EitaNis/iitk-compressor-bearing

IITK Reciprocating Air Compressor — Bearing Benchmark

The equipment-axis benchmark for compressor bearing-fault detection: acoustic recordings from a real single-stage reciprocating air compressor, belt-driven by a 5 HP / 415 V / 50 Hz / 1440 rpm direct-on-line induction motor under pressure-switch duty cycling. Eight labelled states — healthy, a bearing fault, and six compressor-native confounders (flywheel, piston ring, rider belt, leaking inlet valve, leaking outlet valve, non-return valve).

Derived from Verma et al. (2016) and repackaged as parquet with gain-invariant spectral features. Published as a companion to a motor-current/vibration benchmark measured on a centrifugal pump: that one has the right sensors on the wrong machine, this one has the right machine with only a microphone.

Provenance and licence

• Original source: https://www.iitk.ac.in/idea/datasets/
• Licence: air-compressor-dataset-bsd-2-clause-style (https://www.iitk.ac.in/idea/datasets/)
• Attribution required: Nishchal K. Verma, R. K. Sevakula, S. Dixit and A. Salour, "Intelligent Condition Based Monitoring using Acoustic Signals for Air Compressors", IEEE Transactions on Reliability, vol. 65, no. 1, pp. 291-309, 2016. Copyright 2016 Nishchal K. Verma, R. K. Sevakula, S. Dixit and A. Salour.

Why this redistribution is permitted

The archive ships a license.txt granting redistribution explicitly: "Redistribution and use of this dataset, with or without modification, are permitted provided that the following conditions are met: 1. Redistribution in any form must retain the above copyright notice, this list of conditions and the following disclaimer. 2. Redistribution and use in any form must be accompanied by the following citation: [Verma et al. 2016]." That is a BSD-2-Clause-style grant, not a bare citation request, and both conditions are satisfied here: the notice and the citation are reproduced in this card.

Configurations

• benchmark — splits: test
• raw — splits: test

from datasets import load_dataset

ds = load_dataset("EitaNis/iitk-compressor-bearing", "benchmark")
ds = load_dataset("EitaNis/iitk-compressor-bearing", "raw")

Why this sits alongside a current/vibration benchmark

A centrifugal pump imposes a smooth load torque. A reciprocating compressor does not — it imposes a strongly periodic load at the compression-cycle rate, plus valve events, plus pressure-switch duty cycling. For a predictive-maintenance system whose first target is an HVAC compressor, that nuisance structure is the thing most likely to break a bearing detector, and it is structurally absent from pump data.

This dataset supplies the machine. It does not supply the sensors: it is acoustic only, one microphone, so it cannot exercise motor current signature analysis or the ISO 10816-3 vibration-velocity measurand. The two benchmarks are complementary and neither is sufficient alone.

Separability

Bearing vs Healthy, single-feature AUC on the published rows (0.5 is chance; values near 0 mean the feature separates with the opposite sign):

feature	AUC (Bearing vs Healthy)
band_0_200_frac	0.955
band_200_500_frac	0.157
band_500_1000_frac	0.001
band_1000_2000_frac	0.065
band_2000_4000_frac	0.000
kurtosis	0.961
crest_factor	0.860
spectral_centroid_frac	0.017

Read these with suspicion, not satisfaction. Several features separate almost perfectly, which is the same pattern that earned CWRU its "too easy" reputation. The recordings carry no session, date or microphone-position metadata, so it is not possible to verify that each class was not recorded in one sitting — and if it was, a model can learn the sitting rather than the fault. The features published here are all gain-invariant (fractional band energies and distribution shape) specifically to remove the most obvious version of that shortcut, but spectral-shape confounds would survive it. Treat a high score here as necessary, not sufficient.

Sample rate is unresolved

The source paper describes 50 kHz recordings; the files hold exactly 50,000 samples; the WAV headers stamp 16 kHz. There is no measurable energy above ~4 kHz, consistent with the paper's own preprocessing (filenames are preprocess_ReadingN.wav).

nominal_sample_rate_hz records the header value and sample_rate_is_unverified is true on every row. Nothing in the published features depends on it: fractional band energies and shape statistics are invariant to a uniform rescaling of the frequency axis. Any work needing absolute frequencies — a compression-cycle harmonic, a bearing defect frequency — must resolve the rate first against the 1440 rpm motor and the belt ratio.

What is and is not labelled

• Class balance: bearing 225, flywheel 225, healthy 225, leaking_inlet_valve 225, leaking_outlet_valve 225, non_return_valve 225, piston_ring 225, rider_belt 225.
• fault_type is bearing_wear for the bearing class, healthy for healthy, and null for the six confounders — they are neither, and forcing them into one would put mislabelled rows in the benchmark. is_confounder marks them.
• There is no severity grading and no run-to-failure trajectory, so severity, remaining-useful-life and time-to-threshold are unexercised here.
• No session or microphone-position metadata exists, so a grouped split is not possible. This is a real limitation, not an omission.

Citation

@article{verma2016aircompressor,
  title   = {Intelligent Condition Based Monitoring using Acoustic Signals for Air Compressors},
  author  = {Verma, Nishchal K. and Sevakula, Rahul K. and Dixit, Sonal and Salour, Al},
  journal = {IEEE Transactions on Reliability},
  volume  = {65},
  number  = {1},
  pages   = {291--309},
  year    = {2016},
  doi     = {10.1109/TR.2015.2459684}
}