README.md

---
license: cc-by-nc-4.0
task_categories:
  - tabular-classification
  - tabular-regression
  - time-series-forecasting
language:
  - en
tags:
  - synthetic
  - predictive-maintenance
  - pdm
  - rul
  - remaining-useful-life
  - prognostics
  - phm
  - condition-monitoring
  - oil-and-gas
  - rotating-equipment
  - fft
  - vibration
  - lubrication
  - thermal
  - iso-17359
  - iso-13373
  - iso-14224
  - api-rp-670
  - api-rp-691
  - api-rp-580
  - iogp
  - degradation-modeling
pretty_name: "OIL-039 — Synthetic Predictive Maintenance Dataset (Sample)"
size_categories:
  - 100K<n<1M
---

# OIL-039 — Synthetic Predictive Maintenance Dataset (Sample)

A schema-identical preview of **OIL-039**, the XpertSystems.ai synthetic
predictive-maintenance and prognostics dataset for oil & gas rotating and
stationary assets. The full product covers ~12,000 assets across a 730-day
horizon with high-fidelity 7-band FFT decomposition. This sample is the
generator's `sample` mode (250 assets × 90 days × 2 samples/day) covering
all 13 product tables, with **pre-built per-timestamp RUL labels and 7d/30d
failure probabilities** ready for PHM model training.

> **Built by** XpertSystems.ai — Synthetic Data Platform
> **Contact** [pradeep@xpertsystems.ai](mailto:pradeep@xpertsystems.ai) · [xpertsystems.ai](https://xpertsystems.ai)
> **License** CC-BY-NC-4.0 (sample); commercial license available for the full product.

---

## OIL-039 vs OIL-038 — what's different

OIL-039 and OIL-038 are complementary upstream-asset PdM products covering
different research workloads:

| Dimension | **OIL-039** (this dataset) | OIL-038 (equipment-failure events) |
|---|---|---|
| Primary workload | **PHM / RUL prognostics** | **Failure-event analytics + reliability KPIs** |
| ML labels | Per-timestamp RUL + 7d/30d failure probs | Per-asset 30d/90d failure probability |
| FFT decomposition | 4-band (sample) / 7-band (full) | None |
| Failure-event detail | Compact (3 tables) | Rich (3 equipment groups × 16 tables) |
| Telemetry tables | 6 modalities (vibration + FFT + thermal + lubrication + pressure + acoustic) | 5 modalities (vibration + thermal + lubrication + environmental + alarms) |
| Time density | 2 samples/day at sample scale | 1 sample/day at sample scale |
| Best for | Time-series prognostics, RUL regression, fault-signature classification | Reliability KPI benchmarking, MTBF/MTTR fitting, multi-modal anomaly detection |

Buy or download **both** for full PHM + reliability coverage. They share the
upstream-asset and ISO 14224 / API RP 580 / API RP 670 calibration heritage.

---

## What's inside

13 CSV tables covering the complete PdM data plane: equipment master →
6-modality telemetry (vibration + FFT + thermal + lubrication + pressure +
acoustic) → health scores → RUL labels → failure probabilities →
maintenance work orders → failure & downtime events.

| Table | Rows (sample) | What it represents |
|---|---:|---|
| `equipment_master.csv` | 250 | 10-type asset master with criticality, MTBF, maintenance strategy |
| `vibration_signatures.csv` | 45,000 | RMS, peak, kurtosis, crest factor, sensor quality |
| `fft_spectra.csv` | 180,000 | 4-band FFT (1x, 2x, bearing, cavitation) × time × asset |
| `temperature_anomalies.csv` | 45,000 | Temperature, thermal gradient, anomaly score, thermal state |
| `lubrication_analysis.csv` | 45,000 | Viscosity, particle count, water ppm, contamination |
| `pressure_telemetry.csv` | 45,000 | Pressure, transient flags, flow rate |
| `acoustic_signals.csv` | 45,000 | Acoustic dB, ultrasonic energy, cavitation score |
| `equipment_health_scores.csv` | 45,000 | Per-timestamp health index, degradation, severity band |
| `remaining_useful_life.csv` | 45,000 | Predicted RUL hours/days + 5-class RUL bucket |
| `predictive_labels.csv` | 45,000 | **7d + 30d failure probability + target failure mode + root cause** |
| `maintenance_workorders.csv` | ~110 | 7-type repair categories with labor hours, parts cost |
| `failure_events.csv` | ~55 | IOGP severity (major/critical/catastrophic) + production loss |
| `downtime_events.csv` | ~55 | Downtime hours, production impact USD, restart success |

Total: ~610,000 rows, ~36 MB. The full OIL-039 product is ~140 million rows.

---

## Calibration sources

Every distribution and ratio is anchored to **named public references**. The
validation scorecard (see below) re-scores observed vs. target for 10
industry-anchored metrics, every one citing its source. Highlights:

- **SAE ARP4761 / API RP 691** — rotating equipment design MTBF benchmarks.
- **ISO 17359** Condition monitoring + **ISO 13373-1** Vibration monitoring —
  crest factor and kurtosis severity bands.
- **API RP 670** Machinery protection systems — FFT decomposition standards.
- **ISO 10816 / 20816** Mechanical vibration evaluation.
- **API RP 580** Risk-based inspection — criticality-tier distributions.
- **Reliability Web** Maintenance Strategy Survey — proactive maintenance share.
- **ARC Advisory Group** Predictive Maintenance Maturity Survey — sensor
  detection share benchmarks.
- **IOGP Safety Performance Indicators Report** — incident severity pyramid.
- **ISO 14224:2016** Reliability and Maintenance Data — work-classification.
- **ISO 45001** Clause 10.2 — work-order closure benchmarks.
- **PHM Society** conventions — synthetic PdM dataset label quality norms.

---

## Validation scorecard

The wrapper ships a 10-metric scorecard (`validation_scorecard.json`) that
re-scores the dataset on every generation. Default seed 42 result:

| ID | Metric | Target | Observed | Source |
|---|---|---|---:|---|
| M01 | Median design MTBF (hours) | 4,000–15,000 | **4,506** | SAE ARP4761 / API RP 691 |
| M02 | Vibration crest factor (mean) | 3–7 | **4.92** | ISO 17359 / ISO 13373-1 |
| M03 | Proactive maintenance share (floor) | ≥ 0.45 | **0.624** | Reliability Web survey |
| M04 | Criticality tier ≥ 3 share | 0.60–0.80 | **0.692** | API RP 580 RBI |
| M05 | IOGP severity pyramid — major share | 0.45–0.75 | **0.679** | IOGP Safety Performance |
| M06 | Sensor-based detection share (floor) | ≥ 0.40 | **0.566** | ARC Advisory PdM |
| M07 | Work-order close rate (floor) | ≥ 0.65 | **0.748** | ISO 45001 / CCPS |
| M08 | Repair-type taxonomy coverage (floor) | ≥ 7 | **7** | ISO 14224:2016 |
| M09 | FFT frequency-band coverage (floor) | ≥ 4 | **4** | ISO 17359 / API 670 |
| M10 | Pre-built ML label quality (mean) | 0.92–0.98 | **0.959** | PHM Society |

**Grade: A+ (100/100). Verified across seeds 42, 7, 123, 2024, 99, 1.**

---

## Suggested use cases

- **Remaining Useful Life (RUL) regression** — `remaining_useful_life.csv`
  provides per-timestamp RUL targets in hours/days plus 5-class RUL buckets
  (<7d / 7–30d / 30–90d / 90d–1y / >1y). 45,000 timestamps × 250 assets
  is right-sized for LSTM, transformer, and gradient-boosting prognostics
  baselines.
- **Fault-signature classification from FFT** — 4-band FFT spectra
  (1x, 2x, bearing, cavitation) × `fault_signature` labels enables direct
  bearing-fault, cavitation, and misalignment classification training.
- **7-day + 30-day failure probability** — `predictive_labels.csv` carries
  both horizons calibrated via sigmoid on degradation index. Useful for
  early-warning vs. medium-term planning model comparisons.
- **Multi-modal degradation modeling** — 6 telemetry modalities are
  per-timestamp aligned per asset (vibration + FFT + thermal + lubrication +
  pressure + acoustic), enabling true multi-modal fusion research.
- **Maintenance-reset event detection** — degradation trajectories include
  stochastic "reset points" simulating maintenance interventions; useful
  for change-point detection and survival analysis with competing-risks
  models.
- **PHM Society challenge-style benchmarking** — pre-built target labels
  (target_failure_mode + target_root_cause) follow PHM Society conventions
  for end-to-end prognostics evaluation.
- **Maintenance strategy ROI quantification** — 4 strategies (preventive /
  condition_based / run_to_failure / reliability_centered) × workorder &
  downtime tables enable strategy-vs-availability ROI modeling.

---

## Loading

```python
from datasets import load_dataset

# Load equipment master
master = load_dataset(
    "xpertsystems/oil039-sample",
    data_files="equipment_master.csv",
    split="train",
)
# Load RUL labels and predictive labels for prognostics training
rul = load_dataset(
    "xpertsystems/oil039-sample",
    data_files="remaining_useful_life.csv",
    split="train",
)
labels = load_dataset(
    "xpertsystems/oil039-sample",
    data_files="predictive_labels.csv",
    split="train",
)
# Load multi-modal telemetry (each ~45K rows)
vibration = load_dataset(
    "xpertsystems/oil039-sample",
    data_files="vibration_signatures.csv",
    split="train",
)
```

Or with pandas directly:

```python
import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="xpertsystems/oil039-sample",
    filename="fft_spectra.csv",
    repo_type="dataset",
)
df = pd.read_csv(path)
```

All 13 tables join on:

- `equipment_id` → master ↔ all telemetry ↔ labels ↔ workorders ↔ failures
- `equipment_id` + `timestamp` → 6-modality telemetry per-row alignment
- `failure_id` → failure events ↔ downtime events
- `workorder_id` → maintenance work orders

---

## Schema highlights

**`equipment_master.csv`** — `equipment_id`, `facility_id`, `asset_type`
(10-class: centrifugal_pump / reciprocating_compressor / centrifugal_compressor
/ gas_turbine / electric_motor / control_valve / pipeline_segment / separator
/ heat_exchanger / gearbox), `facility_type` (8-class), `region` (8-class),
`manufacturer` (6-class), `install_date`, `asset_age_years`,
`criticality_score` ∈ {1, 2, 3, 4, 5}, `hazardous_service`, `offshore_flag`,
`maintenance_strategy` ∈ {preventive, condition_based, run_to_failure,
reliability_centered}, `design_mtbf_hours`, plus 6 baseline telemetry
reference values per asset.

**`vibration_signatures.csv`** — `timestamp`, `equipment_id`, `asset_type`,
`rpm`, `vibration_rms`, `vibration_peak`, `kurtosis` (ISO 13373-1
impulsive-fault indicator), `crest_factor` (ISO 17359 healthy 3–5 vs.
faulty 5–9), `severity_band` ∈ {normal, watch, warning, critical, failure},
`sensor_quality`.

**`fft_spectra.csv`** — 4 frequency bands at sample scale (1x, 2x, bearing,
cavitation), 7 in full product (adds 3x, bearing_inner, bearing_outer,
gear_mesh). Each row carries `dominant_frequency_hz`, `harmonic_amplitude`,
`spectral_energy`, and `fault_signature` (matched to the asset's target
failure mode when the band-specific multiplier exceeds the degradation
threshold).

**`remaining_useful_life.csv`** — `predicted_rul_hours`, `predicted_rul_days`,
`rul_confidence` ∈ [0, 1], `rul_bucket` ∈ {<7d, 7-30d, 30-90d, 90d-1y, >1y}.
PHM Society-style per-timestamp RUL targets.

**`predictive_labels.csv`** — `failure_probability_7d`,
`failure_probability_30d`, `maintenance_priority` ∈ {1, 2, 3, 4, 5},
`target_failure_mode` (50 distinct modes across asset types),
`target_root_cause` (12-class), `label_quality` ∈ [0, 1].

**`failure_events.csv`** — `severity` ∈ {major, critical, catastrophic}
(IOGP pyramid), `detected_by` ∈ {vibration_alarm, temperature_alarm,
operator_round, predictive_model, shutdown_trip}, `production_loss_bbl`,
`safety_impact_flag`.

---

## Calibration notes & limitations

In the spirit of honest synthetic data, a few things buyers of the sample
should know:

1. **Aggressive 90-day degradation simulation.** The sample window compresses
   a full degradation trajectory into 90 days for ML utility, so the
   `severity_band` distribution is skewed toward warning/critical/failure
   (≈ 70% combined), with only ~2% of timestamps in `normal`. This is
   **intentional** — it provides positive-class density for failure
   classifiers and RUL regressors. For studies that require steady-state
   healthy operations, filter to `degradation_score < 0.30` or use the
   early window (first 14 days). The full product simulates 730 days with
   slower drift and recovers a healthy `normal`-band majority.

2. **Vibration RMS units differ from OIL-038.** Mean vibration RMS in this
   dataset is ~0.39 (different unit normalization than OIL-038's ~7.14 in
   mm/s ISO 10816 units). Crest factor (mean 4.92) and kurtosis (mean 7.1)
   are validated against ISO 13373-1 / ISO 17359 instead, since they're
   dimensionless and directly comparable across calibrations. For absolute
   ISO 10816 vibration severity classification, use OIL-038.

3. **FFT fault_signature label sparsity.** The `fault_signature` column in
   `fft_spectra.csv` is 99.6% "none" in the sample. The label is set only
   when the band-specific multiplier exceeds a degradation-dependent
   threshold, which is rare in the sample window. For ML use,
   **derive your own threshold** from `harmonic_amplitude` × `spectral_energy`
   on the matched-fault-mode band, or use the full product's 7-band
   decomposition which exposes 3 additional fault signatures.

4. **Lubrication water-ppm trajectory.** Median water ppm is ~375 across
   the sample (above ISO 4406 clean threshold of 200) because the generator's
   water content formula `80 + 600 × degradation` puts most degraded assets
   above the clean threshold. This is consistent with the aggressive
   degradation simulation (point 1). For "healthy lubrication" baselining,
   filter to `contamination_level < 0.2`.

5. **Per-timestamp RUL skew.** The RUL bucket distribution at sample scale
   is heavily weighted toward `30-90d` and `90d-1y` (≈83% combined) with
   only ~17% in the <7d and 7-30d buckets that ML teams care about most
   for early warning. For balanced training, **oversample on
   `rul_bucket ∈ {<7d, 7-30d}`** or use the full product (730-day window
   exposes more imminent-failure windows per asset).

6. **Workorder + failure event counts.** Sample mode produces ~110
   workorders and ~55 failure events. These are sparse on purpose
   (modeling realistic event rates over 90 days at 250 assets) but limit
   small-sample statistics on severity/severity-pyramid metrics — the
   scorecard's M05 (IOGP major share) tolerance is intentionally widened
   to ±0.15 for this reason. The full product recovers tight pyramid
   ratios at production scale.

7. **Deterministic seeding.** All 13 tables are deterministic on `--seed`.
   Catalog default is seed 42. Seed sweep verifies Grade A+ across
   {42, 7, 123, 2024, 99, 1}.

---

## Commercial / full product

The full **OIL-039** product covers ~12,000 assets × 730 days × 8
samples/day (~140 million telemetry rows total), with high-fidelity 7-band
FFT decomposition, slower-drift degradation trajectories that recover
healthy `normal`-band majority, dense ground-truth failure labels with
balanced RUL bucket distributions, and configurable maintenance-strategy
mode-packs for ROI quantification. Available under commercial license —
contact [pradeep@xpertsystems.ai](mailto:pradeep@xpertsystems.ai).

XpertSystems.ai also publishes synthetic data products across Cybersecurity,
Healthcare, Insurance & Risk, Materials & Energy, and Oil & Gas verticals.
Catalog: [huggingface.co/xpertsystems](https://huggingface.co/xpertsystems).