xpertsystems/oil044-sample

OIL-044 — Synthetic Autonomous Oilfield Dataset (Sample)

A schema-identical preview of OIL-044, the XpertSystems.ai synthetic autonomous oilfield operations dataset for autonomous robotics, drone inspection, edge AI inference, remote operations, predictive maintenance, and human-in-the-loop decision-support AI training. The full product covers 5,000 assets / 500,000 telemetry rows / 100,000 robotic operations / 50,000 edge AI decisions. This sample is HF-sized (500 assets, ~75K rows total) covering all 8 product tables.

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.

---

What OIL-044 does that nothing else in the catalog does

OIL-044 is the catalog's first autonomous-systems / edge AI SKU. Where OIL-042 (Digital Twin) models steady-state operations and OIL-043 (Scenario Simulation) models perturbations, OIL-044 models the autonomous decision-making layer that sits on top: SAE-J3016-style autonomy-level assets, ISO 10218 / ISO 13482 robotic operations, drone inspection missions, edge AI inference with calibrated confidence and human-override flags, and remote-operator sessions across healthy / degraded / partitioned network conditions.

This is the substrate that autonomous-systems researchers, robotics SaaS vendors, edge AI platform teams, and human-AI interaction researchers have been waiting for: a coherent, joinable dataset where robotic ops, drone inspections, edge AI decisions, and human overrides share asset_id and timestamps for cross-layer correlation research.

Buyer Persona	Use Case
Robotics SaaS Vendor	Robotic operations success modeling, fleet analytics
Edge AI Platform Team	Confidence calibration, human-override-trigger learning
Autonomous Systems Researcher	SAE J3016 autonomy-level performance benchmarking
Human-AI Interaction Researcher	Override decision modeling, network-conditional autonomy
Industrial IoT Vendor	OT network health × autonomous-decision correlation
Drone Inspection Vendor	Anomaly detection rate calibration across drone types
C-suite AI Demo	"Show me autonomous oilfield AI in 60 seconds"

---

What's inside

8 CSV tables organized around an asset_id master key: autonomous asset inventory → robotic operations → drone inspection missions → edge AI decisions → predictive maintenance → remote-operator sessions → sensor telemetry → pre-built ML labels.

Table	Rows (sample)	What it represents
autonomous_assets.csv	500	6-class asset taxonomy × 5-tier autonomy level × 3-state operational status
robotic_operations.csv	10,000	5-class robot × 6-class task × 3-class execution status + battery + failure prob
drone_missions.csv	3,000	3-class drone × mission type × flight duration + anomaly detected + collision risk
edge_ai_decisions.csv	6,000	5-class decision × confidence score + inference latency + human override flag
predictive_maintenance.csv	4,000	Degradation + failure probability + remaining days + action taken
remote_control_sessions.csv	2,500	Latency ms + 3-state network health + commands sent per session
equipment_telemetry.csv	50,000	5-class sensor (pressure/temp/flow/vibration/RPM) with calibrated value distributions
autonomous_labels.csv	500	Pre-built ML labels: autonomy risk score + intervention probability + 4-tier criticality

Total: ~76,000 rows, ~9 MB. The full OIL-044 product is ~700K rows.

---

Calibration sources

Every distribution and ratio is anchored to named public references. Highlights:

• NIST AI-RMF 1.0 (NIST AI 100-1) + ISO/IEC TR 24028 — autonomous-AI confidence calibration conventions.
• ISO 22989 AI concepts and terminology — autonomous decision and human-oversight conventions.
• NIST SP 1011 Autonomy Levels for Unmanned Systems + NIST Robotic Systems Test Methods — autonomous-mission success benchmarks.
• ISO 10218 Industrial robots — failure-probability and safety norms.
• ISO 13482 Service robots — autonomous-machinery safety conventions.
• ISO 10816 / ISO 20816 Mechanical vibration evaluation — vibration sensor severity bands.
• IEC 62443 Industrial network security — OT network health KPIs.
• 3GPP Industrial IoT KPIs — remote-link availability conventions.
• SAE J3016 Levels of driving automation (5-level scale) — applied analogously here to industrial asset autonomy.
• AUVSI + FAA Part 107 + Oil & Gas drone-inspection survey data — 3-class drone taxonomy.
• ISA-95 / OPC UA + ISO 14224 — sensor classification.

---

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	Asset-Type Taxonomy (floor)	≥ 6	6	XpertSystems autonomous-oilfield
M02	Robot-Type Taxonomy (floor)	≥ 5	5	ISO 10218 / ISO 13482 / IADC
M03	Drone-Type Taxonomy (floor)	≥ 3	3	AUVSI / FAA Part 107
M04	Edge AI Decision-Type (floor)	≥ 5	5	NIST AI-RMF / ISO 22989
M05	Sensor-Type Taxonomy (floor)	≥ 5	5	ISA-95 / OPC UA / ISO 14224
M06	Vibration Sensor Mean (mm/s)	3.5–5.5	4.49	ISO 10816 Class III
M07	Robotic Success Rate (floor)	≥ 0.87	0.919	NIST SP 1011
M08	Network Healthy Rate (floor)	≥ 0.87	0.909	IEC 62443 / 3GPP IIoT
M09	Edge AI Confidence (mean)	0.89–0.95	0.917	NIST AI-RMF / ISO 24028
M10	Human-Override Rate	0.03–0.07	0.053	ISO 22989 / NIST AI-RMF

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

The scorecard intentionally focuses on NIST AI-RMF / ISO 22989 calibration anchors — autonomous-systems standards where the synthetic data must faithfully represent the standard's defined ranges to be useful for regulator-compliant decision-support AI training.

---

Suggested use cases

• Human-override prediction modeling — edge_ai_decisions.csv has per-decision human_override_required binary flag plus confidence_score, enabling training of override-trigger models for NIST AI-RMF Govern-1 human-oversight workflows.
• Confidence calibration research — edge_ai_decisions.csv confidence distribution (mean 0.92, calibrated against NIST 100-1) is ground truth for calibration-error studies, Platt scaling, and isotonic regression benchmarking.
• Autonomous-mission success classification — robotic_operations.csv 3-class status (SUCCESS / PARTIAL_SUCCESS / FAILED) × battery level × task duration × failure probability. Train mission-success classifiers with 92% positive class.
• Drone-inspection anomaly detection — drone_missions.csv provides anomaly_detected binary flag × drone type × collision risk score, suitable for AUVSI/FAA-aligned inspection-quality ML.
• Network-conditional autonomy modeling — remote_control_sessions.csv × edge_ai_decisions.csv joinable on asset_id supports network-aware human-AI handoff research (when does the link quality justify falling back to edge AI vs human operator).
• Telemetry-driven predictive maintenance — equipment_telemetry.csv has calibrated normal distributions per sensor type, joinable with predictive_maintenance.csv for degradation modeling.
• Pre-built autonomy-risk ML labels — autonomous_labels.csv provides asset-level autonomy_risk_score, intervention_probability, and 4-tier criticality_level, ready for downstream regression or multi-class classification.
• Cross-vertical autonomous-systems methodology — OIL-044 schemas apply analogously to manufacturing, mining, ports, and warehouse autonomy research; buyers can use the same data plane for non-O&G autonomous research.

---

Loading

from datasets import load_dataset

assets = load_dataset(
    "xpertsystems/oil044-sample",
    data_files="autonomous_assets.csv",
    split="train",
)
edge_ai = load_dataset(
    "xpertsystems/oil044-sample",
    data_files="edge_ai_decisions.csv",
    split="train",
)
labels = load_dataset(
    "xpertsystems/oil044-sample",
    data_files="autonomous_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/oil044-sample",
    filename="robotic_operations.csv",
    repo_type="dataset",
)
df = pd.read_csv(path)

All 8 tables share asset_id as the master join key, supporting cross-table joins:

• autonomous_assets ⨝ autonomous_labels (1:1) — every asset has labels
• autonomous_assets ⨝ robotic_operations (1:N) — ~20 operations per asset
• autonomous_assets ⨝ drone_missions (1:N) — ~6 missions per asset
• autonomous_assets ⨝ edge_ai_decisions (1:N) — ~12 decisions per asset
• autonomous_assets ⨝ equipment_telemetry (1:N) — ~100 telemetry rows per asset

---

Schema highlights

autonomous_assets.csv — asset_id, asset_type (6-class: DRILLING_RIG / PIPELINE_STATION / REFINERY_UNIT / COMPRESSOR / LNG_TERMINAL / OFFSHORE_PLATFORM), autonomy_level ∈ {1, 2, 3, 4, 5} (SAE J3016-like 5-tier scale: 1=Driver Assistance, 5=Full Automation), operational_status ∈ {ACTIVE, STANDBY, MAINTENANCE}, location_lat, location_lon.

robotic_operations.csv — operation_id, robot_id, asset_id, robot_type (5-class: INSPECTION_ROBOT / PIPE_CRAWLER / AUTONOMOUS_TRUCK / ROBOTIC_ARM / VALVE_CONTROLLER), task_type (6-class), execution_status ∈ {SUCCESS, FAILED, PARTIAL_SUCCESS}, task_duration_minutes, battery_level, failure_probability, timestamp.

drone_missions.csv — drone_type (3-class: THERMAL_DRONE / VISUAL_INSPECTION_DRONE / LEAK_DETECTION_DRONE), mission_type, flight_duration_minutes, anomaly_detected (binary), collision_risk_score.

edge_ai_decisions.csv — decision_type (5-class: SHUTDOWN / CONTINUE_OPERATION / ESCALATE / DISPATCH_ROBOT / REQUEST_HUMAN_OVERRIDE), confidence_score ∈ [0, 1] (calibrated to NIST AI-RMF norms), inference_latency_ms, human_override_required (binary), timestamp.

remote_control_sessions.csv — latency_ms (calibrated to satellite + terrestrial Edge mix), network_health ∈ {HEALTHY, DEGRADED, PARTITIONED}, commands_sent, timestamp.

equipment_telemetry.csv — 5-class sensor_type with calibrated normal distributions:

• PRESSURE: mean 1,200 psi, σ 80
• TEMPERATURE: mean 85 °C, σ 12
• FLOW_RATE: mean 420 bpd, σ 55
• VIBRATION: mean 4.5 mm/s, σ 1.1 (ISO 10816)
• RPM: mean 1,800 RPM, σ 250

autonomous_labels.csv — pre-built ML labels: autonomy_risk_score ∈ [0, 1], intervention_probability ∈ [0, 1], criticality_level ∈ {LOW, MEDIUM, HIGH, CRITICAL}.

---

Calibration notes & limitations

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

1. Edge AI inference latency uses uniform random 10–3,000 ms with median ~1,500 ms. This is much slower than industry "edge AI" (<100 ms for real-time control). The synthetic distribution is intentionally wide for ML training utility — it covers the full range from on-asset edge inference (<100 ms) through fog computing (200–500 ms) through cloud-fallback (1–3 s). For pure on-device edge AI work, filter to inference_latency_ms < 100 to recover a real-time-control subset. The scorecard validates confidence and override rates instead of latency for this reason.

2. Predictive maintenance action distribution is uniform 25% each across NONE / INSPECTION / PART_REPLACEMENT / EMERGENCY_SHUTDOWN. Industry mature operations sustain EMERGENCY_SHUTDOWN at ≤5%. This uniform distribution is intentional for ML training utility (balanced multi-class target). For realistic action-distribution work, threshold from degradation_score with custom mapping (e.g., NONE when score < 0.3, INSPECTION 0.3–0.5, PART_REPLACEMENT 0.5–0.7, EMERGENCY_SHUTDOWN > 0.7), or use the OIL-038 / OIL-039 PdM SKUs which carry calibrated action distributions.

3. Operational status uniform across ACTIVE / STANDBY / MAINTENANCE (~33% each). Industry mature is 70–85% ACTIVE. This is by design to give all 3 status classes equal ML training density at sample scale.

4. Criticality level uniform across LOW / MEDIUM / HIGH / CRITICAL (~25% each). Industry mature criticality distributions are pyramid- shaped (most LOW). The uniform distribution gives balanced multi-class training; for pyramid-shaped sampling, threshold from autonomy_risk_score directly.

5. HF preview sizing — default generator sizing is 5K assets / 500K telemetry / 100K robotic operations producing ~150 MB. The HF preview is reduced to 500 assets / 50K telemetry / 10K robotic operations, ~9 MB. All schemas, taxonomies, and scorecard calibrations are preserved at the smaller scale. For higher-density studies, override the underlying generator's --n-assets / --n-telemetry-rows flags.

6. Sensor telemetry uses simple Gaussian per sensor type. There's no cross-modal coupling, no temporal autocorrelation, no degradation trajectory linkage. For multi-modal anomaly detection with realistic covariance, use OIL-038 (16-modality with degradation trajectories) or OIL-039 (RUL prognostics with sigmoid-calibrated degradation). OIL-044 is optimized for autonomous-decision and human-oversight research, not multi-modal anomaly detection.

7. Battery level uniform 5–100%. Real fleet battery distributions are bimodal (charging stations + active duty). For realistic battery analytics, use the full product or condition on execution_status (low battery is correlated with FAILED status in the full generator).

8. Deterministic seeding. All 8 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-044 product covers 5,000 assets × 100,000 robotic operations × 50,000 edge AI decisions × 500,000 telemetry rows across a 1-year horizon (700K rows total), with calibrated industry-pyramid distributions for operational status, criticality, and predictive maintenance actions, plus realistic battery / fleet bimodal distributions and edge-vs-cloud latency segmentation. 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.