Industries / Industrial equipment and services / Industrial machinery / Bearings
Monitoring in use
Monitoring in use means watching bearings inside running machines, mostly through vibration, sound, temperature and motor current, so that a failing bearing can be replaced before it seizes.
6
Milestones achieved, from 1969 to 2023.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1969
Listening for shock pulses
Before instruments, machine condition was judged by putting an ear to a wooden rod or screwdriver held on the machine. SPM Instrument of Sweden says the Danish shipowner A.P. Møller found the cargo pumps of its tankers breaking down too often without warning, that Eivind Søhoel patented the shock pulse method in 1969, and that the company, formed in 1970, had sold 40,000 of its portable 43A meters by 1980. The sensor is tuned to resonate at about 32 kHz, so it responds to the weak shocks of a rolling element striking a flaw and filters out slower machine vibration. A 2024 study in the journal Sensors lists the method among the commercial bearing diagnostics in use, beside Emerson's PeakVue and Rockwell Automation's spike energy.
Credit
Eivind Søhoel; SPM Instrument AB
2007
The filter band chosen automatically
Envelope analysis is the usual way to find a damaged bearing in a vibration signal. Each time a ball strikes a flaw, the structure rings at a high frequency, so the signal is filtered to a band around that ringing and its outline, the envelope, is traced; the outline repeats at a rate that points to the damaged part, such as the outer or the inner ring. Choosing the band took skill. Jérôme Antoni defined spectral kurtosis in 2006, a measure of how spiky a signal is within each frequency band, and in 2007 published a fast way to compute it over many bands at once, the kurtogram, which points to the spikiest band. MathWorks added a kurtogram function to its signal processing software in release R2018a, and its bearing diagnosis example uses it to make an outer ring fault visible in the envelope.
Credit
Jérôme Antoni; Robert B. Randall, his co-author on the bearing diagnostics tutorial MathWorks cites
2012
A contest on worn-out bearings
For the IEEE PHM 2012 data challenge, the FEMTO-ST Institute in Besançon, France, ran ball bearings on its PRONOSTIA rig under heavy radial load until vibration passed 20 g (twenty times the acceleration of gravity), so the damage grew on its own and was not cut in by hand. Teams got 6 complete runs and had to predict the remaining life of 11 more from cut-off records; lives ranged from about 1 to 7 hours. The organizers wrote that the textbook fault frequencies did not work, because damage reached balls, rings and cage at the same time, and that L10 life ratings did not match the lives measured. A.L.D. Ltd. of Israel and the University of Maryland's CALCE center won.
LessonMethods tuned on single, clean flaws can fail on bearings that wear out on their own.
Credit
Patrick Nectoux, Rafael Gouriveau, Kamal Medjaher, Noureddine Zerhouni and colleagues, FEMTO-ST Institute; the IEEE Reliability Society
2013
A wind bearing, recorded failing
Eric Bechhoefer, Brandon Van Hecke and David He published vibration from the high-speed shaft bearing of a 2 MW wind turbine, 6 seconds a day for 50 days, during which an inner race fault developed and the bearing failed in service. MathWorks uses the record in its worked example of predicting a bearing's remaining life.
Credit
Eric Bechhoefer, Brandon Van Hecke and David He (PHM Society conference, 2013)
2016
Cut flaws mislead the models
Most research tested methods on flaws cut into bearings on purpose, by spark erosion or drilling. Christian Lessmeier and colleagues at Paderborn University ran 108 bearings of type 6203 in an accelerated life test with heavy load and thin oil, found 33 damaged spots on 18 of them, and published vibration and motor current from 32 bearings: 6 healthy, 12 with cut flaws and 14 with damage grown in the life test. Classifiers trained on the cut flaws named the grown damage correctly at best 75% of the time from vibration and 60.9% from motor current; trained on grown damage, they reached 98.5% and 93.3%. The authors concluded that training data with artificial damage "is apparently not sufficient".
LessonA model is only as good as the damage it learned from.
Credit
Christian Lessmeier, James Kuria Kimotho, Detmar Zimmer and Walter Sextro, Paderborn University
2023
Eleven pulp-mill failures
Adam Lundström and Mattias O'Nils of Mid Sweden University published vibration recorded at a pulp mill from 2019 to the end of 2022: 11 cases, each with about four months of measurements up to a fault, ten of them bearing faults before the bearing was changed. They wrote that almost all open bearing data came from laboratories, and found that the mill's faults differed widely in how they developed, in shaft speed and in vibration level. In 2026 Stamatis Apeiranthitis and colleagues compared 41 laboratory runs from five public data sets with 10 runs from industrial and ship machinery, and found the laboratory runs steadier in operating conditions and smoother in how they degraded. In 2024 Lundström, O'Nils and Faisal Qureshi trained a model on seven public laboratory data sets plus healthy records from the mill, without any of the mill's faults, and it classified the mill's measurements correctly 77.47% of the time on average.
data.mendeley.comapi.semanticscholar.orgapi.semanticscholar.orgdoi.org
Credit
Adam Lundström and Mattias O'Nils, Mid Sweden University and SCA; Stamatis Apeiranthitis and colleagues
Watch for
Sensors inside the bearing
Sentea, a Belgian maker of fibre-optic sensing, says SKF chose it to co-develop and produce the SKF Core, one unit for sensing, processing and communication in the next generation of SKF Insight bearings, which report load, speed, temperature and vibration from inside the bearing. Sentea says its chip-based interrogator, the part that reads the optical fibres, is ready for serial production, and SKF planned to show live data at the AMB trade fair in Stuttgart on 15 to 19 September 2026. In 2020 SKF called a bearing with built-in fibre-optic load sensing the world's first commercial one, and no field results were reported then either.
Achieved whena customer or laboratory outside SKF publishes load or condition data from these bearings in service, checked against separate sensors or against what was found when the bearings were opened.
Claimed by
Sentea and SKF, September 2026
One model for all vibration
Giuseppe Mannone, Paula Fischer and Martin Dazer of the University of Stuttgart's Institute of Machine Components trained VibFM, a transformer model, on about 400 hours of unlabeled vibration from 16 open data sets, so that it can be adapted to a new machine with few labeled examples. On Paderborn's bearing data, which it had not seen, the adapted model named the damage correctly 85.5% of the time and its fixed features with a small add-on 75.1%, against 42% to 54% for the same network trained from scratch; the authors note that the original Paderborn paper reported up to 98.3% with classical features.
Achieved whensomeone outside the team tests it on bearings that failed in service, such as the pulp-mill cases, and it beats envelope analysis and classical features there.
Claimed by
the authors, July 2026
Open
01
Lab-trained models, plant-scored
A public scorecard that runs bearing fault detectors trained only on laboratory data over every public record of a bearing that failed in a working machine, and scores each on how early it warned and how many false alarms it raised. Start with the 11 pulp-mill cases and the 2 MW wind turbine bearing, then add cases donated by plants, each with the bearing's damage confirmed when it was opened.
The briefCredit · unclaimed02
Open bearing frequency table
A free, machine-readable table that gives, for common bearing designations from the main makers, the internal geometry (number of balls or rollers, their diameter, the pitch diameter and the contact angle) and the four defect frequencies per shaft revolution that vibration analysts look for: outer ring, inner ring, rolling element and cage. Each value carries its source.
The briefCredit · unclaimed
Open 01 · the brief
Lab-trained models, plant-scored
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public scorecard that runs bearing fault detectors trained only on laboratory data over every public record of a bearing that failed in a working machine, and scores each on how early it warned and how many false alarms it raised. Start with the 11 pulp-mill cases and the 2 MW wind turbine bearing, then add cases donated by plants, each with the bearing's damage confirmed when it was opened.
Why it is still open
Field records are few. A curated list of 93 public bearing and machinery data sets, updated in September 2026, marks only five as recorded in working machines, among them the 11 pulp-mill cases and the single wind turbine bearing; most of the rest have flaws cut in or damage grown on a rig. The pulp-mill data's own authors scored one model trained on laboratory data on the 11 mill cases in 2024, by the share of measurements classified correctly, not by how early each fault was flagged; the 2026 comparison of laboratory and operational runs scored no detector; and the VibFM model was tested for transfer only on Paderborn's laboratory data. No shared scorecard was found that runs many detectors over every public field case.
Built with what exists
The BearingDatasets Python library, which loads Paderborn, XJTU-SY, FEMTO and other sets in one format; envelope analysis with a kurtogram as the baseline; PyTorch for trained detectors, with small models fine-tuned on one graphics chip (see the Note "One GPU to fine-tune"); phone photos of each opened bearing.
First step
Load the 11 pulp-mill cases and the wind turbine record, run the envelope and kurtogram baseline and two networks trained on Paderborn data, and note the day each first flags a fault.
Done looks like
At least ten detectors scored on every public field case and on at least 30 newly donated cases from at least three plants; for each detector, the median warning time before removal, the share of failures missed and false alarms per machine-year, all published with code.
Open 02 · the brief
Open bearing frequency table
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A free, machine-readable table that gives, for common bearing designations from the main makers, the internal geometry (number of balls or rollers, their diameter, the pitch diameter and the contact angle) and the four defect frequencies per shaft revolution that vibration analysts look for: outer ring, inner ring, rolling element and cage. Each value carries its source.
Why it is still open
Every analyst needs these numbers, and they sit inside paid software. The help pages of Metra's VibroMetra system describe a built-in database of more than 25,000 bearing types with their damage frequencies. GitHub searches in November 2026 for bearing frequency databases found nothing, and those for calculators found tools that need the geometry typed in, such as SDT North America's bearing frequency calculator.
Built with what exists
Makers' published catalogs and data sheets; pdfplumber to read their tables; an AI model to draft entries from catalog pages, every value checked by a person (see the Note "Seven more flagged"); bought bearings cut open by a machine shop, and a vernier caliper to count and measure the parts; Python to compute the frequencies.
First step
Collect what three makers publish for their 6200 series of deep-groove ball bearings, and cut open one bearing of each of five sizes to check the count and size of the balls.
Done looks like
At least 2,000 designations from at least five makers, each with its source; 100 checked against bearings cut open, with the error in each frequency published and under 1%; the table released under an open license.