Hugging Face's logo

Datasets:
xpertsystems
/
oil040-sample

Tasks:
Tabular Classification
Tabular Regression
Time Series Forecasting
Languages:
English
Size:
100K<n<1M
Tags:
Synthetic
vibration
fft
condition-monitoring
predictive-maintenance
sensor-data
License:
Dataset card Data Studio Files
xet
 Community
oil040-sample / README.md
pradeep-xpert's picture
pradeep-xpert
Upload folder using huggingface_hub
e9e288e verified
preview code
|
Raw
History Blame Contribute Delete
14.9 kB
metadata 
license: cc-by-nc-4.0
task_categories:
  - tabular-classification
  - tabular-regression
  - time-series-forecasting
language:
  - en
tags:
  - synthetic
  - vibration
  - fft
  - condition-monitoring
  - predictive-maintenance
  - sensor-data
  - oil-and-gas
  - rotating-equipment
  - iso-10816
  - iso-20816
  - iso-17359
  - iso-13373
  - iso-14224
  - iso-4406
  - api-rp-670
  - api-rp-580
  - iso-18436
  - signal-processing
  - 3-axis-vibration
  - harmonic-analysis
pretty_name: OIL-040 — Synthetic Vibration & Sensor Dataset (Sample)
size_categories:
  - 100K<n<1M

OIL-040 — Synthetic Vibration & Sensor Dataset (Sample)

A schema-identical preview of OIL-040, the XpertSystems.ai synthetic vibration-and-sensor dataset for oil & gas rotating equipment condition monitoring. The full product covers ~15,000 assets across a 365-day horizon with 96-bin FFT spectra. This sample is a custom HF preview (80 assets × 30 days × 4 samples/day, 32-bin FFT) covering all 12 product tables, optimized for ISO 10816 / ISO 13373 / API RP 670 vibration analytics work.

Built by XpertSystems.ai — Synthetic Data Platform Contact pradeep@xpertsystems.ai · xpertsystems.ai License CC-BY-NC-4.0 (sample); commercial license available for the full product.

OIL-040 vs OIL-038 vs OIL-039 — what's different

OIL-038, OIL-039, and OIL-040 are three complementary upstream-asset PdM products covering different research workloads:

Dimension OIL-040 (this dataset) OIL-039 (PHM/RUL) OIL-038 (failure events)
Primary focus 3-axis vibration + FFT signal processing Per-timestamp RUL prognostics Failure-event analytics + reliability KPIs
3-axis vibration (X/Y/Z) Yes (horizontal-dominant) RMS only RMS only
FFT spectra 32-bin (sample), 96-bin (full) 4-band (sample), 7-band (full) None
ISO 10816 calibration Yes — median RMS in normal/alert band Different unit normalization ISO 10816 absolute units
Sensor pack tiers 4-tier (basic / standard / advanced / edge_ai) None None
Pre-built labels anomaly + fault_class + severity + rare_event + 30d target RUL hours/days + 7d/30d failure prob 30d/90d failure probability
Best for Signal-processing ML, FFT-based fault classification, ISO 10816 severity work RUL regression, prognostics benchmarks Reliability KPI fitting, MTBF/MTTR

Buy or download all three for complete PdM coverage. They share the upstream-asset, ISO 14224 / API RP 580 / API RP 670 calibration heritage.

What's inside

12 CSV tables covering 3-axis vibration + 5 supporting telemetry modalities

  • workorders + failures + per-record health/RUL/labels + 32-bin FFT spectra.
Table Rows (sample) What it represents
equipment_master.csv 80 12-type asset master with sensor pack, criticality, primary fault mode
vibration_timeseries.csv 9,600 3-axis (X/Y/Z) RMS mm/s + crest factor + kurtosis per timestamp
temperature_telemetry.csv 9,600 Thermocouple temperature, gradient, overheat flag
pressure_telemetry.csv 9,600 Pressure psi, delta, spike flag
acoustic_signals.csv 9,600 Acoustic dB, ultrasonic energy, anomaly score
lubrication_analysis.csv 9,600 Viscosity index, contamination, water ppm, lubrication risk
maintenance_workorders.csv ~80 6-type maintenance work orders with priority, downtime, notes quality
failure_events.csv ~5 Per-failure mode + severity + repair cost + production loss
health_scores.csv 9,600 Per-record health index, degradation score, condition state
remaining_useful_life.csv 9,600 Per-record RUL days, 30d failure probability, maintenance recommendation
vibration_labels.csv 9,600 Anomaly + fault class + severity + rare event flag + 30d target
fft_spectra.csv ~307,000 32-bin FFT (5–1000 Hz) with rotational harmonics + fault-defect frequency energy

Total: ~388,000 rows, ~40 MB. The full OIL-040 product is ~80 million rows with 96-bin FFT decomposition.

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:

  • ISO 10816 / ISO 20816 Mechanical vibration evaluation — vibration severity bands (normal / alert / alarm / shutdown) for Class II rotating equipment.
  • ISO 17359 Condition monitoring of machines — crest factor severity bands.
  • ISO 13373-1 / ISO 13373-2 Vibration condition monitoring — kurtosis, spectrum analysis.
  • ISO 18436-2 Vibration analyst certification conventions — horizontal / vertical / axial axis amplitude relationships.
  • API RP 670 Machinery Protection Systems — FFT decomposition standards and rotational harmonic boost relationships (1x, 2x, 3x, 4x rpm).
  • API RP 580 Risk-Based Inspection — criticality-tier distributions.
  • ISO 14224:2016 Reliability and Maintenance Data — equipment taxonomy and maintenance work classification.
  • ISO 4406:2021 Hydraulic fluid power: cleanliness code thresholds.
  • ARC Advisory PdM Maturity Survey + ISA-95 / OSDU — advanced sensor pack deployment baselines.
  • Noria Lubrication Practices — water content thresholds.

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 Vibration RMS median (mm/s) 1.5–4.5 2.38 ISO 10816 / ISO 20816
M02 Crest factor (mean) 2–6 3.30 ISO 17359 / ISO 13373-1
M03 Kurtosis (mean) 2–6 4.35 ISO 13373-1
M04 Lubrication water ppm (ceiling) ≤ 250 112 ISO 4406 / Noria
M05 Horizontal-axis dominance (X/RMS) 0.85–1.15 1.006 ISO 18436-2 / API RP 670
M06 Criticality tier ≥ 3 share 0.60–0.80 0.775 API RP 580 RBI
M07 Maintenance-type coverage (floor) ≥ 6 6 ISO 14224:2016
M08 FFT bin coverage (floor) ≥ 32 32 API RP 670 / ISO 13373-2
M09 Asset-type taxonomy (floor) ≥ 12 12 ISO 14224
M10 Advanced sensor pack share (floor) ≥ 0.25 0.388 ARC Advisory / OSDU

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

Suggested use cases

  • FFT-based fault classification — 32-bin FFT spectra include classic rotational harmonic peaks (1x, 2x, 3x, 4x rpm) and fault-specific defect frequencies (6.3x rpm bearing tone, 14x rpm gear-mesh tone, 420 Hz cavitation/surge band). Train CNN-on-spectrogram or 1D-conv classifiers.
  • ISO 10816 vibration severity classification — per-record RMS in mm/s is calibrated to the standard's Class II band, enabling direct alert / alarm / shutdown classifier training without unit conversion.
  • 3-axis anomaly detection — X/Y/Z axis decomposition with the classic horizontal-dominant ratio (X > Y > Z) makes this dataset suitable for geometry-aware anomaly models and axis-mixing experiments.
  • Crest factor + kurtosis impulsive fault detection — both metrics are ISO-calibrated and per-record, enabling bearing-fault and gear-mesh detection benchmarking against ISO 13373-1 thresholds.
  • Multi-modal sensor fusion — 6 telemetry modalities (vibration + temperature + pressure + acoustic + lubrication + health) are per-record_id-aligned for tight multi-modal experiments.
  • Sensor-pack tier ROI — sensor_pack field (basic / standard / advanced / edge_ai) on each asset enables ROI quantification of advanced PdM hardware against detection rate and failure cost.
  • Rare-event detection — rare_event_flag in vibration_labels.csv flags spike events (vibration ×1.7–3.8 multipliers) calibrated to fault-mode-dependent rates; useful for imbalanced-class ML training.
  • RUL regression — rul_days is per-record and calibrated against health index + failure probability; alternative to OIL-039's RUL bucket formulation.

Loading 

from datasets import load_dataset

master = load_dataset(
    "xpertsystems/oil040-sample",
    data_files="equipment_master.csv",
    split="train",
)
vibration = load_dataset(
    "xpertsystems/oil040-sample",
    data_files="vibration_timeseries.csv",
    split="train",
)
fft = load_dataset(
    "xpertsystems/oil040-sample",
    data_files="fft_spectra.csv",
    split="train",
)
labels = load_dataset(
    "xpertsystems/oil040-sample",
    data_files="vibration_labels.csv",
    split="train",
)

Or with pandas directly:

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="xpertsystems/oil040-sample",
    filename="fft_spectra.csv",
    repo_type="dataset",
)
df = pd.read_csv(path)

All 12 tables join on:

  • equipment_id → master ↔ all telemetry ↔ FFT ↔ workorders ↔ failures ↔ labels
  • record_id → tight per-timestamp join across all 6 telemetry modalities + labels + health + RUL
  • timestamp → temporal join across asset/record streams

The shared record_id makes multi-modal fusion experiments straightforward: join on record_id to get every modality at the same instant for the same asset.

Schema highlights

equipment_master.csv — equipment_id, facility_id, facility_type (7-class: upstream / offshore / midstream / refinery / lng / petrochemical / tank_farm), asset_type (12-class: centrifugal_pump / reciprocating_compressor / gas_turbine / steam_turbine / electric_motor / gearbox / pipeline_booster / drilling_mud_pump / lng_refrigeration_compressor / refinery_process_pump / blower / offshore_lift_motor), manufacturer (6-class), model, install_date, age_years, criticality ∈ {1, 2, 3, 4, 5}, rated_rpm, bearing_count, sensor_pack ∈ {basic, standard, advanced, edge_ai}, plus 4 baseline reference values per asset and primary_fault_mode (12-class).

vibration_timeseries.csv — record_id, equipment_id, timestamp, rpm, axis_x_mm_s, axis_y_mm_s, axis_z_mm_s (horizontal-dominant ISO 18436 convention), vibration_rms_mm_s (ISO 10816 unit), crest_factor (ISO 17359), kurtosis (ISO 13373-1).

fft_spectra.csv — Per-record × 32-bin FFT decomposition (5–1000 Hz linear), each row has frequency_hz, amplitude, phase_angle. Amplitude includes base lognormal noise + rotational harmonic boosts at {1x, 2x, 3x, 4x} rpm/60 + fault-defect frequency boosts: bearing tone at 6.3x rpm (bearing_wear, lubrication_loss), gear-mesh tone at 14x rpm (gear_mesh_wear), and 420 Hz cavitation/surge band.

vibration_labels.csv — record_id, equipment_id, timestamp, anomaly_label (binary), fault_class (12-class), severity_level (5-class: normal / low / medium / high / critical), rare_event_flag, target_failure_30d.

maintenance_workorders.csv — maintenance_type (6-class: inspection / lubrication / bearing_replacement / alignment / sensor_calibration / overhaul), priority (4-class: low / medium / high / emergency), downtime_hours, technician_notes_quality ∈ {complete, partial, missing}.

Calibration notes & limitations

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

  1. Custom HF preview sizing. The default generator sample mode produces ~326 MB (250 assets × 30 days × 24 samples/day × 32 FFT bins = ~1.9M FFT rows). The HF preview is reduced to 80 assets × 30 days × 4 samples/day to stay under 50 MB while preserving every table, schema, and the scorecard's industry-anchored calibration validity. For higher time-density studies, override sizing with the underlying generator's --samples-per-day and --n-assets flags, or use the commercial full product.

  2. Anomaly label rate is ~99%. In vibration_labels.csv, the anomaly_label (binary) is set to 1 whenever condition_state != normal, and the severity-label thresholds combined with the fail_prob distribution put 99% of records in low/medium/high/critical bands. This is a labeling-convention artifact, not a positive-class density claim. For binary anomaly classification work, use severity_level directly (5-class) or threshold failure_probability_30d > 0.70 to recover a balanced positive class (5% of records). The full product ships a threshold-tuned binary label variant.

  3. Overheat flag is 0 in the sample. temperature_telemetry.csv's overheat_flag triggers above 115°C, but at the 30-day window most assets don't reach that threshold. For overheat-detection studies, lower the threshold to 95°C in your downstream pipeline, or use the full product's 365-day window which exposes more thermal-overload events.

  4. Only 9 of 12 fault modes appear at sample scale. With 80 assets and 55% normal primary fault, only 9 of the 12 fault modes are represented in any given seed's sample. For full taxonomy coverage, use multiple seeds and concatenate, or use the full product (15,000 assets sees all 12 fault modes with statistically representative density).

  5. Small failure-event count. With 80 assets × 30 days, the sample produces ~5–10 failure events depending on seed. Failure-severity distributions are not reliably estimable at this scale (small-sample variance). For severity-pyramid analytics, use OIL-038 (rich failure-event tables) or OIL-039 (sigmoid-calibrated 7d/30d probabilities).

  6. FFT amplitude scale. FFT amplitudes are in normalized units derived from vibration_rms_mm_s × harmonic-boost factors. They are NOT absolute G or m/s². For absolute-unit FFT work, calibrate against the vibration_rms_mm_s baseline.

  7. Deterministic seeding. All 12 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-040 product covers 15,000 assets × 365 days × 24 samples/day × 96-bin FFT decomposition (80 million rows total), with threshold-tuned binary anomaly labels, full 12-fault-mode coverage at production scale, and longer-horizon thermal-overload events. Available under commercial license — contact 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.