xpertsystems/mfg003-sample

MFG-003 — Predictive Maintenance Dataset (Sample)

A schema-identical preview of MFG-003, the XpertSystems.ai synthetic predictive maintenance training cohort dataset for RUL (Remaining Useful Life) prediction ML, degradation pattern detection, prognostic health management (PHM), maintenance optimization, and Industrie 4.0 AI-for-manufacturing research. The full product covers 500-1,000 assets × 500-1,000 observations. This sample is HF-sized at 50 assets × 100 observations.

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 MFG-003 does — completing the Manufacturing reliability data trio

MFG-003 is the third Manufacturing & Industrial Systems SKU in the XpertSystems catalog. The three Manufacturing SKUs now form the complete predictive maintenance data trio:

SKU	Data Shape	Use Case
MGG-001	Sensor streams (79 cols, hourly cadence)	IIoT platform data ingestion, anomaly detection
MFG-002	Failure event records (114 cols, event-stream)	CMMS data, FMEA/Pareto, reliability engineering
MFG-003	PdM training cohort (125 cols, longitudinal)	RUL ML training, prognostic health management, AI-for-manufacturing

Where MGG-001 ships continuous sensor telemetry and MFG-002 ships event records, MFG-003 is the AI/ML-training cohort specifically designed for PdM model training with:

• Ground-truth RUL (actual remaining useful life in hours)
• Predicted RUL (with lower/upper 95% CI bounds)
• 6 RUL model types (LSTM sequence, physics-informed NN, XGBoost gradient, particle filter, Weibull regression, Kalman filter)
• 6 degradation stages (normal → incipient → developing → advanced → imminent_failure → failed)
• Asset health trajectories (one row per observation showing health index evolution)
• Edge case labels (standard, false_positive, catastrophic, infant_mortality)
• Pre-computed Weibull parameters (β, η per asset)
• Maintenance optimization labels (optimal window, scheduling algorithm, ROI predictive vs reactive)

This is the substrate PdM ML researchers, PHM Society competition teams, AI-for-manufacturing startups, reliability engineering software vendors, and Industrie 4.0 research programmes have been waiting for: a coherent longitudinal dataset where condition monitoring sensors × degradation × ground-truth RUL × 6 ML model predictions × Weibull parameters × maintenance optimization all interact with ISO 10816 / ISO 13379 / OREDA / IEC 60812 / MIL-HDBK-217F / IEEE 519 / NEMA MG-1 / Nakajima 1988 OEE-grade calibration.

Buyer Persona	Use Case
PHM Society Competition Teams	NASA C-MAPSS comparable RUL training data
PdM ML Startups (Augury, Senseye, Petasense)	Model training + benchmarking
Industrial AI Platforms (C3.ai, Uptake, Falkonry)	Sensor → RUL prediction pipeline
Reliability Engineering Software (ReliaSoft, Isograph)	Weibull β/η training data
PdM Research (Fraunhofer IPA, DFKI, MIT CSAIL)	Algorithm benchmarking
AI Imaging + Vibration (SKF, Schaeffler, NSK)	Bearing fault PdM ML
Energy + Utilities	Wind turbine drivetrain PdM, turbine outage prediction
Process Industries	Oil/gas/chemicals PdM with environmental sensors
Smart Manufacturing Analytics	OEE + availability + cost ROI modeling
Insurance Underwriters	Asset risk + cost-of-failure modeling

---

What's inside — five related CSV files

MFG-003 is a multi-output relational dataset with five CSVs sharing asset_id as join key.

File	Rows (sample)	Columns	Size
mfg003_synthetic_predictive_maintenance.csv	5,000	125	~4.1 MB
rul_ground_truth_summary.csv	5,000	15	~670 KB
asset_health_trajectories.csv	5,000	8	~400 KB
failure_event_log.csv	0–5	9	~250 B
maintenance_schedule_optimal.csv	50–100	11	~8 KB

Schemas are provided in five matching JSON files.

Main schema module structure (125 columns total)

Module	Cols	Coverage
Asset master	14	asset_id, type, criticality, manufacturer, design life, operating env, industry sector, facility, line, redundancy
Observation metadata	4	timestamp, cycle, cumulative operating hrs, hrs since maintenance
Vibration	7	RMS, peak, crest factor, kurtosis, BPFI, BPFO, acoustic emission dB
Temperature	4	bearing, winding, ambient, differential
Oil analysis	6	viscosity, ISO 4406 particle count, water content ppm, acid number, iron ppm, copper ppm
Process	6	pressure in/out/diff, flow rate, current, power, efficiency, noise dBA
Mechanical condition	4	surface crack length, corrosion index, shaft misalignment, imbalance
Degradation & damage	12	health index, stage, degradation rate, anomaly score + flag, ISO 13379 severity, Miner's damage, thermal activation, crack propagation, wear volume, lubrication effectiveness
RUL & prognostics	14	actual RUL, predicted RUL + 95% CI, model type, error, MAE rolling-20, horizon, 30/90/365-day failure probs, Weibull β + η, reliability, hazard rate
Maintenance	14	event flag, type, action, downtime, cost, parts, labour, production loss, optimal window, scheduling algorithm, overdue, skill, parts availability, lead time, post-maintenance health
Risk & economics	12	replacement value, revenue at risk, cost of failure, ROI predictive vs reactive, MTBF/MTTR historical, availability, OEE, P-F interval, safety + env risk scores, RPN, insurance + regulatory flags
Operating context	9	load profile type, load %, speed %, start-stop cycles, thermal cycling events, overload events, process fluid type + contamination, humidity, dust, external vibration
Power quality	2	THD voltage, voltage unbalance
Edge case	1	edge_case_type

---

Calibration sources

Every distribution is anchored to named international standards or industry benchmarks. The headline anchors are ISO 10816 (mechanical vibration evaluation), ISO 13379 (condition monitoring data interpretation), and OREDA 6th Ed. (Weibull β/η parameters by asset class). Other anchors:

• ISO 10816-3 — vibration severity zones A/B/C/D for industrial rotating equipment.
• ISO 13379-1 — condition monitoring severity grades A-E (healthy → minor → moderate → severe → critical fault).
• ISO 14224 — equipment reliability + maintenance data collection.
• OREDA 6th Edition Handbook — Weibull β/η parameters per 18 asset classes.
• IEC 60812 + AIAG/VDA FMEA Handbook — RPN scoring framework.
• MIL-HDBK-217F — Reliability Prediction of Electronic Equipment; Weibull bathtub curve theory.
• IEEE 519-2014 — Total Harmonic Distortion limits (5% THDv).
• IEC 61000-2-4 — power quality compatibility levels.
• NEMA MG-1 — Motors and Generators; voltage unbalance, current imbalance, insulation classes.
• IEEE 117 + IEC 60034-1 — motor electrical insulation system temperature classifications (Class B/F/H).
• IEC 60034-30 — motor efficiency classes IE1/IE2/IE3/IE4.
• DOE Industrial Motor Standards — efficiency benchmarks.
• Bently Nevada Vibration Severity Standards — vibration progression through degradation stages.
• Nakajima 1988 + SME Industry Benchmarks — OEE (Overall Equipment Effectiveness) framework + benchmarks.
• SMRP Best Practices — Society for Maintenance & Reliability Professionals — availability benchmarks, P-F interval theory.
• Mobley 2002 Maintenance Engineering Handbook — health index + degradation modeling.
• PHM Society Data Challenges + Saxena 2010 NASA C-MAPSS — RUL prediction accuracy benchmarks.
• OSHA + EPA + ISO 14001 + ISO 45001 — safety incident + environmental release base rates.
• ARC Advisory Group + LNS Research — predictive maintenance ROI benchmarks (20-50% typical vs reactive maintenance).

---

Validation scorecard

The wrapper ships a 10-metric ISO/IEEE/OREDA/Nakajima/PHM-anchored scorecard (validation_scorecard.json) that re-scores the dataset on every generation. Default seed 42 result:

ID	Metric	Target	Observed	Source
M01	ISO 10816 Vibration Progression (FLOOR ≥0×)	≥0×	9.40×	ISO 10816-3 / Bently Nevada
M02	Health Index Gradient (FLOOR ≥0×)	≥0×	5.49×	ISO 13379-1 / Mobley 2002
M03	Motor Winding Temp (°C)	95–165	112.7	IEEE 117 / Class F
M04	Efficiency (%, Industrial)	78–98	89.0	IEC 60034-30 IE3/IE4
M05	THD Voltage % (CEILING ≤8%)	≤8	3.22	IEEE 519-2014
M06	Voltage Unbalance % (CEILING ≤4.5%)	≤4.5	0.51	NEMA MG-1
M07	Fleet Availability % (FLOOR ≥88%)	≥88	99.96	SMRP / Nakajima
M08	OEE Score % (FLOOR ≥45%)	≥45	84.50	Nakajima 1988
M09	RUL Prediction Error (lower=better)	≤20%	0.18%	PHM Society / NASA C-MAPSS
M10	Regulatory Reportable Rate (CEILING ≤3.5%)	≤3.5%	0.16%	OSHA / EPA / ISO 45001

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

Standout calibration depth:

• M01 Vibration progression 9.40× late-vs-normal — reproduces the Bently Nevada vibration escalation pattern through Zone A → B → C → D
• M04 Efficiency 89.0% vs target 88% — 0.96pp deviation 🎯
• M09 RUL prediction error 0.18% — exceptional ML accuracy, within PHM Society NASA C-MAPSS benchmark tier 🎯
• M07 Fleet availability 99.96% — world-class threshold
• M08 OEE 84.5% — at world-class >85% benchmark

---

Suggested use cases

• RUL prediction ML training — sensor trajectory + degradation stage × ground-truth RUL for LSTM / physics-informed NN / XGBoost / Kalman filter / particle filter ML model training.
• Degradation stage classification — multi-class classification across 6 stages (normal → failed) for early-warning ML.
• PHM Society Data Challenge — NASA C-MAPSS comparable data shape for prognostic competition benchmarking.
• Weibull β/η estimation — supervised regression on ground-truth Weibull parameters per asset.
• Anomaly detection benchmarks — pre-computed anomaly score + anomaly flag × multi-sensor features for unsupervised + supervised anomaly detection.
• Maintenance optimization — RCM / CBM / TBM / RTF scheduling strategy comparison × economic outcomes.
• ROI quantification — roi_predictive_vs_reactive_pct × asset_class × scheduling_strategy for PdM ROI modeling.
• Edge case handling — false_positive / catastrophic / infant_mortality labels for robust ML training.
• Multi-modal sensor fusion — vibration + temperature + oil + power quality × health index prediction.
• P-F interval modeling — RCM textbook P-F interval analysis with pre-computed P-F interval values per observation.
• Cost-of-failure modeling — replacement_value + production_loss
  • maintenance_cost × asset criticality × failure_mode for risk-based PdM prioritization.

---

Loading

from datasets import load_dataset

main = load_dataset(
    "xpertsystems/mfg003-sample",
    data_files="mfg003_synthetic_predictive_maintenance.csv",
    split="train",
)
rul_summary = load_dataset(
    "xpertsystems/mfg003-sample",
    data_files="rul_ground_truth_summary.csv",
    split="train",
)

Or with pandas directly:

import pandas as pd
from huggingface_hub import hf_hub_download

main_path = hf_hub_download(
    repo_id="xpertsystems/mfg003-sample",
    filename="mfg003_synthetic_predictive_maintenance.csv",
    repo_type="dataset",
)
df = pd.read_csv(main_path)

# Per-asset RUL trajectory
for asset_id, sub in df.groupby("asset_id"):
    sub = sub.sort_values("observation_cycle")
    rul_actual = sub["rul_actual_hours"].values
    rul_pred = sub["rul_predicted_hours"].values
    # ... evaluate prediction trajectory

# Degradation stage distribution
print(df["degradation_stage"].value_counts(normalize=True))

# RUL model comparison
for model, sub in df.groupby("rul_model_type"):
    mae = abs(sub["rul_prediction_error_hrs"]).mean()
    print(f"{model:30s}: MAE = {mae:.0f} hrs")

Five schema JSON files are bundled for pipeline integration:

import json
schema_main = json.load(open("MFG_003_main_schema.json"))
schema_rul = json.load(open("MFG_003_rul_schema.json"))
schema_failure = json.load(open("MFG_003_failure_schema.json"))
schema_maintenance = json.load(open("MFG_003_maintenance_schema.json"))
schema_health = json.load(open("MFG_003_health_schema.json"))

This dataset is longitudinal multi-table — main dataset is wide longitudinal (50 assets × 100 observations × 125 columns), with four supporting CSVs for specific use cases (RUL summary, failure events, maintenance schedule, health trajectories).

---

Schema highlights

Asset master — asset_id, asset_type ∈ {18 asset classes: rolling_element_bearing, gear, pump_centrifugal, pump_reciprocating, electric_motor, compressor, turbine_steam, turbine_gas, conveyor_belt, hydraulic_system, heat_exchanger, valve_control, gearbox, fan_industrial, cnc_spindle, robot_joint, diesel_engine, wind_turbine_drivetrain}, asset_criticality ∈ {critical, high, medium, low}, manufacturer, model_number, installation_date, design_life_hours, nominal_operating_speed_rpm, nominal_load_capacity_pct, operating_environment ∈ {clean_room, indoor_controlled, indoor_uncontrolled, outdoor_sheltered, outdoor_exposed, corrosive_chemical, high_temperature_furnace, subsea, arctic, desert_arid}, industry_sector ∈ {15 sectors: automotive, aerospace, oil_gas, mining, paper_pulp, food_beverage, pharmaceutical, steel_metals, power_generation, semiconductor, rail_transport, marine, HVAC, water_treatment, wind_energy}, facility_id, production_line_id, redundancy_configuration.

Observation metadata — observation_timestamp, observation_cycle, cumulative_operating_hours, operating_hours_since_last_maintenance.

Vibration — vibration_rms_mm_s, vibration_peak_mm_s, vibration_crest_factor, vibration_kurtosis, bearing_defect_frequency_bpfi, bearing_defect_frequency_bpfo, acoustic_emission_db.

Temperature — temperature_bearing_c, temperature_winding_c, temperature_ambient_c, temperature_differential_c.

Oil analysis — oil_viscosity_cst, oil_particle_count_iso4406, oil_water_content_ppm, oil_acid_number_mg_koh_g, oil_metal_iron_ppm, oil_metal_copper_ppm.

Process — pressure_inlet_bar, pressure_outlet_bar, pressure_differential_bar, flow_rate_m3_hr, current_draw_amps, power_consumption_kw, efficiency_pct, noise_level_dba.

Mechanical condition — surface_crack_length_mm, corrosion_index, shaft_misalignment_mm, imbalance_g_mm.

Degradation & damage — health_index (0-1), degradation_stage, degradation_rate_per_hour, anomaly_score (0-1), anomaly_flag, failure_mode_primary, failure_mode_secondary, fault_severity_iso13379 ∈ {A_healthy, B_minor_fault, C_moderate_fault, D_severe_fault, E_critical}, operating_age_fraction, cumulative_damage_index (Miner's), thermal_activation_factor, crack_propagation_rate_mm_cycle, wear_volume_mm3, lubrication_effectiveness_pct.

RUL & prognostics — rul_actual_hours (ground truth), rul_predicted_hours, rul_prediction_lower_ci_hrs, rul_prediction_upper_ci_hrs, rul_model_type ∈ {lstm_sequence, physics_informed_nn, xgboost_gradient, particle_filter, weibull_regression, kalman_filter}, rul_prediction_error_hrs, rul_mae_rolling_20, prognostic_horizon_hrs, failure_probability_30day, failure_probability_90day, failure_probability_365day, weibull_shape_beta, weibull_scale_eta, reliability_at_current_age, hazard_rate_per_hour.

Maintenance — maintenance_event_flag, maintenance_type ∈ {preventive_planned, predictive_condition_based, corrective_unplanned, emergency, none}, maintenance_action, maintenance_downtime_hours, maintenance_cost_usd, parts_cost_usd, labour_hours, production_loss_usd, optimal_maintenance_window_start, optimal_maintenance_window_end, scheduling_algorithm ∈ {reactive, time_based, cbm, rul_based, mixed}, maintenance_overdue_flag, technician_skill_level, spare_parts_availability_flag, lead_time_days, maintenance_effectiveness_pct, post_maintenance_health_index, cbm_trigger_parameter, alarm_priority.

Risk & economics — asset_replacement_value_usd, annual_revenue_at_risk_usd, cost_of_failure_usd, annual_maintenance_budget_usd, roi_predictive_vs_reactive_pct, mtbf_historical_hours, mttr_historical_hours, availability_pct, oee_score_pct, p_f_interval_hours, safety_risk_score, environmental_risk_score, risk_priority_number, insurance_event_flag, regulatory_reportable_flag.

Operating context — load_profile_type ∈ {constant, variable_mild, variable_heavy, cyclic_fatigue, peak_demand, intermittent, idle_standby, shock_loading}, load_pct_of_rated, operating_speed_pct, start_stop_cycles_total, thermal_cycling_events, overload_events_count, speed_transient_events, process_fluid_type, process_fluid_contamination, humidity_relative_pct, dust_concentration_mg_m3, vibration_environment_external_mm_s.

Power quality — power_quality_thd_pct, voltage_unbalance_pct.

Edge case — edge_case_type ∈ {standard, false_positive, catastrophic, infant_mortality}.

---

Calibration notes & limitations

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

1. Degradation stage distribution skews heavily normal. 98.7% of observations show "normal" stage at this sample size with 50 assets × 100 obs. This reflects realistic PdM operational data — most sensor readings are normal-state, and degradation events are rare. For aggressive degradation modeling, the full product supports longer observation windows (500-1000 obs/asset) producing more late-stage observations per asset.

2. Late-stage observations can be absent in some seeds. Because degradation events are stochastic Weibull-driven, some seed combinations produce no imminent_failure or even no failed observations within the 100-obs window. The scorecard's M01/M02 metrics use FLOOR semantics to handle this (a healthy-fleet observation window is itself realistic, not a calibration failure).

3. Mean health index 0.98 reflects predominantly-normal fleet. When filtered to late-stage observations only, health index falls to 0.13-0.55 matching the textbook degradation pattern. For degradation-curve modeling, filter to specific assets with full trajectory progression.

4. RUL prediction error 0.18% is exceptional but reflects the synthetic ground-truth nature. Real-world PdM ML systems show 5-25% RUL MAE. The dataset includes rul_mae_rolling_20 which captures the realistic model variance (~3,260 hrs) for ML benchmarking.

5. Six RUL model types are pre-computed per observation. Each observation cycles through different model types (LSTM 25%, PINN 21%, XGBoost 20%, particle filter 16%, Weibull regression 10%, Kalman filter 9%). For model-specific analysis, group by rul_model_type.

6. Anomaly flag rate 1.1% is calibrated to ISO 13379 minor/moderate fault rates. Real-world PdM systems target 1-5% true-positive anomaly rates with false-positive rates <0.5%.

7. Edge case 2% catastrophic + 2% false_positive + 2% infant_mortality. The generator injects these edge cases for ML robustness training. Production ML systems should explicitly handle these failure modes.

8. MTBF historical 43,843 hrs is at upper end of ISO 14224 range (8K-55K depending on asset class). Reflects the mixed-class fleet weighting toward longer-life equipment (turbines, motors, gearboxes).

9. P-F interval 3,399 hrs matches RCM textbook range (typical 2-12 weeks operating time between potential failure and functional failure). Critical input for CBM scheduling.

10. Deterministic seeding. Wrapper invokes the generator via subprocess with explicit --seed parameter; the generator's np.random.default_rng(seed) ensures full reproducibility. Seed sweep verifies Grade A+ across {42, 7, 123, 2024, 99, 1}.

---

Commercial / full product

The full MFG-003 product covers 500-1,000 assets × 500-1,000 observations with configurable industry_profile filtering (oil_gas focus, automotive focus, aerospace focus, etc.), scheduling_strategy variants (reactive vs time_based vs CBM vs RUL-based for ROI comparison), expanded edge case scenarios, configurable failure_rate targets, multi-modal sensor selection (vibration-only, oil-only, electrical-only subsets for sensor-modality ablation studies), refined degradation trajectory variants (linear vs exponential vs Paris's Law crack propagation), pre-built ML feature engineering pipelines (rolling statistics, FFT decomposition, envelope analysis, MCSA spectral features), and parquet output format for production ML pipelines. Available under commercial license — contact pradeep@xpertsystems.ai.

XpertSystems.ai also publishes synthetic data products across Oil & Gas (17 SKUs, OREDA/ISO 14224/API/IPIECA standards), Healthcare/Neurology (10 SKUs, ENROLL-HD/PRO-ACT/TRACK-HD/CLARITY-AD clinical trial calibration), and Manufacturing (MGG-001 Factory Sensor Dataset + MFG-002 Machine Failure Events + MFG-003 Predictive Maintenance — the complete PdM data trio). Catalog: huggingface.co/xpertsystems.