Upload README.md

README.md

@@ -1,48 +1,581 @@
-# DRMHB examples
+---
+pretty_name: "DRMHB — Dynamic Rotating Machinery Health Benchmark"
+tags:
+  - synthetic-data
+  - vibration
+  - condition-monitoring
+  - predictive-maintenance
+  - anomaly-detection
+  - industrial
+  - rotating-machinery
+  - time-series
+  - multivariate-time-series
+  - anndata
+  - vibframe
+  - vibsynth
+size_categories:
+  - 100K<n<1M
+---
 
-The Hugging Face repository distributes **DRMHB in VibFrame form only**. AnnData
-files are generated locally with the companion
-[`vibframe-anndata`](https://413hq.github.io/vibframe-anndata-docs/latest/)
-package.
+# DRMHB — Dynamic Rotating Machinery Health Benchmark
 
-Validated package version for these examples:
+> **Status: pre-release.**  
+> The benchmark definition, source VibFrame and AnnData conversion pipeline are in the final validation stage. Public data artifacts will be uploaded after the final consistency checks.
 
-```bash
-python -m pip install "vibframe-anndata==0.2.0"
+> **VibFrame → AnnData tooling:** the companion package [`vibframe-anndata`](https://413hq.github.io/vibframe-anndata-docs/latest/) provides the supported conversion workflow from VibFrame to AnnData. The repository file `createanndata.py` is a minimal example that creates a **featureless AnnData** from the raw VibFrame, preserving the raw signal representation so that features can be materialized later. For the full API, configuration options and processing semantics, see the [package documentation](https://413hq.github.io/vibframe-anndata-docs/latest/).
+
+**DRMHB** is a synthetic benchmark for machine-learning research in **condition monitoring of rotating machinery**. It represents **180 days of operation of a fleet of 12 nominally equivalent variable-speed motor–centrifugal-pump units**, with machine individuality, multiple normal operating regimes, dynamic transitions, healthy out-of-distribution operation and progressive mechanical degradation.
+
+The benchmark is designed primarily for:
+
+- unsupervised and semi-supervised anomaly detection;
+- normality modelling under multiple operating regimes;
+- early fault detection;
+- machine-to-machine generalisation;
+- domain shift across nominally equivalent assets;
+- healthy-OOD versus mechanical-fault discrimination;
+- progressive and interacting fault analysis;
+- vibration representation learning;
+- waveform/spectrum feature engineering;
+- temporal and streaming evaluation.
+
+All measurements are **synthetic**. DRMHB is not presented as field data and should not be interpreted as a substitute for validation on real industrial machinery.
+
+---
+
+## Industrial scenario
+
+DRMHB simulates a standardised fleet of **12 variable-speed centrifugal motor–pump units**, identified as `M01` to `M12`.
+
+Each unit represents the same equipment class:
+
+```text
+Variable-frequency drive (VFD)
+        │
+Electric motor
+        │
+Mechanical coupling
+        │
+Centrifugal pump
+        │
+Process line
+```
+
+The monitored location is the **pump drive-end bearing housing**, with three vibration directions:
+
+- `pump_DE_H` — radial horizontal;
+- `pump_DE_V` — radial vertical;
+- `pump_DE_A` — axial.
+
+The machines share the same nominal architecture, but small deterministic variations are applied to parameters such as structural response, damping, transfer gains and degradation parameters. They are therefore members of the same fleet rather than numerical clones.
+
+The synthetic vibration source is **VibSynth**, developed by TWave. The source data are exported using the **VibFrame 0.2.0** contract and subsequently converted to AnnData for machine-learning workflows.
+
+---
+
+## Time horizon and acquisition cadence
+
+The simulated interval is:
+
+**2026-01-01 00:00:00 UTC → 2026-06-29 23:40:00 UTC**
+
+for a total of **180 days**.
+
+A snapshot is produced every **20 minutes**:
+
+- **12,960 snapshots per machine**;
+- **155,520 fleet-level snapshots**.
+
+The temporal history is divided into three intended evaluation regions:
+
+| Region | Approx. days | Content | Intended use |
+|---|---:|---|---|
+| Calibration | 0–30 | Healthy operation | Learn normality |
+| Normal holdout | 31–45 | Healthy unseen operation | Threshold calibration / false-positive analysis |
+| Future stream | 46–180 | Healthy operation, transitions, OOD and progressive faults | Streaming evaluation |
+
+Random row-wise train/test splits are discouraged because neighbouring samples from the same evolving machine are strongly related.
+
+---
+
+## Multi-regime normality
+
+The fleet operates under several legitimate process conditions.
+
+The scenario includes five regime families:
+
+| ID | Name | Interpretation |
+|---|---|---|
+| `R1` | ECO | Reduced-demand / reduced-speed operation |
+| `R2` | NOMINAL | Standard production duty |
+| `R3` | PEAK | Temporary high-demand operation |
+| `R4` | THROTTLED | Similar rotational speed to nominal operation but different hydraulic load |
+| `R5` | RECIRCULATION | Low-flow / recirculation-type operation |
+
+Not every machine uses every regime. This intentionally creates **partially shared operating domains** across the fleet.
+
+A key property of the benchmark is that normal behaviour is **not reducible to rotational speed alone**. Different hydraulic/process states can occupy similar RPM regions while producing different vibration responses.
+
+---
+
+## Dynamic operating transitions
+
+Operating transitions are treated as part of **healthy normality**.
+
+The revised scenario does not move between regimes using a simple direct interpolation of speed and load. Instead, the truth contains dynamic state variables such as:
+
+- commanded speed;
+- true rotational speed;
+- speed rate;
+- true load;
+- flow;
+- valve position;
+- torque;
+- thermal state;
+- recirculation / low-flow instability state.
+
+These variables evolve with lag and memory, allowing forward and reverse transitions to follow different trajectories.
+
+The current 180-day fixture contains **22,179 transition snapshots**, approximately 14% of the complete stream.
+
+This is intentional: a health detector should not raise an alarm merely because a healthy asset is accelerating, decelerating or changing process duty.
+
+---
+
+## Signal acquisition
+
+Each snapshot contains both spectral and time-domain information.
+
+### Spectra
+
+For each machine and timestamp, spectra are available for all three monitored directions and two processing modes:
+
+- `VEL_1K` — velocity spectrum, 0–1,000 Hz;
+- `ACC_10K` — acceleration spectrum, 0–10,000 Hz.
+
+Each spectrum contains **1,601 bins**.
+
+This gives:
+
+- 6 spectra per snapshot;
+- **933,120 spectral signals** across the complete fleet.
+
+### Waveforms
+
+Unlike the earlier development fixture, the current DRMHB revision contains a synchronized triaxial acceleration waveform at **every snapshot**.
+
+The waveform sampling rate is:
+
+**51.2 kHz**
+
+For storage efficiency, steady/quasi-steady snapshots persist approximately **1.8 shaft revolutions** rather than the complete acquisition buffer.
+
+Measured on the current AnnData fixture:
+
+- **155,520 waveform snapshots**;
+- **466,560 waveform channels/signals**;
+- **2,107,792,704 waveform samples**;
+- median stored length ≈ **3,825 samples**;
+- median stored duration ≈ **74.7 ms**;
+- median stored shaft coverage ≈ **1.800 revolutions**.
+
+Because the stored duration follows shaft period, slower operating points contain more waveform samples than faster ones.
+
+### Waveform-first spectral generation
+
+The short stored waveform is **not** used to approximate the high-resolution spectrum.
+
+The generation chain is:
+
+```text
+physical / operating state
+        ↓
+full internal waveform acquisition
+        ├── high-resolution spectral products
+        └── persisted short waveform slice
+```
+
+The spectrum and persisted waveform therefore originate from the **same simulated acquisition context and physical state**, while the short 1.8-revolution slice is retained primarily for time-domain ML, transient morphology and compact raw-signal studies.
+
+The FFT of the stored short slice is not expected to reproduce the persisted high-resolution spectrum.
+
+### Synthetic tachometer information
+
+The AnnData raw representation also stores synchronized tachometer edge information associated with the waveforms, enabling order-aware and shaft-referenced analyses without requiring RPM to be inferred only from the vibration signal.
+
+---
+
+## Progressive fault modelling
+
+Mechanical degradation is not represented only by a generic monotonic amplitude curve.
+
+Each fault family carries its own physical or semi-physical latent state, and the observable vibration response depends on both that state and the current operating condition.
+
+The simulated fault families are:
+
+### Imbalance
+
+Imbalance is represented through an evolving unbalance state with magnitude and phase. Its observable 1X response depends on shaft speed and the structural transfer path rather than on a direct `severity × constant` rule.
+
+### Angular misalignment
+
+Angular misalignment includes a geometric misalignment state coupled to operating load and thermal history. This introduces memory and regime dependence into the resulting harmonic response.
+
+### Bearing inner- and outer-race degradation
+
+Bearing faults use a progressive defect state with crack/spall-like evolution and impact/ring-down behaviour. The resulting BPFO/BPFI-related observables are intentionally not perfectly monotonic with latent degradation.
+
+### Mechanical looseness
+
+Looseness evolves through support/preload and clearance-related states. Intermittent impacts emerge from the nonlinear contact condition rather than from a purely time-based on/off schedule.
+
+### Resonance
+
+Resonance is represented through evolving structural parameters such as effective natural frequency, damping and frequency ratio. Observable vibration can increase strongly near a resonance crossing and decrease again even while the underlying degradation state continues to evolve.
+
+### Interacting faults
+
+Some machines contain two simultaneous mechanisms. These are not intended to behave as two completely independent additive labels; the simulated physical states can influence the manifestation or evolution of the other mechanism.
+
+---
+
+## Fault distribution
+
+Three machines remain mechanically healthy throughout the full horizon, while the remaining units develop one or two progressive faults.
+
+| Machine | Fault 1 | Fault 2 |
+|---|---|---|
+| `M01` | Imbalance | — |
+| `M02` | Healthy | — |
+| `M03` | Angular misalignment | — |
+| `M04` | Bearing outer-race fault | — |
+| `M05` | Imbalance | Bearing outer-race fault |
+| `M06` | Mechanical looseness | — |
+| `M07` | Healthy | — |
+| `M08` | Angular misalignment | Mechanical looseness |
+| `M09` | Bearing inner-race fault | — |
+| `M10` | Bearing inner-race fault | Resonance |
+| `M11` | Healthy | — |
+| `M12` | Resonance | — |
+
+The evaluation sidecar preserves the separate state and severity of each active mechanism.
+
+---
+
+## Healthy operational OOD
+
+Not every unusual observation is a fault.
+
+DRMHB contains deliberately unusual operating periods on machines that remain mechanically healthy. These events are outside the normal calibration operating distribution but are not degradation.
+
+This allows evaluation of the distinction:
+
+```text
+statistical novelty ≠ mechanical fault
 ```
 
-## 1. Create the featureless AnnData
+A useful health-monitoring model should ideally remain sensitive to progressive mechanical degradation without treating every unseen speed/load combination or legitimate process transition as a fault.
+
+---
+
+## Ground truth
+
+The source VibFrame stores evaluation truth separately from the learning representation.
+
+The current snapshot truth contains **155,520 aligned rows across 12 machine partitions** and includes both high-level labels and detailed latent states.
+
+Examples include:
+
+- machine identity;
+- temporal split;
+- operating regime;
+- transition flag, origin, destination and progress;
+- commanded and true speed;
+- speed rate;
+- load;
+- flow;
+- valve position;
+- torque;
+- thermal state;
+- recirculation state;
+- healthy OOD flag;
+- fault type and severity;
+- mechanism-specific physical states;
+- selected truth-side diagnostic observables;
+- reproducibility seed.
+
+The current truth representation contains **89 columns** after expanding mechanism-specific physical-state payloads in the reference EDA.
+
+### Leakage warning
+
+Ground truth is provided for **evaluation, interpretation and reproducibility**.
+
+Fields such as:
+
+- `health_state`;
+- `fault_*`;
+- latent physical degradation state;
+- `ood_operational`;
+- evaluation split information;
+
+must not automatically be used as model inputs in unsupervised anomaly-detection experiments.
 
-```bash
-python examples/01_create_featureless_anndata.py \
-    DRMHB.vibframe.zip \
-    DRMHB.float32.h5ad
+---
+
+## Dataset scale
+
+Current validated AnnData geometry:
+
+| Item | Value |
+|---|---:|
+| Machines | 12 |
+| Duration | 180 days |
+| Snapshot cadence | 20 min |
+| Snapshots per machine | 12,960 |
+| Fleet snapshots | 155,520 |
+| Spectral signals | 933,120 |
+| Spectral bins per signal | 1,601 |
+| Waveform snapshots | 155,520 |
+| Waveform signals | 466,560 |
+| Waveform samples | 2,107,792,704 |
+| Waveform sample rate | 51.2 kHz |
+| Median stored waveform | ~1.8 shaft revolutions |
+| Transition snapshots | 22,179 |
+| Base featureless AnnData size | ~13.85 GiB |
+| Base AnnData feature matrix | `(155520, 0)` |
+
+The base `.h5ad` intentionally contains **no derived features in `X`**. Features can be materialized incrementally from the persisted raw representations.
+
+---
+
+## AnnData representation
+
+The reference machine-learning artifact is a featureless raw AnnData representation generated with `vibframe-anndata`.
+
+Conceptually:
+
+```text
+obs
+├── snapshot metadata
+├── machine
+├── timestamp
+├── speed
+└── signal-availability flags
+
+X
+└── initially (n_snapshots, 0)
+
+var
+└── feature metadata after feature materialization
+
+obsm
+├── raw_spectra
+├── raw_spectra_lengths
+├── raw_spectra_positions
+├── raw_spectra_speed_hz
+├── raw_waveforms
+├── raw_waveforms_lengths
+├── raw_waveforms_positions
+├── raw_waveforms_speed_hz
+├── raw_waveforms_tacho_rising
+└── raw_waveforms_tacho_falling
+
+uns
+└── VibFrame / machine / channel / provenance metadata
 ```
 
-This is the recommended starting point for ML workflows. It ingests the raw
-VibFrame signals and metadata but leaves `X` with zero variables/features.
+Derived features can be added later without reopening the original VibFrame archive.
+
+---
+
+## Recommended evaluation protocol
+
+The benchmark is designed around **temporal normality modelling**.
+
+A recommended anomaly-detection protocol is:
+
+1. fit preprocessing and models only on the healthy calibration period;
+2. use the normal holdout period for threshold calibration;
+3. evaluate once on the future stream;
+4. count healthy transitions as normal;
+5. count healthy OOD operation as normal unless the research question explicitly concerns novelty detection;
+6. report both classification/ranking quality and detection timing.
+
+Useful metrics include:
+
+- AUROC;
+- AUPRC;
+- F1;
+- false-positive rate on ordinary healthy operation;
+- false-positive rate during healthy transitions;
+- false-positive rate during healthy OOD;
+- detection delay;
+- severity at first sustained detection;
+- per-machine and per-fault performance.
+
+---
 
-## 2. Inspect the generated AnnData
+## Compact reference baseline
 
-```bash
-python examples/02_inspect_anndata.py DRMHB.float32.h5ad
+The reference EDA includes four deliberately simple unsupervised baselines.
+
+All models are fitted only on calibration data. Thresholds are set using the **99th percentile of normal-holdout anomaly scores** and then evaluated on the future stream.
+
+| Model | AUROC | AUPRC | F1 | Healthy FPR | Transition FPR | OOD FPR | Median delay |
+|---|---:|---:|---:|---:|---:|---:|---:|
+| Robust distance | 0.792 | 0.815 | 0.360 | 1.03% | 0.83% | 0.00% | 8.11 d |
+| PCA reconstruction | 0.825 | 0.865 | 0.600 | 0.76% | 0.72% | 0.99% | 10.54 d |
+| Isolation Forest | **0.884** | **0.920** | **0.773** | 1.39% | 4.44% | 2.38% | 3.61 d |
+| One-Class SVM | 0.775 | 0.848 | 0.679 | 0.84% | 2.68% | 4.96% | **1.65 d** |
+
+These numbers are intended as a **sanity-check baseline, not a leaderboard**.
+
+They illustrate a useful property of DRMHB: different simple models exhibit different trade-offs between early detection, transition robustness and OOD robustness.
+
+---
+
+## Evidence of non-trivial healthy structure
+
+The reference EDA also checks whether healthy operation collapses into a trivially low-dimensional manifold.
+
+Using a compact engineered feature representation on healthy calibration data:
+
+- 10 principal components are required for 80% explained variance;
+- 14 for 90%;
+- 17 for 95%;
+- a TwoNN intrinsic-dimension estimate gives approximately **12.15**.
+
+This does not prove industrial realism, but it provides evidence that the benchmark's healthy operating space is not simply a one- or two-dimensional RPM trajectory.
+
+---
+
+## Suggested research uses
+
+DRMHB is particularly suitable for research on:
+
+- unsupervised anomaly detection;
+- semi-supervised health monitoring;
+- one-class learning;
+- early fault detection;
+- progressive-fault modelling;
+- machine-invariant representation learning;
+- fleet-level normality modelling;
+- domain generalisation;
+- regime-conditioned anomaly detection;
+- transition-aware monitoring;
+- healthy-OOD discrimination;
+- interacting faults;
+- raw waveform learning;
+- spectral representation learning;
+- order-aware vibration analysis;
+- classical feature engineering;
+- explainable condition monitoring;
+- temporal benchmark methodology.
+
+---
+
+## Important limitations
+
+DRMHB is a controlled synthetic benchmark and has important limitations:
+
+1. **All measurements are synthetic.**
+2. The plant and its operating history are fictitious.
+3. The simulated motor–pump train represents an equipment class rather than a digital twin of one specific commercial machine.
+4. Reduced physical and semi-physical models cannot reproduce every structural, hydraulic, tribological, electrical or environmental non-linearity present in a real plant.
+5. The stored ~1.8-revolution waveform is intentionally short and is **not** a substitute for the high-resolution spectrum.
+6. Spectral and waveform products are generated from the same acquisition context, but users should follow the documented processing semantics rather than assume arbitrary signal-processing equivalence.
+7. Ground truth is unusually complete because the scenario is simulated; real deployments do not provide this level of observability.
+8. Benchmark performance must not be interpreted as evidence of production readiness.
+9. Findings should ideally be validated on independent field datasets before making claims about real industrial performance.
+
+---
+
+## Reproducibility
+
+The scenario is deterministic from a master seed and derived sub-seeds.
+
+The reference scenario uses:
+
+```text
+20260910
 ```
 
-The script opens the H5AD in backed read-only mode, prints `obs`, `obsm` and
-`uns` structure, and does not load the full raw payload into RAM.
+Sub-seeds are derived for machine identity, operating dynamics, degradation mechanisms and signal realizations.
+
+The aim is to make the benchmark reproducible while retaining machine-to-machine and snapshot-to-snapshot variability.
+
+---
+
+## Provenance
+
+DRMHB was generated with **VibSynth**, synthetic vibration generation tooling developed by **TWave**, and serialized using **VibFrame 0.2.0**.
+
+The AnnData representation is produced through the `vibframe-anndata` conversion pipeline.
+
+The benchmark was developed in the context of collaboration involving:
+
+- TWave;
+- University of Oviedo;
+- Fundación Universidad de Oviedo.
+
+---
+
+## Repository status
 
-## 3. Materialize a small feature set
+- [x] Industrial scenario defined
+- [x] Dynamic multi-regime simulation generated
+- [x] Waveform-first signal generation enabled
+- [x] Dense waveform acquisition validated
+- [x] Dynamic transitions validated
+- [x] Progressive physical/semi-physical fault states validated
+- [x] Ground-truth alignment validated
+- [x] VibFrame → AnnData conversion validated
+- [x] Reference EDA executed
+- [x] Compact ML baseline executed
+- [ ] Final public artifacts uploaded
+- [ ] Loading examples finalized
+- [ ] License finalized
+- [ ] First public release tagged
 
-```bash
-python examples/03_add_features.py \
-    DRMHB.float32.h5ad \
-    DRMHB.example-features.h5ad
+---
+
+## Citation
+
+A formal publication citation will be added if/when a corresponding paper is released.
+
+For preliminary use, cite the dataset repository and make its synthetic nature explicit.
+
+```bibtex
+@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}
+}
 ```
 
-This demonstrates the second stage of the package: deriving features from raw
-signals already stored in AnnData, without reopening the original VibFrame.
+---
+
+## License
+
+**To be defined before the first public release.**
+
+No reuse rights should be inferred until an explicit dataset license is added to the repository metadata.
+
+---
+
+## Contact and reporting results
+
+Questions and reproducibility reports are welcome through the repository discussions.
 
-For the complete API, configuration schema, supported metrics and limitations,
-see:
+When reporting results, please state:
 
-https://413hq.github.io/vibframe-anndata-docs/latest/
+- which temporal split was used;
+- whether the model was trained only on calibration data;
+- whether raw spectra, waveforms or derived features were used;
+- whether RPM, regime or other operating metadata were supplied to the model;
+- whether healthy transitions were considered normal;
+- whether healthy OOD periods were considered normal or anomalous;
+- whether evaluation was per-machine or fleet-level;
+- how detection delay was defined.