This fault dataset index curates access routes, not the data, and it says so up front

The repository holds indexes and notes, not the data

The repository is a README with twenty-one numbered sections, most pointing at a markdown file under `doc/`, plus an `images/` directory with the figures, a `papers/` directory and a Word document named `Rotating-machine-fault-data-set.docx` at the root.

What the repository does not hold is the data. The author's own framing in the introduction is that the datasets shared here all come from their original research groups and that the author's contribution is to organise the routes to obtain them. That distinction matters, and it is the difference between a data repository and a map of one.

The scope is rotating mechanical fault data for condition monitoring, prognostics and rolling element bearing diagnosis, which is visible in the topic list as well as in the section titles. The entries span the United States, Germany, France, Canada, Italy, Brazil and China, and they cover both fault diagnosis on healthy and damaged bearings and the harder problem of degradation over time, with several entries labelled as bearing life degradation data rather than simple fault classes.

The README marks a subset of entries with a star symbol, described in the introduction as indicating the more widely used datasets, and the first four such entries are the Case Western Reserve bearing data, MFPT, Paderborn and FEMTO-ST. A reader who wants a baseline to test a method against will start with those.

The maintenance picture is modest. The last push to the repository was on 2026-04-09, the repository is not archived, and there are no GitHub releases. The project is a curated list rather than a piece of software, and the licence is MIT.

The Case Western Reserve entry is the baseline every paper uses

The Case Western Reserve bearing data centre entry is the one that opens the curated set, and the note explains its role better than a dataset card would.

It says the data was produced at Case Western Reserve University, that it is the most widely used standard dataset in papers on bearing vibration signal processing and fault diagnosis, that the fault features are distinct, that there is a lot of reference literature available, and that it can serve as a basic validation dataset for methods. It also points out that many GitHub projects use this dataset as their example, and says a usage summary will be added later, with material welcome.

That is a candid and useful framing. The CWRU dataset is so widely used that a new method's performance on it is close to meaningless as a novelty claim, and the note's phrasing, a basic validation dataset, says so without being unkind about it.

The next entry, MFPT, describes the Mechanical Failure Prevention Technology society data assembled by Eric Bechhoefer, chief engineer at NRG Systems, released to promote research into bearing analysis. The note lists what the data contains, and the structure is instructive for anyone choosing a dataset: normal bearing data, outer race fault data under different loads, inner race fault data under different loads, and three real fault cases.

The load variation and the inclusion of real field failures are the two features that separate this set from a laboratory fault injection dataset. A method tested only on clean injected faults has not been tested on the case where the damage was real.

The Paderborn entry, from Paderborn University in Germany, is credited to Christian Lessmeier, Enge-Rosenblatt, Bayer and Zimmer, with the experiment designed in 2014, and the FEMTO-ST entry describes the dataset used during the PHM IEEE 2012 data challenge.

The notes on individual datasets carry the engineering detail

The IMS bearing degradation entry is the clearest example of what this index provides that a bare download link does not, and it reads like a rig specification sheet.

The data was collected on an endurance test rig at the University of Cincinnati and released in 2014. The rig has four double row bearings of type Rexnord ZA-2115, a stationary speed of 2000 rpm, a load of 6000 lbs applied onto the shaft and bearing by a spring mechanism, and PCB 253B33 high sensitivity quartz ICP accelerometers. The entry credits the team of Professor Jie Lee.

Every one of those numbers changes what you can do with the data. The spring mechanism load is why the data is a degradation series rather than a fixed condition set, the bearing type tells you the geometry a model is expected to learn, the accelerometer model tells you the frequency response you are working with, and the constant speed means the classic resampling-free diagnostics papers are not applicable here.

The Southeast University gearbox entry is even more detailed. It comes from a Drivetrain Dynamic Simulator and contains two subdatasets, bearing and gear data, with two working conditions where the rotating speed and load configuration are set to 20-0 and 30-2. Each file has eight rows of signals, and the entry names them: motor vibration, vibration of the planetary gearbox in the x, y and z directions, motor torque, and vibration of the parallel gearbox in x, y and z. It also says which rows are effective, namely rows 2, 3 and 4.

That last detail, the note that only certain rows carry usable planetary gearbox signals, is exactly the sort of thing that is unwritten in the associated paper and decisive when you build a dataset loader.

Some entries cannot be downloaded at all, and the index says so

The index is honest about the entries that are broken, which is more useful than a link that returns a dead page.

The entry for the third Industrial Big Data Innovation Competition states that the data required entering the competition, that use of the data requires authorisation from the copyright holder, and that it contains rotor part separation data for multiple compressors and turbines. It also says the competition has ended and the data can no longer be downloaded, that a white paper is expected from the organisers, and that authorised sharing would be possible. In other words, the entry records a dead end with a reason, rather than pretending the link works.

The Acoustics and Vibration Database entry is different in kind, a pointer to a public-interest site collecting vibration fault datasets, and the entry for a collection of mechanical equipment fault diagnosis data and technical material points to a Chinese site with a large amount of such material. The NASA CoE prognostics data repository entry links the NASA data and analytics repository, and there is a separate entry for the data and documentation collection.

The introduction also sets the general expectation. It says that some dataset downloads may be difficult or may already be unavailable, and that the author can be contacted in those cases, while adding that if the copyright holder has requirements, the data will not be provided. That last clause is the important one and it is the ethical core of the project: the author is offering help finding a route, not a copy of the data.

The introduction also asks users to have actually tried before asking, and says that people who simply take without trying are not welcome, while inviting new public data sources.

The Ottawa and Turin entries show the harder data shapes

Two entries stand out for the data shapes they represent, and both illustrate why a fixed-protocol dataset is not enough.

The University of Ottawa entry contains vibration signals collected from bearings in different health conditions under time-varying rotational speed. There are 36 datasets in total, and for each dataset there are two experimental settings, the bearing health condition and the varying speed condition. The health conditions are healthy, an inner race defect and an outer race defect. The speed conditions are increasing speed, decreasing speed, increasing then decreasing, and decreasing then increasing, which gives twelve distinct situations. To ensure the authenticity of the data, three trials were collected for each experimental setting, giving the 36 datasets. Each dataset has two channels, Channel_1 measured by an accelerometer and Channel_2 measured by an encoder for speed, all sampled at 200,000 Hz over ten seconds.

Variable speed is the hard case for bearing diagnosis, because the fault impulse train is no longer at a fixed frequency, and having the speed signal in a second channel is what makes resampling or order tracking possible. The entry links both the dataset on Mendeley Data and two published papers, one on multiple time-frequency curve extraction and one on tachometer-free and resampling-free diagnostics under time-varying speed.

The Politecnico di Torino DIRG BearingData entry describes data collected on a drilling rig in the DIRG laboratory of the Department of Mechanical and Aerospace Engineering, designed specifically for testing high speed aviation bearings. Acquisition is at variable speed, radial load and damage level, together with temperature measurements, and it is available as open access data through an FTP link to the university, with the paper describing the test rig.

The Brazil MAFAULDA entry from the Federal University of Rio de Janeiro links the data page and two published papers on similarity-based modelling for rotating machine failure classification and on differential evolution optimisation for a stacked sparse autoencoder applied to bearing fault diagnosis.

The range across these entries, from a stationary laboratory rig to a drilling rig with variable speed and load, is the real value of the index. A method that only works on one of these shapes will fail on the others, and the index is how you find out which shapes exist.

Editorial conclusion

Adopt this repository as a starting map for a fault diagnosis project, because the per-dataset notes on rig geometry, sensor setup, fault types and sampling rate are the things you otherwise discover by reading several papers. Do not treat it as a data host, since it holds no datasets itself and only links to publishers, some of which require registration or authorisation. Verify first that the download route for each dataset still works, that the licence terms of the original publisher allow your use, and that the sampling rate and channel layout in the note match the published paper before you design a pipeline around them.

Frequently asked questions

Official sources

1. hustcxl/Rotating-machine-fault-data-set on GitHub
2. Issues
3. License: MIT
4. README