Predictive Maintenance Fleet

• Remaining useful life (RUL) regression

  with rul_days or true_rul_days, and failure within 30 days classification with fails_within_30d (5.6% positive, the realistic imbalance).

• Survival analysis

  239 runs are right-censored at the cut-off and 204 were ended early by a repair or replacement; Weibull, Cox and Kaplan-Meier on runs.csv with load, climate and failure mode as covariates.

• Anomaly detection and alarm tuning

  the alarms are rule-based (ISO zone, temperature, pressure, current, debris); beat them with a model, and measure false alarms against lead time.

• Failure mode diagnosis

  from the sensor signature: vibration for bearings, imbalance and misalignment, temperature for lubrication, pressure for impellers and seals, current for windings, oil debris for gears.

• Maintenance strategy and cost

  run-to-failure against condition-based against age-based replacement; breakdowns cost a median $30,991 against $9,796 for a condition-based repair.

• Dashboards and teaching

  MTBF, MTTR, availability, backlog, PM compliance, cost by site and criticality in Power BI, Tableau or SQL; a reliability engineering course or a CMMS demo.

• runs.failure_modeThe true failure mode of the life, which sets the signals that move.8 values: bearing wear, shaft misalignment, seal leak…
• runs.true_life_daysHow long the life lasts to failure if nothing intervenes; a run caught or replaced early ends before it.52.8 to 3,979.8, mean 428.9
• runs.true_failure_atWhen the machine would fail, including for runs that were caught or renewed first.2024-01-11 19:12:00 to 2035-08-01 22:00:00
• runs.warning_signsWhether the failure gives signs a sensor can see, or comes suddenly.true on 82.2% of rows
• runs.pf_onsetThe share of the life at which the signals start to rise (the P on the P-F curve); near 1 for sudden failures.0.55 to 0.98, mean 0.69
• readings.life_usedThe share of the run's true life used at the reading.0 to 1, mean 0.46, 0.78% empty
• readings.rul_daysDays left until the run ends, however it ends.0 to 3,979.4, mean 289.6, 0.78% empty
• readings.true_rul_daysDays left until the true failure: the remaining-useful-life target.0 to 3,979.4, mean 325.6, 0.78% empty
• readings.fails_within_30dWhether the machine truly fails within 30 days: the classification target.true on 5.6% of rows

• Signals stay flat for most of a life, then rise along a P-F curve: median vibration on vibration-type failures goes 1.62 -> 1.63 -> 2.50 -> 8.62 mm/s from new to the last 3% of life.
• 91% of warned and 7% of sudden failures gave a sustained alarm first, median 46 days ahead. 25% of failures were sudden: some faults give no warning a sensor can see.
• Harder-worked machines live shorter (bearings with the cube of load): Spearman -0.57 between load factor and true life; hot sites wear machines out sooner (median true life 275 against 333 days).
• Near the end of a life, bearing temperature is 62 C against 42 C for bearing and lubrication failures against other modes; discharge pressure falls to 78% of nominal against 96% for impeller and seal failures.
• Breakdowns keep a machine down a median 51 h against 6 h for planned work; lost production is $20,565,604 of the $22,049,444 maintenance cost.
• Runs: 328 failures, 123 caught by monitoring, 81 age-based replacements, 239 still running at the cut-off. Sensor dropouts: about 0.8% of readings.

• every key resolves; one reading per asset per day for all 731 days; runs of an asset never overlap and each starts after the last repair ended;
• the run in force is rebuilt for every reading from the run table, and every label (run age, life used, remaining life, true remaining life, fails within 30 days) recomputes from it;
• no sensor value while the asset is down; pressure only on pumps and compressors, oil debris only on gearboxes and compressors; ISO 10816-3 zone, alarm and alarm reason recompute from the readings;
• one work order per run ended in the window, its type from how the run ended, its times from the run, its costs re-adding; inspection findings recompute from the measured vibration and follow each asset's PM interval;
• the physics above: P-F curve, mode-specific signals, load and climate against life, criticality against catches, sudden failures, false alarm rate.

• readingsOne reading per asset per day for two years: vibration RMS and crest factor, bearing temperature, motor current, discharge pressure, oil debris, load, ambient, ISO 10816-3 zone, alarm and its reason; plus labels (run age, life used, remaining life to the run's end and to the true failure, fails within 30 days)175,440 rows · 24 cols
• inspectionsRoute-based vibration checks at each asset's PM interval, with the measured value and the finding2,042 rows · 7 cols
• runsEvery life of every asset, from install or repair to the next failure or renewal: failure mode, true life, whether it gave warning signs, how it ended, downtime, and whether it was still running at the cut-off771 rows · 19 cols
• work_ordersOne per run that ended in the window: breakdown, condition-based or preventive, priority, raised/started/completed, component replaced, found condition, labour hours, parts, labour and lost-production cost532 rows · 19 cols
• assetsPumps, motors, fans, gearboxes and compressors: type, criticality A/B/C, rated power and current, nominal pressure, install year, load factor, duty hours, online monitoring or route-based, PM interval240 rows · 15 cols
• sitesFour plants in different climates: ambient temperature and its seasonal swing, spares lead time, cost of an hour of downtime4 rows · 7 cols

Is this real data?

No. This is synthetic data generated by software. No row describes a real person, company, patient, store, machine or transaction. The patterns are modelled to be realistic and the statistics quoted are measured on these files, but they do not describe any real population or market. Use it for learning, testing, demos, benchmarks and prototyping, not as evidence about the real world. Provided as is, without warranty.

What do I get?

One zip of 6.1 MB: 6 linked tables and 179,029 rows as CSV, a README of what each table holds and how the data behaves, INTEGRITY.json with the 62 audit checks, and RECIPE.json, the Misata blueprint that made these exact rows (seed 20240101).

Can I try it before buying?

Yes. The free preview (303 KB) is a slice of the same data with the keys intact, plus the README, so you can load it and check it fits before you pay.

How was it checked?

62 of 62 checks pass, re-run on the delivered files by an independent script with plain pandas: every key resolves; one reading per asset per day for all 731 days; runs of an asset never overlap and each starts after the last repair ended; the run in force is rebuilt for every reading from the run table, and every label (run age, life used, remaining life, true remaining life, fails within 30 days) recomputes from it; no sensor value while the asset is down; pressure only on pumps and compressors, oil debris only on gearboxes and compressors; ISO 10816-3 zone, alarm and alarm reason recompute from the readings; one work order per run ended in the window, its type from how the run ended, its times from the run, its costs re-adding; inspection findings recompute from the measured vibration and follow each asset's PM interval; the physics above: P-F curve, mode-specific signals, load and climate against life, criticality against catches, sudden failures, false alarm rate.

Can I use it commercially?

Yes: in any project, course, benchmark, demo or product, commercial or not. You may not resell or redistribute the dataset itself as a dataset.

What should I know before using it?

ISO 10816-3 zone limits (2.3 / 4.5 / 7.1 mm/s) are applied to every asset for simplicity; real limits depend on machine class and foundation. true_rul_days and true_failure_at are ground truth no plant could know; use rul_days for a realistic label and the true values to study how censoring biases a model. pf_onset is the fraction of the true life at which degradation became visible (about 0.985 for a sudden failure). Lives are compressed so two years hold enough failures to learn from: the median life of a failure mode is 260 to 760 days, several times shorter than most plants see (a well-kept bearing or gear set often runs for years). The shape of each life (the P-F curve, the load and climate effects, the censoring) is what the data is built to teach, not its length; do not quote the MTBF as a benchmark. lost_production_cost is downtime hours times the site's hourly rate, scaled by criticality (A 1, B 0.45, C 0.12). Sites, assets and models are invented; no real plant, manufacturer or equipment record is represented.

Can I get a bigger or different version?

Yes. RECIPE.json runs in Misata Studio or through Misata's MCP server to make a variant, or ask us to build one to your spec.

Want it bigger, in another setting, or with your own columns? Have us build it or make it in Studio.