Industries / Industrial equipment and services / Industrial services / Vibration analysis
Reading faults
Reading faults is working out, from a vibration recording, which part of a machine is failing and how badly.
4
Milestones achieved, from 1973 to 2012.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1973
Echo math reads gear sidebands
In 1963 B. P. Bogert, M. J. R. Healy and J. W. Tukey described the cepstrum, the spectrum of the logarithm of a spectrum, as a way to find echoes in time series. In a 1973 Brüel & Kjær application note, R. B. Randall applied it to gearboxes: many gear faults make the tooth-meshing vibration rise and fall, which puts evenly spaced extra peaks, called sidebands, around the meshing frequency, and the cepstrum turns that whole family into one peak at its spacing. On a gearbox between a gas turbine and an alternator it confirmed an 85 Hz modulation that was "much more marked" before the gearbox internals were replaced.
LessonRandall reported that other cepstra taken under different conditions gave a less clear result, and advised comparing only signals processed under identical conditions.
Credit
R. B. Randall, Brüel & Kjær; B. P. Bogert, M. J. R. Healy and J. W. Tukey
1989
Spectra that follow the shaft
An order is a multiple of the shaft's running speed, and when the speed changes, an order moves across an ordinary spectrum. Ron Potter and Mike Gribler described computed order tracking, which resamples the recorded signal at equal steps of shaft angle in software, in place of the tracking filters and frequency synthesizers used before; they wrote that it follows rapid changes in speed with no time delay. Xiao Bing Wu wrote in 2013 that most commercial signal processing software has it built in, giving Brüel & Kjær's PULSE 7702 order tracking software as an example.
Credit
Ron Potter and Mike Gribler
2009
Gear indicators put to test
Time synchronous averaging adds up many turns of a gear's vibration, each started at the same shaft angle, so that vibration not tied to that gear averages away. FM4, from R. M. Stewart's 1977 report, measures how peaky the averaged signal is once the normal meshing tones are removed, to catch one damaged tooth; NA4 was later developed at NASA Glenn to stay sensitive as damage spreads. The US Army Research Laboratory at NASA Glenn, with Boeing in Mesa, Arizona, scored both on 24 sets of test gears, 17 of which ended with pitted, flaking teeth: at the thresholds chosen, NASA's own NA4 caught 99% of faulty readings with under 1% false alarms, FM4 caught 61% to 77%, and the overall vibration level 59% to 74% with 14% to 16% false alarms.
LessonFM4 was built for a single damaged tooth and lost sensitivity as pitting spread over many teeth, so an indicator has to match the damage it is meant to find.
Credit
R. M. Stewart; David G. Lewicki, Paula J. Dempsey, Gregory F. Heath and Perumal Shanthakumaran
2012
Blind test on real damage
A test gearbox in the US Gearbox Reliability Collaborative was damaged in a wind plant by two losses of oil, then rerun on the National Renewable Energy Laboratory's 2.5 MW test stand before it was taken apart. NREL gave the vibration recordings to 16 partners, seven universities and nine companies, who had two months and no knowledge of the damage. Teardown showed seven instances of damage that all partners agreed vibration could detect; the best partner found five of the seven, and most partners missed more than they falsely flagged. NREL released the recordings of a healthy and the damaged gearbox in 2014 so that others could test their methods.
Lessonon damage that grew in service, not faults cut in by hand, the best methods of 2012 still missed some, and NREL concluded there was "room for the industry to improve".
Credit
NREL (Shawn Sheng, editor) and the round robin partners
Watch for
Language models over diagnostic tools
Donghwan Kim reports that across five diagnostic data sets, including the Case Western bearing vibration set, conflicting outside information overturned large language model judgments that had been correct, and that a model given a specialist tool's answer scored below the tool alone: 77.43% against 83.33% on a set from the Tennessee Eastman simulated chemical process.
Achieved whenthe work passes peer review and another group repeats the comparison on recordings from working machines.
Claimed by
Donghwan Kim (arXiv preprint), October 2026
Open
01
Score the textbook fault rules
A public scorecard for the rules analysts are taught for reading a spectrum, each written as code and run on recordings whose fault is known. Fluke's guide gives examples: unbalance shows as a peak at once per turn (1X) growing in the radial directions but not the axial one; misalignment as raised 1X along the shaft and twice per turn (2X) across it; looseness as many harmonics of 1X. Report each rule's hit rate and false-alarm rate, and whether a threshold tuned on one data set holds on another.
The briefCredit · unclaimed02
Replay the gearbox into monitors
A test that plays NREL's recordings from its healthy and damaged test gearboxes into commercial condition monitoring systems, and publishes which of the seven damage instances found at teardown each system flags and how many false alarms it raises on the healthy recordings.
The briefCredit · unclaimed
Open 01 · the brief
Score the textbook fault rules
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public scorecard for the rules analysts are taught for reading a spectrum, each written as code and run on recordings whose fault is known. Fluke's guide gives examples: unbalance shows as a peak at once per turn (1X) growing in the radial directions but not the axial one; misalignment as raised 1X along the shaft and twice per turn (2X) across it; looseness as many harmonics of 1X. Report each rule's hit rate and false-alarm rate, and whether a threshold tuned on one data set holds on another.
Why it is still open
Rule-based software weights its rules without publishing the evidence. Erbessd Instruments gives one unbalance rule a positive predictive weight of 2.0 and a negative one of 10.0, with no data shown for either. The two open-source rule systems found on GitHub, one a student prototype that simulates its sensors with knobs, report no score on recorded data. A web search in November 2026 found no study that scored the classical rules on a public fault data set.
Built with what exists
MaFaulDa, a free database of 1,951 recordings from a SpectraQuest fault simulator covering normal running, imbalance, horizontal and vertical misalignment and bearing faults; NREL's gearbox recordings; Python with NumPy and SciPy, with computed order tracking where speed varies.
First step
Code the three Fluke rules and run them on MaFaulDa's 49 normal, 333 imbalance and 498 misalignment recordings.
Done looks like
A public table of hit and false-alarm rates for at least ten named rules on at least two public data sets, with code that anyone can rerun.
Open 02 · the brief
Replay the gearbox into monitors
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test that plays NREL's recordings from its healthy and damaged test gearboxes into commercial condition monitoring systems, and publishes which of the seven damage instances found at teardown each system flags and how many false alarms it raises on the healthy recordings.
Why it is still open
NREL released the data in 2014 because it found "very few validation and verification efforts on commercial wind turbine CM systems". Its 2012 round robin scored the analysis methods of 16 partners, without their names, not the products sold to wind farms. A web search in November 2026 found no system-by-system results on the data.
Built with what exists
The NREL benchmarking files on the Open Energy Data Initiative; a calibrated laboratory shaker carrying each system's own sensor, or an analog output card for systems that accept a voltage input; Python or MATLAB to resample the files to each system's input rate.
First step
Download both data sets, replay one healthy and one damaged file into one system, and check that the spectrum it records matches the spectrum computed from the file.
Done looks like
For each system tested, a table of the seven damage instances flagged or missed and the false alarms raised on the healthy files, published with the maker's name or with its consent.