--- license: cc-by-nc-4.0 task_categories: - tabular-classification - tabular-regression - time-series-forecasting language: - en tags: - synthetic - oil-and-gas - upstream - iot - rig-telemetry - predictive-maintenance - condition-monitoring - remaining-useful-life - sensor-fusion - anomaly-detection - xpertsystems pretty_name: "OIL-012 — Synthetic Rig Sensor IoT Dataset (Sample)" size_categories: - 100K ROP = 35 + 1.2·WOB + 0.08·RPM − 0.0018·Torque + N(0, 8) > Torque = 5200 + 75·WOB + 18·RPM + 750·sin(t) + N(0, 850) This produces strong physics coupling: **WOB↔ROP Pearson r ≈ 0.92, WOB↔Torque r ≈ 0.87, RPM↔Torque r ≈ 0.82** — matching the Teale/Pessier MSE physics. **`vibration_analysis.csv`** — 3-axis vibration with **stick-slip-coupled torsional amplification**: > vibration_rms = 0.22 + 0.0035·RPM + 0.16·stick_slip + 0.25·wear + N(0, 0.08) > torsional_vib = vibration_rms × (0.75 + 0.8·stick_slip) + noise Per **API 670** machinery-protection-system thresholds, the alarm zone starts at ~1.8 g RMS (typical for drilling-equipment Class B-C classification per ISO 10816). Sample mean ~0.58 g sits comfortably in the Class A "good" zone. **`power_systems.csv`** — three-phase rig power with motor-load-coupled voltage sag: > voltage = N(480, 12) − 0.25·max(motor_load − 80, 0) > amperage = 80 + 2.1·motor_load + N(0, 25) Voltage centers on **480 V** (US industrial standard 3-phase) with sag under high motor load — matches IEEE 141 power-quality conventions. **`thermal_monitoring.csv`** — bearing temperature coupled to motor load, vibration, and wear: > bearing_temp = 145 + 0.17·motor_load + 25·vibration_rms + 18·wear + N(0, 6) Per **API 670** §4.5, bearing alarm threshold is 210°F (99°C) and trip at 250°F (121°C). Sample mean ~175°F is well below alarm — realistic for properly-maintained rotating equipment. **`rul_labels.csv`** — remaining useful life and failure probability computed from a **sigmoid risk model**: > risk_logit = −4.4 + 2.4·vibration_rms + 0.032·max(bearing_temp − 190, 0) > + 0.018·max(motor_load − 75, 0) + 2.2·stick_slip > + 1.5·env_stress + 1.1·(1 − health) + 0.8·wear > failure_probability = sigmoid(risk_logit) > RUL_hours = (1 − failure_probability) × 1500 − 450·wear + N(0, 80) Conforms to **ISO 13374 (Condition Monitoring & Diagnostics)** data architecture: features → state detection → diagnosis → prognosis. **`sensor_fusion_features.csv`** — multi-sensor synthesis for ML serving layer: > maintenance_priority = 0.5·failure_prob + 0.3·anomaly_score > + 0.2·(1 − subsystem_health) Designed as **ML-ready features** for downstream condition-based maintenance dashboards. **`telemetry_streams.csv`** — **sparse event-stream format** (one row per sensor reading), filtered by per-rig packet-loss model. This is the realistic SCADA/edge-gateway pattern (vs. the dense matrix in the per- subsystem tables above). At packet_loss ~0.3% sample-wide, ~99.7% of sensor readings are emitted. --- ## Suggested use cases 1. **Remaining Useful Life regression** — predict `remaining_useful_life_hr` from the dense per-subsystem feature tables (drilling + vibration + thermal + power). Standard PHM/RUL benchmark target. 2. **Sigmoid failure-probability classification** — binary or probabilistic classifier on `failure_probability` (threshold > 0.5) from upstream features. 3. **3-class severity classification** — multi-class classifier on `severity` (medium/high/critical) from alarm-event features for alarm prioritization. 4. **Stick-slip detection** — regress `stick_slip_index` from drilling parameters (WOB/RPM/torque). Strong torque-coupled signal. 5. **Anomaly detection on telemetry streams** — autoencoder / isolation-forest training on `telemetry_streams.csv` against `anomaly_score` ground truth in `sensor_fusion_features.csv`. 6. **Sensor calibration drift detection** — regress `calibration_score` decay over time from `sensor_health.csv`. 7. **Operating state classification (4-class)** — classifier on `operating_state` (drilling/tripping/circulating/maintenance) from drilling parameters and motor load. 8. **Predictive maintenance scheduling** — sequence-to-sequence prediction of maintenance windows from `maintenance_priority_score` time series. 9. **Multi-table relational ML** — entity-resolution and graph neural-network learning across the 15 joinable tables via `rig_id` + `event_ts`. --- ## Loading ```python from datasets import load_dataset ds = load_dataset("xpertsystems/oil012-sample", data_files="drilling_parameters.csv") print(ds["train"][0]) ``` Or with pandas: ```python import pandas as pd rigs = pd.read_csv("hf://datasets/xpertsystems/oil012-sample/rigs_master.csv") drilling = pd.read_csv("hf://datasets/xpertsystems/oil012-sample/drilling_parameters.csv") vibration = pd.read_csv("hf://datasets/xpertsystems/oil012-sample/vibration_analysis.csv") rul = pd.read_csv("hf://datasets/xpertsystems/oil012-sample/rul_labels.csv") joined = drilling.merge(vibration, on=["event_ts", "rig_id"]).merge(rul, on=["event_ts", "rig_id"]) ``` --- ## Reproducibility All generation is deterministic via the integer `seed` parameter (driving both `random.seed` and `np.random.default_rng`). A seed sweep across `[42, 7, 123, 2024, 99, 1]` confirms Grade A+ on every seed in this sample. --- ## Honest disclosure of sample-scale limitations This is a **sample** product calibrated for IoT and predictive-maintenance ML research, not for live rig operations. A few notes: 1. **The `equipment_failures.csv` table is empty in the sample run** (and sparse in the full product). This is **by design**: failures fire only when `failure_probability > 0.86`, which requires concurrent high-vibration + hot-bearing + high-stick-slip + degraded-health conditions — i.e., genuinely pathological operation. At sample scale (30 rigs × 1 day) such conditions almost never occur, matching real industry data (modern drilling rigs do not fail daily). For ML training on failure-mode classification, the `alarm_events.csv` table (which uses a lower 0.72 threshold) provides the positive examples; for RUL regression, `rul_labels.csv` provides continuous targets. The full product (15K rigs × 7 days × 60s) generates a substantial failures table. 2. **`shock_event_flag` is 0 throughout the sample** in `vibration_analysis.csv` for the same reason — the threshold of `vibration_rms > 1.85 g` (API 670 alarm zone) is not reached at sample scale. 3. **All telemetry rows are quality_flag="good"** because `calibration_score` stays above 0.86 at sample scale (slow wear model). Full-product runs with `--days 30+` produce degraded-quality examples for calibration-drift ML. 4. **Telemetry stream is sparse-formatted**, one row per sensor reading. This means joining telemetry to dense per-subsystem tables requires a pivot/aggregate first. For dense time-series training, use the per-subsystem tables directly; for sensor-fusion / multi-stream training, the sparse format is canonical. 5. **3 of 8 rig types are underrepresented at sample scale**: jackup (3% of rigs), unconventional_shale (7%), land_drilling (10%). With only 30 rigs in a 8-class uniform draw, each class is sampled ~4 times in expectation. The full product (15K rigs) gives clean per-class statistics. 6. **`maintenance_priority_score` distribution is calibrated for workflow triage**, not ground-truth failure prediction. It's a weighted average of failure_probability, anomaly_score, and (1 − subsystem_health) — useful as an ML feature target, but should not be confused with actual maintenance scheduling decisions. 7. **Packet loss is ~0.3%** in the sample (per `--packet-loss-rate 0.0018` default). Real SCADA / edge-gateway systems see 0.5-2% loss; adjust `--packet-loss-rate` higher if training models that need realistic gap-handling. --- ## Full product The **full OIL-012 dataset** ships at **15,000 rigs × 7 days × 60s** (prod mode) producing several billion telemetry rows with substantial populated `equipment_failures` and `shock_event_flag` tables, full calibration-drift histories, and basin-conditioned operator behavior priors — licensed commercially. Contact XpertSystems.ai for licensing terms. 📧 **pradeep@xpertsystems.ai** 🌐 **https://xpertsystems.ai** --- ## Citation ```bibtex @dataset{xpertsystems_oil012_sample_2026, title = {OIL-012: Synthetic Rig Sensor IoT Dataset (Sample)}, author = {XpertSystems.ai}, year = {2026}, url = {https://huggingface.co/datasets/xpertsystems/oil012-sample} } ``` ## Generation details - Sample version : 1.0.0 - Random seed : 42 - Generated : 2026-05-22 12:28:49 UTC - Rigs : 30 - Days simulated : 1 - Telemetry freq : 600s (144 timesteps per rig) - Rig types : 8 (land, offshore deepwater, jackup, arctic, MPD, HPHT, automated smart, unconventional shale) - Basins : 10 (Permian, Eagle Ford, Bakken, GoM, North Sea, Middle East, Brazil Pre-Salt, Marcellus, W Canada, N Africa) - Subsystems : 10 (top drive, mud pump, rotary table, drawworks, BOP, power system, hydraulic system, hoisting system, drillstring, compressor) - Sensor types : 18 (with redundancy on 4 critical channels = 22/rig) - Operating states : 4 (drilling, tripping, circulating, maintenance) - Calibration basis : API 670, API 16D, API RP-7G, API RP-13B-1, API RP-541, Teale (1965), SPE 21943, SPE 178850, ISA-18.2, EEMUA 191, IEEE 141, ISO 10816, ISO 13374, Rystad, Spears, IHS Markit - Overall validation: 100.0/100 — Grade A+