xpertsystems/oil022-sample

OIL-022 — Synthetic Shutdown & Turnaround Dataset (Sample)

SKU: OIL022-SAMPLE · Vertical: Oil & Gas / Downstream Refining License: CC-BY-NC-4.0 (sample) · Schema version: oil022.v1 Sample version: 1.0.0 · Default seed: 42

A free, schema-identical preview of XpertSystems.ai's enterprise refinery shutdown & turnaround dataset for maintenance planning ML, inspection optimization, schedule slippage prediction, restart readiness assessment, turnaround cost forecasting, and RBI (risk-based inspection) analytics. The sample covers 1,200 turnaround campaigns across 15 refineries with 2,250 pieces of equipment in 10 global regions, with 165,114 rows linked across 15 tables.

OIL-022 is the third downstream (refining) SKU in the catalog, complementing OIL-019 (refinery process operations) and OIL-020 (product yields + economics) with maintenance, inspection, and turnaround operations specialization.

---

What's in the box

File	Rows	Cols	Description
refineries_master.csv	15	6	Refinery catalog: 10 regions × 4 operator types × Nelson complexity × capacity × PSM maturity
equipment_master.csv	2,250	11	Equipment inventory: 14 classes × 14 units × 5 material families × 4 RBI categories per API 580
shutdown_campaigns.csv	1,200	12	Turnaround campaigns: 5 shutdown types × planned/actual duration × schedule slippage × scope complexity
corrosion_monitoring.csv	11,250	9	Per-equipment 5-point time-series: UT/RT/Guided Wave measurements + 10-class corrosion mechanisms per API 570 + NACE
maintenance_work_orders.csv	39,048	12	Per-campaign WOs: 12 maintenance types × 4 priorities × 4 statuses × QA/QC flags
inspection_findings.csv	21,479	11	API 510 RBI findings: wall thickness, corrosion rate, remaining life, anomaly score per API 580/581
turnaround_schedule.csv	36,000	9	Critical path tasks: predecessor logic, 10 craft types, planned hours, schedule risk score
workforce_allocation.csv	24,000	8	Contractor allocations: 350 contractors × 10 craft types × shift hours × fatigue risk
permit_to_work.csv	15,771	8	OSHA 1910.119 PSM permits: 7 permit types × 4 hazard levels × isolation/gas test/approval delay
equipment_failures.csv	1,031	7	10 failure modes × 7 root causes × downtime + startup-detection flag
catalyst_replacement.csv	1,045	7	Reactor catalyst events: 6 catalyst types × activity % × age days × replacement cost
startup_readiness.csv	9,600	6	8-step startup readiness per CCPS: Mechanical Completion → Stability Test + risk scores
turnaround_costs.csv	1,200	7	Per-campaign cost breakdown: labor + material + delay + contractor + total
safety_events.csv	25	7	7-class CCPS events: near miss, first aid, recordable, lost time + severity + corrective action days
shutdown_labels.csv	1,200	9	FEATURE-COUPLED ML labels: 4-class reliability grade (A/B/C/D) + restart success + cost overrun + completion %

Total: 165,114 rows across 15 CSVs, ~12.5 MB on disk.

---

Calibration: industry-anchored, honestly reported

Validation uses a 10-metric scorecard with targets sourced exclusively to named industry standards: API 510 (Pressure Vessel Inspection Code), API 570 (Piping Inspection Code), API 580/581 (Risk-Based Inspection), NACE TM0274 (corrosion measurement), OSHA 29 CFR 1910.119 (Process Safety Management — PSM), AFPM Reliability and Maintenance Benchmarking Survey, Solomon Associates Refinery Performance Survey, IPA (Independent Project Analysis) Turnaround Cost Performance Database, OGCI turnaround safety statistics, ANSI/AICHE CCPS (Center for Chemical Process Safety) guidelines, EIA-820 Refinery Capacity Report, Nelson Complexity Index (Oil & Gas Journal).

Sample run (seed 42, n_turnarounds=1,200, refineries=15):

#	Metric	Observed	Target	Tolerance	Status	Source
1	avg refinery capacity bpd	217625.3333	220000.0	±80000.0	✓ PASS	EIA-820 Refinery Capacity Report — mean capacity for mixed global refinery portfolio (US median ~135K BPD, largest US refineries 600K+ BPD, Indian/Chinese mega-refineries 400-1200K BPD; portfolio mean ~220K BPD)
2	avg complexity index	9.6007	9.5	±2.0	✓ PASS	Nelson Complexity Index (Oil & Gas Journal) + Solomon Associates Refinery Performance Survey — mean Nelson Complexity for global refinery portfolio (simple hydroskimmers 4-6, modern conversion refineries 9-12, deep-conversion / petrochemical 12-16)
3	avg planned duration days	23.9475	26.0	±6.0	✓ PASS	AFPM Reliability and Maintenance Benchmarking Survey + Solomon Associates — mean planned turnaround duration for mixed scope portfolio (pitstop 10d, planned TA 26d, major TA 42d; portfolio mean ~26d weighted by frequency)
4	avg schedule slippage pct	12.6093	11.5	±5.0	✓ PASS	IPA (Independent Project Analysis) Turnaround Cost Performance Database + AFPM — mean schedule slippage across refinery turnaround portfolio (8-15% typical for well-planned, 20%+ indicates poor planning per IPA benchmarks)
5	avg corrosion rate mpy	6.5791	5.0	±3.0	✓ PASS	API 570 (Piping Inspection Code) + NACE TM0274 — mean corrosion rate for refinery piping portfolio (2-8 mpy normal for moderate service; >10 mpy triggers RBI high-risk classification per API 581)
6	avg work order completion pct	89.9395	90.0	±5.0	✓ PASS	AFPM Reliability and Maintenance Benchmarking Survey — mean work order completion rate during refinery turnarounds (85-95% typical; >95% indicates either conservative scoping or schedule pressure)
7	anomaly flag rate	0.0321	0.032	±0.015	✓ PASS	ANSI/AICHE CCPS process safety management + AFPM operational data — typical anomaly/deviation rate for refinery work order execution (2-5% of WOs exhibit execution anomalies per CCPS safety reporting)
8	slippage reliability pearson correlation	-0.6386	-0.55	±0.15	✓ PASS	IPA Turnaround Cost Performance + AFPM — expected strong inverse correlation between schedule slippage and reliability grade score (generator formula: reliability_score = 100 - slippage0.7 - findings0.8 - completion_penalty*1.2). Validates feature-coupled label generation.
9	corrosion remaining life pearson correlation	-0.5538	-0.5	±0.15	✓ PASS	API 510 + API 580/581 (Risk-Based Inspection) — expected inverse correlation between corrosion rate and remaining life (RBI formula: remaining_life = (wall_thickness - retirement_limit) / corrosion_rate). Validates generator's API 510 RBI physics.
10	equipment class diversity entropy	0.9986	0.97	±0.03	✓ PASS	API 580 RBI equipment classification + Solomon Associates equipment census — 14-class equipment diversity benchmark covering pressure vessels, heat exchangers, fired heaters, compressors, pumps, piping, tanks, valves, reactors, columns, boilers, cooling towers, instrument loops, normalized Shannon entropy

Overall: 100.0/100 — Grade A+ (10 PASS · 0 MARGINAL · 0 FAIL of 10 metrics)

---

Schema highlights

refineries_master.csv — 10-region global refinery portfolio with Nelson Complexity Index:

Region	Complexity Notes
US Gulf Coast	High complexity (11-13) — deep conversion + petrochemical
US West Coast	Moderate (8-10) — fluid catalytic cracking
North Sea	Moderate-high (10-12) — sour crude processing
Middle East	Mixed (6-14) — both simple export + deep conversion
India / China	Mega-refineries (10-15) — petrochemical integrated
Brazil	Moderate (8-11) — heavy/sour crude
North/West Africa	Lower (5-9) — export-grade simple refineries
SE Asia	Moderate (8-12)
Western Europe	Higher (10-13) — declining capacity, high specs

shutdown_campaigns.csv — 5 shutdown types per AFPM nomenclature:

Type	Weight	Base Duration
Planned Turnaround	58%	26d
Major Turnaround	16%	42d
Pitstop	14%	10d
Emergency Shutdown	6%	7d
Regulatory Outage	6%	18d

Schedule slippage applied stochastically: actual_duration = planned × (1 + slippage/100) with ~3.5% anomaly rate adding 12-45 extra slippage percentage points.

inspection_findings.csv — API 510 RBI physics implemented:

wall_thickness = nominal − (age × corrosion_rate × 0.0254/2.0) + noise [API 570 form] remaining_life = (wall_thickness − retirement_limit) / corrosion_rate [API 580/581] anomaly_score = f(metal_loss, criticality) clip(0, 1) repair_required_flag = (anomaly_score > 0.68) [API 510 trigger]

The sample's corrosion-rate↔remaining-life Pearson correlation is r ≈ −0.55 — strong inverse coupling validates API 510 RBI physics (higher corrosion rate → shorter remaining life).

maintenance_work_orders.csv — slippage-coupled WO execution:

est_hours = lognormal(2.65, 0.7) # typical 14-23 hours overrun_factor = N(1 + slippage/180, 0.22) # slippage drives overrun actual_hours = est_hours × overrun_factor

WO status distribution per AFPM benchmark: 90% Completed / 5% Deferred / 4% In Progress / 1% Cancelled.

permit_to_work.csv — OSHA 29 CFR 1910.119 PSM permit types:

Type	Notes
Hot Work	Welding/cutting requires gas test
Confined Space	Vessel entry requires gas test
Line Break	Piping isolation requires gas test
Electrical Isolation	LOTO
Working at Height	Fall protection
Cold Work	Routine maintenance
Excavation	Underground services

shutdown_labels.csv — FEATURE-COUPLED ML labels (unlike OIL-019/020 pure-random labels):

reliability_score = 100 − slippage × 0.7 − high_risk_findings × 0.8 − max(0, 95 − completion_pct) × 1.2 reliability_grade = 'A' if score ≥ 90 else 'B' if ≥ 80 else 'C' if ≥ 70 else 'D' restart_success = (ready_count ≥ 7) AND (reliability_score > 72) AND (rng > risk × 0.12)

The slippage↔reliability Pearson correlation is r ≈ −0.64 in the sample — strong inverse coupling validates feature-coupled labels per IPA/AFPM turnaround performance benchmarks.

---

Suggested use cases

1. API 510 remaining life regression — predict remaining_life_years from corrosion rate + wall thickness + criticality features. Strong physics signal: corrosion-life inverse r ≈ −0.55.
2. Reliability grade classification — 4-class ordinal classifier on reliability_grade (A/B/C/D) from slippage + findings + completion features. Strong feature coupling — models WILL learn meaningful patterns (unlike OIL-019/020 pure-random labels).
3. Schedule slippage regression — predict schedule_slippage_pct from scope_complexity + shutdown_type + equipment criticality features per IPA turnaround benchmark.
4. Anomaly score regression — predict anomaly_score from corrosion rate + wall loss + criticality per API 580/581 RBI.
5. Restart success binary classification — predict restart_success_flag from readiness + reliability + risk features.
6. Cost overrun prediction — predict cost_overrun_flag from slippage + scope complexity features per IPA cost performance.
7. Permit-to-work hazard classification — 4-class hazard level classifier per OSHA 1910.119 PSM.
8. Turnaround cost regression — predict total_cost from labor + material + delay components per AFPM cost benchmarks.
9. Safety event classification — 7-class CCPS event type classifier (rare events; see Honest Disclosure §3).
10. Multi-table relational ML — entity-resolution and graph neural- network learning across the 15 joinable tables via turnaround_id, equipment_id, refinery_id, work_order_id.

---

Loading

from datasets import load_dataset
ds = load_dataset("xpertsystems/oil022-sample", data_files="inspection_findings.csv")
print(ds["train"][0])

Or with pandas:

import pandas as pd
ref     = pd.read_csv("hf://datasets/xpertsystems/oil022-sample/refineries_master.csv")
camps   = pd.read_csv("hf://datasets/xpertsystems/oil022-sample/shutdown_campaigns.csv")
find    = pd.read_csv("hf://datasets/xpertsystems/oil022-sample/inspection_findings.csv")
labels  = pd.read_csv("hf://datasets/xpertsystems/oil022-sample/shutdown_labels.csv")

# Multi-table join for ML feature engineering:
joined = (labels
    .merge(camps, on="turnaround_id")
    .merge(ref, on="refinery_id"))
# Now you have reliability grade alongside slippage + complexity + region + capacity

---

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 turnaround/maintenance ML research, not for live planning decisions. Several notes:

1. Work order completion is ~90% vs declared 93% benchmark — generator samples status uniformly with [0.90, 0.05, 0.04, 0.01] weights, so per-WO completion is 90% by construction. Per-campaign aggregate completion (used in labels) averages 89.94% in the sample. For 90%+ scenarios, this is realistic; for top-decile world-class turnarounds (95%+ per Solomon Associates Q1 performers), the sample is biased low. Use the full product or post-process with quartile-conditional completion priors.

2. Cost overrun rate is ~99.75% because the generator's delay_cost = max(0, actual - planned) × U(450K, 2.5M) is non-zero whenever actual exceeds planned (true for ~99% of sample rows given mean 12.6% slippage). Real cost overrun rates depend on budget granularity — 99% is realistic for "any delay = cost overrun" definition but unrealistic for "≥10% budget overrun" definition. Treat cost_overrun_flag as "any slippage" indicator rather than budget-threshold flag.

3. Safety events are very sparse (~2 events per 100 campaigns) reflecting realistic OGCI/CCPS rates. At sample scale (1200 campaigns), this produces only ~25 events — insufficient for class-balanced 7-class safety event ML. For safety event ML, use the full product (45,000+ campaigns generating 500+ events) or merge with the alarm_trip_logs from OIL-019 / OIL-021.

4. Equipment criticality coupling to failure rate is weak in the sample (failed equipment criticality 0.882 vs safe 0.873 — only 0.009 difference). The generator's fail_prob = failure_rate × (0.55 + criticality) × (1 + slippage/100) formula spreads failure probability across most equipment because failure_rate=0.016 is small. Strong physics signal requires larger samples — the full product (90K equipment) shows clearer criticality-failure coupling.

5. Restart success rate is ~82% vs declared target 96% — the generator's restart_success formula penalizes for several conditions (low readiness count, low reliability score, high restart risk). The sample-scale rate is realistic for moderate- complexity turnaround portfolios but lower than world-class benchmarks. Filter to reliability_grade in ['A', 'B'] for high-performing subset analysis.

6. Reliability grade distribution is B-dominant (52% B, 24% A, 19% C, 5% D) reflecting the slippage-coupled formula. This is a meaningful 4-class distribution unlike degenerate single-class outcomes in some other refinery SKUs — both ordinal classification and continuous reliability_score regression are well-supported.

7. Material family is uniformly carbon-steel-dominant (~63%) per declared weights, reflecting refinery construction reality. But material choice is not coupled to service severity (sour service should drive more Cr-Mo / stainless; high-temp service should drive more refractory alloys). For service-conditioned material ML, the full product v1.1 will add unit-conditioned metallurgy.

8. Catalyst events are sparse (~1 per campaign on average) and only fire for reactor equipment or specific unit types (FCCU, Hydrocracker, Reformer, Hydrotreating). For catalyst lifecycle ML, filter to those unit types and use the catalyst_replacement table directly.

---

Cross-references to other XpertSystems OIL SKUs

This SKU completes the 3-SKU downstream refining trilogy:

SKU	Layer	Focus
OIL-019	Downstream — process	Refinery unit operations (CDU/VDU/FCC reactor + control + HX)
OIL-020	Downstream — yield	Refinery crude-to-product yields + economics + emissions
OIL-021	Cross-stream	Equipment performance + condition monitoring + RUL
OIL-022	Downstream — turnaround	Shutdown/turnaround planning + RBI + inspection + workforce (this SKU)

OIL-022 vs OIL-019: OIL-019 simulates steady-state refinery operations (when units are running). OIL-022 simulates transient turnaround operations (when units are shut down for inspection/maintenance). Use OIL-019 for operational ML, OIL-022 for maintenance planning, scheduling, and turnaround cost ML.

OIL-022 vs OIL-021: OIL-021 simulates continuous equipment condition monitoring (vibration, lubrication, thermal). OIL-022 simulates point-in-time inspection findings (UT, RT, guided wave thickness measurements) during scheduled turnarounds. Use OIL-021 for predictive maintenance ML, OIL-022 for RBI / inspection planning ML.

---

Full product

The full OIL-022 dataset ships at 45,000 turnarounds × 150 refineries × 3,000 equipment per refinery (prod mode) producing several hundred million rows with service-conditioned metallurgy, quartile-realistic completion rates (Q1 95%+ / Q4 75-85%), richer safety event populations (500+ events for class-balanced ML), and stronger equipment-criticality failure coupling (large-sample statistical power) — licensed commercially. Contact XpertSystems.ai for licensing terms.

📧 pradeep@xpertsystems.ai 🌐 https://xpertsystems.ai

---

Citation

@dataset{xpertsystems_oil022_sample_2026,
  title  = {OIL-022: Synthetic Shutdown & Turnaround Dataset (Sample)},
  author = {XpertSystems.ai},
  year   = {2026},
  url    = {https://huggingface.co/datasets/xpertsystems/oil022-sample}
}

Generation details

• Sample version : 1.0.0
• Random seed : 42
• Generated : 2026-05-22 20:40:58 UTC
• Refineries : 15
• Equipment per ref : 150 (2250 total)
• Turnaround campaigns: 1200
• Work orders per TA : 40
• Regions : 10 (US Gulf Coast, US West Coast, North Sea, Middle East, India, Southeast Asia, China, Brazil, North Africa, Western Europe)
• Equipment classes : 14 (Pressure Vessel, Heat Exchanger, Fired Heater, Compressor, Pump, Piping Circuit, Storage Tank, Control Valve, Relief Valve, Reactor, Column, Boiler, Cooling Tower, Instrument Loop)
• Unit types : 14 (CDU, VDU, FCCU, Hydrocracker, Delayed Coker, Reformer, Alkylation, Hydrotreating, Sulfur Recovery, Hydrogen Plant, Tank Farm, Utilities, Cooling Water, Flare System)
• Shutdown types : 5 (Planned TA, Major TA, Pitstop, Emergency Shutdown, Regulatory Outage)
• Corrosion mechanisms: 10 (Uniform, Pitting, Sulfidation, Naphthenic Acid, Erosion-Corrosion, CUI, H2S Damage, Amine, Thermal Fatigue, Chloride SCC)
• Calibration basis : API 510, API 570, API 580/581, NACE TM0274, OSHA 29 CFR 1910.119, AFPM RAM Survey, Solomon Associates, IPA Turnaround Cost Database, OGCI, CCPS, EIA-820, Nelson Complexity Index
• Overall validation: 100.0/100 — Grade A+