# Bearing Fault Datasets

## Zenodo identification

- Zenodo record ID: 19329597
- Version DOI: https://doi.org/10.5281/zenodo.19329597
- Concept DOI: https://doi.org/10.5281/zenodo.19329596
- Publication date: 2026-03-30
- Creator: Jakub Rekem

## Overview

The archive contains processed acceleration-signal datasets derived from a simplified Adams multibody model of the Case Western Reserve University bearing test rig and from a selected subset of measured CWRU Bearing Data Center signals.

## Processing summary

Signals were detrended, converted to signal envelopes, resampled to 6 kHz, normalised per signal to the interval [0, 1] and segmented into 0.1 s windows of 600 samples. Variable overlap was used to mitigate class imbalance. Samples were batched in groups of eight and divided into training, validation and test subsets.

## Labels and domains

- Data are stored as TensorFlow tf.data.Dataset objects with float64 tensors.
- Each sample has a five-element label vector.
- Elements 1–3 are one-hot class labels in the order Healthy, inner-race fault (IR), outer-race fault (OR).
- Elements 4–5 are one-hot domain labels in the order Simulation, Experiment.
- Simulated samples use domain label [1, 0]. Experimental samples use [0, 1].
- The unlabelled experimental variant masks the first three elements with zeros; the labelled variant retains the class labels.

## Files

- DATASETS.zip: complete prepared TensorFlow datasets. Extract the archive before loading it.

The exact TensorFlow and Python versions are not recorded in the current metadata. Users should document the environment that successfully loads the serialized objects and consider exporting arrays and split indices to an open format.

## Related work and provenance

- Probable related Zenodo publication: https://zenodo.org/records/22305325
- CWRU Bearing Data Center (URL cited in the Zenodo metadata): http://csegroups.case.edu/bearingdatacenter/home
- Doctoral thesis describing model construction and data splits: https://www.vut.cz/en/students/final-thesis/detail/172180
- Benchmark study: https://doi.org/10.1016/j.ymssp.2015.04.021

Avoid leakage between training, validation and test windows. Before redistribution or derivative publication, verify the licence and attribution requirements of the original CWRU measurements separately from the CC BY licence applied to this prepared dataset.

## Licence

The Zenodo record declares Creative Commons Attribution 4.0 International (CC BY 4.0). Third-party source-data rights may impose additional attribution or reuse conditions. Cite this dataset DOI and the original CWRU source.