DRMHB - Dynamic Rotating Machinery Health Benchmark

DRMHB is a synthetic benchmark for machine-learning research in rotating machinery condition monitoring. It represents 180 calendar days for a fleet of 12 nominally equivalent variable-speed motor-pump units, combining:

• multiple healthy operating regimes;
• explicit startup, shutdown and internal regime transitions;
• machine-to-machine individuality;
• rare healthy out-of-distribution (OOD) operation;
• realistic non-fault disturbances and sensor ageing;
• progressive single and interacting mechanical faults;
• synchronized triaxial spectra, waveforms and tachometer edges;
• a causal calibration/holdout/future evaluation protocol.

All measurements are synthetic. DRMHB is not field data and is not a substitute for validation on real machinery.

The ZIPs under data/ are the 2.6.0 compact campaign and its separate 512 paired controls. The conversion examples use vibframe-anndata==0.3.0 with complete evaluation preservation. The complete EDA, all report figures and baseline tables were rerun on the current campaign and checked on 1 October 2026. Measured identities and checks are in the dataset audit.

Start here: dataset report, conversion examples and complete ground-truth guide.

The challenge

The task is not merely to separate healthy and faulty rows. A useful model must learn fleet normality while avoiding false alarms caused by legitimate process changes. In particular:

• R2 and R4 have similar rotational speed but different hydraulic load;
• forward and reverse transitions follow different paths because state has lag and memory;
• adaptive acquisition heavily oversamples transitions relative to elapsed time;
• equivalent machines retain stable individual vibration fingerprints;
• healthy OOD conditions are statistically unusual but not damaged;
• observable vibration need not increase monotonically with latent severity;
• some machines contain two coupled degradation mechanisms.

Industrial scenario

The simulated plant contains 12 variable-frequency-drive motor-pump sets, identified as M01 through M12. Vibration is measured at the pump drive-end bearing housing in three directions:

• pump_DE_H: radial horizontal;
• pump_DE_V: radial vertical;
• pump_DE_A: axial.

Five active regimes are used:

ID	Name	Nominal meaning
R1	ECO	Reduced speed and demand
R2	NOMINAL	Standard production duty
R3	PEAK	Temporary high-demand duty
R4	THROTTLED	R2-like speed with different hydraulic load
R5	RECIRCULATION	Low-flow recirculating operation

OFF is also an explicit state. Machines operate Monday-Thursday from 08:00 to 22:30 UTC and Friday from 08:00 to 18:00 UTC; weekends remain stopped.

Time horizon, splits and adaptive acquisition

The fixture covers:

2026-01-01 00:00:00 UTC -> 2026-06-29 22:30:00 UTC

The 46,907 observations are divided as follows:

Split	Rows	Intended use
calibration	8,179	Learn healthy normality
normal_holdout	3,723	Calibrate thresholds and false-positive behaviour
future_stream	35,005	Causal evaluation with faults, transitions and OOD

Acquisition is deliberately non-uniform:

Context	Policy	Realized rows
Startup transition	Reference: every 5 minutes	8,729
Shutdown transition	Reference: every 5 minutes	8,784
Internal regime transition	Reference: every 10-15 minutes	14,562
Stable running	Reference: every 2-4 hours, plus reallocated captures	11,424
OFF	Every 8 hours	3,408

The compact profile reallocates 8,000 reference transition observations to steady operation while protecting endpoints, central progress and fast motor edges. There are 32,075 transition observations. Realized steady spacing has a median of 55 minutes; reference cadences are not strict spacing rules after reallocation. Observation count is therefore not proportional to elapsed time. Temporal plots and persistence metrics must use timestamps rather than row index.

Signal acquisition and storage

Every observation begins from a synchronized triaxial acceleration realization of 1.6 seconds at 51.2 kHz. Both stored spectral products are calculated from that complete source realization:

• VEL_1K: velocity spectrum from 0 to 1,000 Hz;
• ACC_10K: acceleration spectrum from 0 to 10,000 Hz.

VEL_1K stores native peak lines at 0.625 Hz spacing. ACC_10K stores the root-sum-square of native peak lines in 6.25 Hz bands, rather than a maximum line amplitude. Under its periodic-Hann convention, RMS is sqrt(sum(data**2)/3); apply the correction once. Each spectrum has 1,601 bins. With three directions and two processing modes, the dataset contains 281,442 spectral signals.

The persisted waveform branch is anti-alias filtered and decimated to 25.6 kHz. Its length is time-based rather than revolution-normalized:

Stored window	Samples/channel	Role
0.25 s	6,400	Ordinary OFF observation
0.40 s	10,240	Ordinary active or transition observation
0.75 s	19,200	Localized event
1.60 s	40,960	Stratified anchor

The current fixture contains 3,240 short OFF windows, 41,125 ordinary windows, 198 event windows and 2,344 full-length anchors. Each observation has H/V/A waveforms, giving 140,721 waveform signals and 1,625,003,520 persisted samples. The separate paired controls use 0.50 s / 12,800-sample windows.

Waveforms are quantized to an int16-equivalent grid and reconstructed as physical-unit float32 values to remain compatible with VibFrame 0.2.

Legitimate novelty and faults

Healthy active observations may contain low-probability disturbances such as RPM drift, load transients, electrical interference, sensor-noise bursts and transient impacts. The current fixture contains 1,756 such rows. Three healthy machines each receive eight operational-OOD episodes, totaling 24 episodes and 1,057 observations. The report lists every configured UTC interval, RPM and valve command.

Fault distribution:

Machine	Fault 1	Fault 2
M01	Imbalance	-
M02	Healthy	-
M03	Angular misalignment	-
M04	Bearing outer-race	-
M05	Imbalance	Bearing outer-race
M06	Mechanical looseness	-
M07	Healthy	-
M08	Angular misalignment	Mechanical looseness
M09	Bearing inner-race	-
M10	Bearing inner-race	Resonance
M11	Healthy	-
M12	Resonance	-

No fault begins before day 60. Fault severity is derived from mechanism-specific physical state, and observable response remains dependent on speed, load, thermal history and structural transfer.

Ground truth and leakage prevention

The source VibFrame keeps evaluation information in separate sidecars. The 0.3.0 workflow preserves both supported observation-aligned views and the original files, without automatically exposing targets to a model:

AnnData location	Information
obsm["ground_truth"]	Snapshot truth, aligned with obs_names
obsm["waveform_ground_truth"]	Per-waveform annotations with explicit raw-channel bindings
uns["vibframe_evaluation"]	Byte-exact original annotation files and source/path/size/SHA-256 manifest
uns["vibframe_anndata"]	Conversion, raw-channel and feature provenance
obsm["raw_*_capture_t"] and *_capture_t_known	Actual capture times and explicit/fallback flags

This includes construction labels, crop offsets and durations, quantization and hashes, machine parameters, scenario configuration and DiagGT documents/tables. All regular files under evaluation/, ground-truth/ and ground_truth/ are preserved, plus root JSON/YAML and machine JSON context. Unknown annotation fields are retained; the original Parquet schema can be recovered.

Snapshot joins use source, machine and integer UTC-microsecond snap_t, not row order. Per-waveform joins also identify the raw channel. Capture t can be later than snap_t because of cropping; it is not a different snapshot.

The current original snapshot-truth tables have 49 columns; the aligned view also contains conversion identity fields. Code must inspect actual fields instead of hard-coding a fixed schema. Do not use split, fault_*, severity, physical states, ood_operational, nuisance labels or DiagGT as unsupervised model inputs. Operational-state knowledge must be an explicitly declared experimental condition. The package never adds these fields to X implicitly.

Preservation is not automatic labelling: DiagGT intervals, consolidated observations and findings retain their original semantics. No nearest-time join is invented, and absence of a diagnosis does not mean healthy. The original archive remains dataset-wide when AnnData observations are sliced.

Distribution and local AnnData workflow

The longitudinal signal artifact is data/DRMHB-compact.vibframe.zip. Paired controls are in data/DRMHB-compact-controls.vibframe.zip; import them to a separate H5AD. H5ADs are local derivatives.

python -m pip install -r examples/requirements.txt

python examples/01_create_featureless_anndata.py \
    data/DRMHB-compact.vibframe.zip derived/DRMHB-compact.v0.3.0.raw.h5ad

python examples/02_inspect_anndata.py \
    derived/DRMHB-compact.v0.3.0.raw.h5ad --waveform-details

python examples/03_add_features.py \
    derived/DRMHB-compact.v0.3.0.raw.h5ad derived/DRMHB-compact.v0.3.0.features.h5ad

python examples/05_read_evaluation.py derived/DRMHB-compact.v0.3.0.features.h5ad

Example 01 explicitly enables ground_truth.scope="all", stores numerical values as float32, and ingests in raw-data blocks of 8 MiB. The featureless base has X.shape == (46907, 0) for the current longitudinal ZIP. The feature requests materialized 468 machine/source-scoped columns in the current EDA, with structural missing values for other machines. They are not 468 dense fleet-wide descriptors; the notebook explains the logical-column alignment.

Example 03 calculates features only from the raw H5AD and checks that original annotations and raw coverage are preserved. Example 02 reads HDF5 metadata and public annotation APIs selectively; it does not assume AnnData backed mode keeps obsm or the evaluation archive out of RAM.

For an existing 0.2.x H5AD, preserve the original source identities and run:

python examples/04_add_complete_ground_truth.py \
    derived/DRMHB-compact.v0.3.0.no-truth.h5ad data/DRMHB-compact.vibframe.zip \
    derived/DRMHB-compact.v0.3.0.enriched.h5ad

no-truth.h5ad represents an existing H5AD created from the same longitudinal ZIP with ground-truth import disabled; adjust the path to your actual file. The command adds annotations without recalculating X, var or raw samples. A transactional temporary copy needs roughly one additional H5AD of free disk. Legacy capture-time companion arrays are not retroactively recovered by this operation; new raw imports populate them. Crop information is still retained in waveform truth.

Storage and memory

The longitudinal ZIP is 7,635,324,956 bytes (7.111 GiB). The measured 0.3.0 featureless H5AD is 8,771,810,437 bytes (8.169 GiB) and the feature H5AD is 8,874,970,373 bytes (8.265 GiB). Both retain 97 original annotation files, totaling 26,874,052 bytes. The audit verifies all original annotation bytes against the ZIP, snapshot identities, waveform lengths and identical raw buffers in both H5ADs. These sizes include full annotations and capture-time matrices.

The default 512 MiB sidecar budget covers original encoded annotation files, not decoded tables or total process RAM. Inspect exact archived sizes with list_evaluation_files(). Raw-array compression remains none, as in the existing examples; no compression benchmark is rerun. Ground-truth conversion does not modify the VibFrame ZIP.

AnnData layout

obs   : snapshot identity, snap_t, machine, source, speed and availability
X     : n_snapshots × n_materialized_features (initially zero columns)
var   : feature descriptors, units, channel identity and provenance
obsm  : ragged raw spectra/waveforms + lengths/positions/speed
        tachometer edges + capture_t/known companions
        ground_truth + waveform_ground_truth (evaluation only)
uns   : vibframe_anndata (package/raw/feature provenance)
        vibframe_evaluation (original sidecar bytes and manifest)

See examples/README.md for precise commands and docs/GROUND_TRUTH.md for target selection, source identity, original DiagGT access and missing-data semantics.

Recommended evaluation protocol

1. Fit preprocessing and models only on calibration.
2. Use normal_holdout to select thresholds.
3. Evaluate once on future_stream.
4. Treat healthy transitions and healthy OOD as negative examples unless the research question explicitly targets novelty.
5. Use timestamps for detection persistence and delay.
6. Report per-machine and per-fault results alongside aggregate metrics.

Recommended metrics include AUROC, AUPRC, F1, ordinary healthy FPR, transition FPR, OOD FPR, wall-clock detection delay and the fraction of faulty machines with a confirmed detection.

The reference EDA confirms an alarm after six continuously positive wall-clock hours; an observation gap longer than 4.5 hours breaks the run.

Compact reference baseline

The values below come from the complete current 2.6.0 EDA using converter 0.3.0. Four deliberately small unsupervised baselines use 39 logical features, fitted on calibration only. Thresholds are the 99th percentile of normal-holdout scores.

Model	AUROC	AUPRC	F1	Healthy FPR	Transition FPR	OOD FPR	Median delay
Robust distance	0.741	0.771	0.237	0.616%	0.668%	0.000%	38.44 d
PCA reconstruction	0.735	0.768	0.329	1.017%	1.142%	3.784%	13.48 d
Isolation Forest	0.815	0.843	0.485	1.371%	1.675%	5.866%	53.25 d
One-Class SVM	0.687	0.774	0.508	0.842%	1.049%	2.744%	36.51 d

These are sanity checks, not leaderboard targets. Median delay excludes undetected machines: robust distance detects 2/9 faulty machines, PCA 3/9, and Isolation Forest and One-Class SVM each 5/9. Report coverage with delay. OOD false positives exceed ordinary healthy FPR for three models.

Healthy calibration requires 6, 9 and 13 PCA components for 80%, 90% and 95% cumulative variance, respectively. The TwoNN intrinsic-dimension estimate is 9.17.

The 512 paired controls contain 256 damage interventions and 256 healthy operating-state interventions (1,024 snapshots). Their timestamps are variant identifiers, not a degradation trajectory. fault_present and diagnostic_eligible (RPM >= 120) are separate truth fields; eligibility does not guarantee detectability. No RUL or failure-time target is defined.

Repository contents

DRMHB/
|- data/DRMHB-compact.vibframe.zip          current longitudinal signal dataset
|- data/DRMHB-compact-controls.vibframe.zip separate 512-pair controls
|- derived/                      local H5ADs and EDA exports (not versioned)
|- examples/                     import, inspect, features, retrofit, evaluation readers
|- notebooks/DRMHB_EDA.ipynb       complete executed 0.3.0 EDA on the 2.6.0 campaign
|- latexdocs/DRMHB.tex            editable report source
|- latexdocs/figures/             promoted EDA images and editable TikZ diagrams
|- latexdocs/provenance/          figure/table and report input identities
|- docs/DRMHB.pdf                 compiled current dataset report
|- docs/provenance/               measured current dataset audit
`- tools/                        audit and report build

The report explains the installation, schedules, acquisitions, degradation, paired controls, evaluation and complete AnnData workflow. The 37 figures and two baseline tables come from the complete current notebook execution; their manifest records hashes and notebook identity. The TikZ week was verified against the current full-horizon process recipe.

Run the audit with the notebook's Python environment:

python tools/audit_dataset.py --deep
python tools/build_report.py --refresh-assets

--deep checks every raw numerical buffer in the base and feature H5ADs. The build requires a complete sequential notebook execution and promotes all report figures/tables together. Rebuild without promotion using python tools/build_report.py; a LaTeX distribution with latexmk and TikZ is required because the report has multiple input files and image assets.

The source VibFrame validation reports have ok=true, strict_ok=false: the warning concerns the open node type CentrifugalPump. The consistency audit does not re-synthesise signals or independently validate the physical model.

Limitations

• The plant, signals and labels are entirely synthetic.
• The hydraulic and mechanical dynamics are reduced models, not CFD/FEM.
• The persisted 25.6 kHz waveform discards content above 12.8 kHz.
• Spectra contain magnitude, not absolute complex phase.
• Persisted waveform length depends on acquisition policy.
• Adaptive sampling biases count-based summaries unless elapsed time is used.
• Evaluation truth is cleaner and more complete than real maintenance labels.
• Performance on DRMHB is not evidence of production readiness.

Results should ideally be validated on independent field datasets.

Citation

@dataset{drmhb_2026,
  title     = {DRMHB: Dynamic Rotating Machinery Health Benchmark},
  author    = {Gonzalez Zapico, Alejandro},
  year      = {2026},
  publisher = {Hugging Face},
  note      = {Synthetic rotating-machinery health-monitoring benchmark generated with VibSynth by TWave}
}

License

DRMHB is released under the Creative Commons Attribution 4.0 International license (CC-BY-4.0). See LICENSE.md.

Reporting results

Please state:

• the temporal split and calibration procedure;
• which raw modalities or derived features were used;
• whether RPM, regime or other operating metadata were supplied;
• how transitions and healthy OOD were scored;
• the wall-clock alarm persistence rule;
• per-machine detection coverage and delay;
• whether evaluation truth was used only after prediction.