-ring harmonics. Our implementation mapped the waveform to 15 prime-frequency anchors on a 7 × 7 spiral, populated neighboring cells, iteratively updated phase and coherence, applied chirality and harmonic weights, selected peaks and converted them to an alarm score. All 30 grids reached the frozen convergence criterion. One produced no eligible prime peak.

The combined model used the baseline and grid score together. For the control, we zeroed measured phase offsets before building the grid and kept the combined model’s fitted weights. The standalone arm used only the grid score. The application describes multiple embodiments. The frozen protocol identifies the specific input mapping, update rule, readout and limits used here. The test concerns steady speeds and laboratory damage. The startup test examines the accelerating part of this lab archive.

Our decision is to keep the existing speed-adjusted workflow. A new field trial would need a specific class of failures and blinded inspection outcomes that could justify revisiting this result. Read the earlier phase-only study.

Switching phase on saved zero false alarms in the first test.

Damage on a bearing’s inner ring can produce a pulse each time a roller crosses it. A conventional check looks for vibration energy around the expected repetition rate as speed changes. We asked whether the timing of those vibrations across short intervals could help an analyst dismiss healthy readings.

Held-out recordings, compared at 44 of 48 damaged recordings detected
AnalysisFaults detectedHealthy recordings flagged
Speed-adjusted baseline44 / 4847 / 48
Baseline with phase alignment44 / 4846 / 48
Same fitted model, phase switched off44 / 4846 / 48

Each method missed four damaged recordings. Nearly every healthy recording triggered an alarm. The one-alarm improvement over the standalone baseline persisted with phase switched off. On this evidence, we would keep the analyst’s existing workflow.

Study design, thresholds and limits

We used 96 two-second recordings from three physical NU206 bearings, before and after artificial inner-ring damage. The subset covers four speed settings, two load settings, one bearing position and two runs. For each test recording, training excluded both that bearing and that speed setting. The dataset samples different speed settings. Field inspections during real operating changes remain an open question.

Each method received the same three-axis acceleration at 20 kHz and the measured shaft RPM. The baseline uses envelope power near the first three harmonics of the expected inner-ring defect frequency. The second model adds a phase-alignment score over eight intervals. For the phase-off row, we kept that model’s weights fixed and replaced the score with its training mean.

We matched detection rates retrospectively using held-out scores. With thresholds set using training data alone, the baseline and phase-on models each detected 36 of 48 damaged recordings and flagged 36 and 34 healthy recordings respectively. With phase off, the model detected 27 and flagged 36. With repeated recordings from just three bearings, this secondary result needs an independent test.

Dataset and rig by Christopher Schnur and colleagues at Saarland University. Source recordings · Experiment design

See how speed changes the signal.

In this synthetic signal, increasing shaft speed brings the pulses closer together. Keep speed fixed and increase pulse jitter to see what irregular timing does. Turn the defect impulse down to zero to see shaft vibration and noise alone.

Synthetic vibration

0 ms125 ms250 ms

Illustration only. The simulated defect repeats 7.76 times per revolution. The measured results above come from the physical bearings.

What did we take from the invention?

Devin Bostick’s U.S. application Data Processing via Prime Coherence describes a method for finding patterns in a waveform. It stores frequency and phase information in a grid, updates each entry using its neighbors, and selects peaks using chirality, phase and harmonic relationships. The USPTO issued a notice of allowance for the amended claims on September 9, 2026.

The first bearing experiment borrowed the phase-alignment score in equations 4 and 5. The Hanoi test exercised the core waveform-to-grid-to-peaks method in a bearing-specific implementation. In motor startup and in both steady-speed comparisons, adding measured timing saved zero false alarms at matched detection.

Why we parked VesselSeed

We wanted reliability engineers to be able to leave more healthy machines running while catching just as many faults. Across the two lab datasets, switching timing off left the false-alarm count unchanged at matched fault detection. We found no reason to offer a bearing-inspection service.

We’ve kept the measured results, the code and the synthetic simulator public so you can inspect the decision. The original rationale explains why we thought the experiment was worth running.