Create README.md

README.md

@@ -0,0 +1,447 @@
+---
+tags:
+- synthetic-data
+- vibration
+- condition-monitoring
+- predictive-maintenance
+- anomaly-detection
+- industrial
+- rotating-machinery
+- time-series
+- anndata
+- vibframe
+- vibsynth
+pretty_name: VibSynth Motor-Pump Fleet 180d
+size_categories:
+- 100K<n<1M
+---
+# VibSynth Motor-Pump Fleet 180d
+
+> **Status: dataset files are not published yet.**  
+> This repository is being prepared as the public dataset card for the final **AnnData (`.h5ad`)** release. The data will be uploaded once the VibFrame → AnnData conversion and validation pipeline is complete.
+
+**VibSynth Motor-Pump Fleet 180d** is a synthetic condition-monitoring dataset representing **180 days of operation of a fleet of 12 equivalent variable-speed motor–centrifugal-pump units**.
+
+The scenario is designed for research on **unsupervised anomaly detection, condition monitoring, operating-regime modelling, progressive fault detection and machine-to-machine generalisation**.
+
+All measurements are **simulated**. No signal in this dataset comes from a real industrial plant or from a real monitored machine. The vibration data were generated with **VibSynth, a synthetic vibration generation tool developed by TWave**.
+
+The aim is not to present synthetic data as field measurements, but to provide a controlled industrially plausible benchmark in which the complete physical scenario and ground truth are known while the learning problem remains non-trivial.
+
+---
+
+## Why this dataset exists
+
+Industrial vibration datasets suitable for machine learning often suffer from one or more of the following limitations:
+
+- very few machines;
+- a single fixed operating condition;
+- abrupt, artificially injected faults;
+- insufficient healthy history;
+- little or no information about transitions between regimes;
+- a lack of machine-to-machine variability;
+- no distinction between an unknown operating condition and an actual mechanical degradation;
+- incomplete ground truth.
+
+This dataset was built specifically to create a more demanding controlled scenario.
+
+The fleet shares the same mechanical architecture, but individual units are not numerical clones. Machines operate under several partially shared process regimes, move naturally between them, and may develop one or more progressive faults at different moments in time.
+
+The first part of every machine's history is guaranteed to be healthy, allowing models to learn normality before faults appear.
+
+---
+
+## Industrial scenario
+
+The simulated installation represents a process plant containing **12 motor–pump assemblies**, identified as `M01` to `M12`.
+
+Each unit consists conceptually of:
+
+```text
+Variable-frequency drive
+        │
+Electric motor
+        │
+     Coupling
+        │
+Centrifugal pump
+        │
+Process line
+```
+
+The twelve units belong to the same equipment family and use the same mechanical structure and sensing arrangement.
+
+Each pump is monitored on the pump drive-end bearing location using three vibration directions:
+
+- `pump_DE_H` — radial horizontal;
+- `pump_DE_V` — radial vertical;
+- `pump_DE_A` — axial.
+
+Small reproducible unit-to-unit tolerances are applied so that each machine has its own healthy vibration signature while remaining representative of the same fleet.
+
+---
+
+## Why the machines do not all use the same operating modes
+
+The pumps are mechanically equivalent, but they are assigned to **different process services inside the same installation**.
+
+This is intentional.
+
+A real plant may purchase the same pump package for several lines while using those units under different hydraulic requirements. For example, one line may routinely need short high-flow periods, another may regulate flow by throttling, and another may periodically recirculate product or coolant.
+
+For this reason, all machines share the two fundamental operating conditions, while additional modes occur only in the subset of units whose process duty requires them.
+
+### Operating regimes
+
+| ID | Name | Base speed | Base load | Physical interpretation |
+|---|---|---:|---:|---|
+| `R1` | ECO | 1,000 rpm | 0.375 | Low process demand and reduced-speed operation |
+| `R2` | NOMINAL | 1,450 rpm | 0.675 | Standard production duty |
+| `R3` | PEAK | 1,700 rpm | 0.900 | Temporary high-flow / high-demand operation |
+| `R4` | THROTTLED | 1,450 rpm | 0.800 | Flow regulation with partially restricted discharge |
+| `R5` | RECIRCULATION | 1,225 rpm | 0.300 | Recirculation, flushing or process stabilisation |
+
+An important part of the scenario is that **operating regime is not reducible to rotational speed alone**.
+
+`R2` and `R4`, for example, use the same nominal speed but represent different hydraulic conditions and loads. A model must therefore deal with several forms of normal behaviour rather than simply map RPM to health.
+
+### Regimes available to each unit
+
+| Machine | Normal regimes |
+|---|---|
+| `M01` | R1, R2, R3 |
+| `M02` | R1, R2, R3 |
+| `M03` | R1, R2, R3, R4 |
+| `M04` | R1, R2, R3 |
+| `M05` | R1, R2, R4 |
+| `M06` | R1, R2, R4 |
+| `M07` | R1, R2, R4, R5 |
+| `M08` | R1, R2, R4 |
+| `M09` | R1, R2, R3, R5 |
+| `M10` | R1, R2, R3, R5 |
+| `M11` | R1, R2, R5 |
+| `M12` | R1, R2, R5 |
+
+The operating history is stateful. Normal dwell periods last approximately **1–12 hours**, and machines do not change state in lockstep.
+
+---
+
+## Transitions are part of normality
+
+Real machinery does not jump instantaneously from one process condition to another.
+
+The dataset therefore contains explicit transitions between regimes. During a transition, rotational speed, load and the corresponding vibration response evolve continuously until the target state is reached.
+
+These periods are considered **legitimate normal operation**, not faults.
+
+This distinction is useful for anomaly-detection research because a detector should ideally avoid raising alarms merely because a healthy machine is accelerating, decelerating or changing process duty.
+
+---
+
+## Timeline
+
+The scenario starts at:
+
+**2026-01-01 00:00:00 UTC**
+
+and spans **180 days**.
+
+The history is divided conceptually into three temporal regions:
+
+| Period | Approx. days | Content | Intended use |
+|---|---:|---|---|
+| Calibration | 0–30 | Healthy operation only | Learn normal behaviour |
+| Normal holdout | 31–45 | Healthy unseen operation | Evaluate false positives |
+| Future stream | 46–180 | Healthy operation, OOD events and progressive faults | Online / temporal evaluation |
+
+No mechanical fault begins before **day 60**.
+
+The splits are temporal by design. Random row-wise train/test splitting is discouraged because it would leak information between neighbouring samples from the same evolving machine.
+
+---
+
+## Measurements
+
+### Spectra
+
+A spectral snapshot is generated every **20 minutes**.
+
+For every machine and timestamp, the dataset contains measurements for the three sensor directions and two processing modes:
+
+- `VEL_1K` — velocity spectrum from 0 to 1,000 Hz;
+- `ACC_10K` — acceleration spectrum from 0 to 10,000 Hz.
+
+Each spectrum contains **1,601 frequency bins**.
+
+This results in:
+
+- 12,960 spectral snapshots per machine;
+- 155,520 fleet-level snapshot timestamps;
+- 6 spectra per snapshot;
+- **933,120 spectral rows** in the source VibFrame fixture.
+
+### Waveforms
+
+A 1-second triaxial acceleration waveform is generated every **6 hours**.
+
+Each waveform channel contains:
+
+- sample rate: **51,200 Hz**;
+- samples: **51,200**;
+- duration: **1 second**.
+
+Across the complete fleet, the source fixture contains **25,920 waveform rows**.
+
+Waveforms and spectra generated at the same timestamp describe the same simulated machine condition, but they are not intended to be an exact FFT pair sample-by-sample.
+
+---
+
+## Progressive mechanical faults
+
+Faults are not introduced as abrupt binary changes.
+
+Each degradation begins at a machine-specific onset time and then grows in severity according to a fault-dependent curve. The underlying health deterioration persists across operating-mode changes.
+
+A machine can therefore be mildly degraded in one regime and show a much clearer vibration response in another.
+
+### Fault plan
+
+| Machine | Fault 1 | Onset day | Fault 2 | Onset day |
+|---|---|---:|---|---:|
+| `M01` | Imbalance | 70 | — | — |
+| `M02` | Healthy | — | — | — |
+| `M03` | Angular misalignment | 83 | — | — |
+| `M04` | Bearing outer-race fault | 96 | — | — |
+| `M05` | Imbalance | 65 | Bearing outer-race fault | 105 |
+| `M06` | Mechanical looseness | 110 | — | — |
+| `M07` | Healthy | — | — | — |
+| `M08` | Angular misalignment | 90 | Mechanical looseness | 125 |
+| `M09` | Bearing inner-race fault | 125 | — | — |
+| `M10` | Bearing inner-race fault | 80 | Resonance | 120 |
+| `M11` | Healthy | — | — | — |
+| `M12` | Resonance | 75 | — | — |
+
+Three machines — `M02`, `M07` and `M11` — remain mechanically healthy throughout the complete 180-day period.
+
+The maximum simulated severity is **0.95**.
+
+Different fault families use different progression shapes, including linear, quadratic, accelerated bearing-type and intermittent-growth profiles.
+
+When two faults coexist, their mechanisms and severities remain separately represented in the evaluation ground truth.
+
+---
+
+## Operational OOD events
+
+Not every unusual operating condition is a mechanical fault.
+
+To test this distinction, the scenario contains three **out-of-distribution operating events** on machines that remain mechanically healthy:
+
+| Machine | Days | Approx. RPM | Load | Interpretation |
+|---|---|---:|---:|---|
+| `M02` | 100–103 | 1,800 | 0.35 | Unusual high-speed / low-load duty |
+| `M07` | 150–152 | 1,100 | 0.92 | Unusual low-speed / high-load duty |
+| `M11` | 118–120 | 1,580 | 0.25 | Unseen speed/load combination |
+
+These episodes intentionally test whether an anomaly detector can distinguish **novel operation** from **mechanical degradation**.
+
+---
+
+## Dataset scale
+
+The validated source VibFrame fixture contains:
+
+| Item | Count |
+|---|---:|
+| Machines | 12 |
+| Duration | 180 days |
+| Fleet spectral snapshots | 155,520 |
+| Spectral rows | 933,120 |
+| Waveform rows | 25,920 |
+| Evaluation truth rows | 155,520 |
+| Spectral bins per row | 1,601 |
+| Samples per waveform row | 51,200 |
+| Source fixture size | ~11.86 GB decimal / ~11.05 GiB |
+
+The Hugging Face release will be distributed after conversion to **AnnData**.
+
+---
+
+## Planned AnnData release
+
+The public artifact is intended to be an **AnnData (`.h5ad`) representation derived from the validated VibFrame source dataset**.
+
+The conversion pipeline is designed so that raw measurements and derived machine-learning features remain conceptually separate.
+
+The final layout will be documented together with the release. The intended organisation is:
+
+- `obs` — snapshot-level metadata;
+- `obsm` — raw or structured high-dimensional measurements such as spectra and waveforms;
+- `X` — derived feature matrix when features have been materialised;
+- `var` — feature definitions;
+- `uns` — dataset, machine and processing metadata.
+
+The exact schema will be frozen only after the VibFrame → AnnData converter has completed validation.
+
+**Until then, no `.h5ad` file in this repository should be considered final.**
+
+---
+
+## Ground truth and label leakage
+
+The source dataset contains detailed evaluation truth, including:
+
+- machine identity;
+- operating regime;
+- transition state;
+- true rotational speed;
+- true load;
+- health state;
+- active fault types;
+- fault onset timestamps;
+- continuous fault severities;
+- OOD operating events;
+- reproducibility seed.
+
+This information exists because the scenario is simulated.
+
+It must not automatically be treated as model input.
+
+For unsupervised anomaly-detection experiments, fields such as `health_state`, `fault_*`, `ood_operational` and the evaluation split should remain isolated from the training features.
+
+The ground truth is intended primarily for **evaluation, interpretability and reproducibility**.
+
+---
+
+## Reproducibility
+
+The scenario uses master seed:
+
+```text
+20260910
+```
+
+Sub-seeds are derived deterministically for machine identity, operating timeline, spectra, waveforms and individual snapshots.
+
+The dataset is therefore intended to provide a reproducible controlled benchmark rather than an opaque collection of generated samples.
+
+---
+
+## Suggested research uses
+
+The dataset is particularly suitable for:
+
+- unsupervised and semi-supervised anomaly detection;
+- one-class learning;
+- fleet-level representation learning;
+- machine-to-machine transfer;
+- operating-regime-aware anomaly detection;
+- progressive fault / early-warning detection;
+- multi-fault behaviour;
+- novelty vs. mechanical-fault discrimination;
+- vibration feature engineering;
+- spectral representation learning;
+- multivariate H/V/A modelling;
+- temporal validation protocols;
+- explainable condition monitoring.
+
+It can also be used to compare classical vibration features against learned representations.
+
+---
+
+## Important limitations
+
+This dataset should not be interpreted as a replacement for field data.
+
+In particular:
+
+1. **All signals are synthetic.**
+2. The plant and its operating history are fictitious.
+3. VibSynth approximates mechanical and vibrational behaviour; it does not reproduce every structural, hydraulic, electrical or environmental non-linearity found in a real installation.
+4. Coincident spectra and waveforms share the same simulated condition but are not the same exact signal realisation.
+5. No keyphasor or tachometer pulse train is available.
+6. Stored spectra contain magnitude only; absolute shaft-referenced phase cannot be recovered.
+7. Waveforms are available every 6 hours rather than at every spectral snapshot.
+8. Fault evolution is intentionally controlled and therefore cleaner than many real industrial degradation histories.
+9. Results obtained on this dataset should ideally be validated later against independent real-world data.
+
+---
+
+## Data provenance
+
+The source fixture was generated using **VibSynth**, TWave's synthetic vibration generation tooling, and stored using the **VibFrame 0.2.0** data contract before conversion to AnnData.
+
+The source fixture passed its contractual validation with:
+
+- **0 errors**;
+- 12 non-blocking warnings related to the open `CentrifugalPump` node type;
+- 8 informational notes.
+
+The warnings do not affect the raw measurement schemas.
+
+---
+
+## Repository status
+
+### Current
+
+- [x] Industrial scenario defined
+- [x] 180-day VibSynth simulation generated
+- [x] VibFrame source fixture validated
+- [x] Ground truth validated
+- [x] Dataset card prepared
+- [ ] VibFrame → AnnData conversion finalised
+- [ ] Final `.h5ad` validated
+- [ ] Dataset files uploaded to Hugging Face
+- [ ] Loading / usage examples added
+- [ ] Versioned first public release
+
+The repository may therefore be visible before the actual dataset files become available.
+
+---
+
+## Citation
+
+A formal citation will be added before the first public release.
+
+For now, if you refer to the dataset in preliminary work, please cite the Hugging Face repository and clearly describe it as a **synthetic VibSynth-generated dataset**.
+
+```bibtex
+@dataset{vibsynth_motor_pump_fleet_180d,
+  title        = {VibSynth Motor-Pump Fleet 180d},
+  author       = {Gonzalez Zapico, Alejandro},
+  year         = {2026},
+  publisher    = {Hugging Face},
+  note         = {Synthetic industrial vibration dataset generated with VibSynth by TWave}
+}
+```
+
+---
+
+## Acknowledgements
+
+This dataset was created in the context of work involving **TWave** and the **University of Oviedo**.
+
+Vibration data were simulated with **VibSynth**.
+
+---
+
+## License
+
+**To be defined before the first public release.**
+
+No reuse rights should be inferred from this placeholder. A concrete dataset license will be selected and added to the Hugging Face metadata before publication.
+
+---
+
+## Contact and feedback
+
+Issues, questions and reproducibility reports are welcome through the Hugging Face repository discussions once the dataset is released.
+
+When reporting results, please indicate:
+
+- whether raw spectra, waveforms or derived features were used;
+- whether rotational speed or regime information was included as an input;
+- which temporal split was used;
+- whether results are per-machine or fleet-level;
+- whether OOD operating periods were treated separately from mechanical faults.