Predictive maintenance
Declare when the machine fails.
The remaining useful life is then exact.
Most public predictive-maintenance data draws every row on its own. That is right for orders and payments and wrong for equipment, because a machine has a history. Misata inverts it: a unit draws a life, damage accumulates toward that exact moment, and every sensor follows the damage. Remaining useful life is solved for rather than annotated, which is what makes the label worth training on.
Try the declaration
Move a slider and the fleet changes. The YAML underneath is the real input, not a mock-up of one, and the maths matches the engine rather than approximating it.
one machine each, followed to failure
mean cycles before failure
0 makes every machine identical
how differently each machine responds
measurement error on every reading
degradations:
- table: readings
units: 8
life_mean: 220
life_std: 45
unit_variation: 0.10
responses:
- column: vibration_mm_s
baseline: 0.8
at_failure: 5.2
shape: exponential
noise: 0.08That is the whole input. Paste it into a schema and misata generate produces the fleet above, with rul_cycles exact on every row.
What the existing datasets cannot give you
| AI4I 2020 | C-MAPSS | This | |
|---|---|---|---|
| A machine you can follow over time | |||
| A remaining-useful-life label | |||
| Wear that never goes backwards | |||
| Change the fleet size or failure mix | |||
| The latent damage state, for checking a model | |||
| Validated against real hardware |
The last row is the honest one. C-MAPSS and AI4I describe real machinery; this describes a declared one. Use it where you need ground truth that real data cannot give you, and cite the real sets where you need real physics.
What holds, measured on the published files
- RUL is exact on all 100 units. It decrements by exactly 1 each cycle and reaches 0 on the failure cycle. No smoothing, no clipping.
- Tool wear never decreases, on 100% of steps. Declare a cumulative quantity monotonic and it stops going backwards, however much noise you add.
- Wear correlates +0.849 with cycle. Something is genuinely progressing toward failure.
- Failure mode is learnable, not decorative. Units failing by tool wear reach a mean 357 minutes against 245 to 261 for the other modes, and heat-dissipation units reach 325.6 K against about 316 K.
- Sensors are correlated but not collinear: mean 0.62, max 0.83. Each unit draws its own susceptibility per sensor, so the fleet does not move as one body.
What it refuses to claim
A generator that claims everything is worth nothing, so these are stated here and on the dataset itself rather than buried.
- The physics is not validated. The damage law is a simplified lumped model, unchecked against XJTU-SY, PRONOSTIA/FEMTO or IMS.
- One damage process per unit. Real machines fail from several interacting mechanisms.
- No per-machine attributes beyond a control type. No location, no maintenance history, no operator.
- Not a drop-in replacement for C-MAPSS in published benchmarks.
Questions
- What makes remaining useful life exact rather than estimated?
- The failure time is drawn before any row exists. Each unit gets a life, damage accumulates toward exactly that moment, and the rows are solved so the countdown lands on zero on the failure cycle. Nothing is labelled after the fact, so rul_cycles decrements by exactly 1 each cycle on all 100 units in the published dataset.
- How is this different from AI4I 2020?
- AI4I has no unit identity and no time index. Each of its 10,000 rows is an independent snapshot of a different product, so there is no machine to follow, no trajectory, and no remaining-life label to predict. It is a good classification dataset, but nothing in it progresses toward failure.
- How is this different from NASA C-MAPSS?
- C-MAPSS has trajectories and RUL and remains the standard benchmark. It is also a fixed turbofan simulation from 2008: you cannot change the fleet size, the failure mix or the noise level. This is generated from a declaration, so you can change any of them and regenerate.
- Is the physics validated against real bearings?
- No, and that is stated on the dataset itself. The damage law is a simplified lumped model and has not been checked against XJTU-SY, PRONOSTIA/FEMTO or IMS run-to-failure data. The defensible claim is that the labels are exact and the trajectory is declared, not that the degradation is faithful to a specific machine.
- Can I use it commercially?
- Yes. The dataset is public domain (CC0) and Misata is MIT licensed. No real machine data was sampled and no model was fitted to real data, so there is no proprietary or personal information in it.
- Can I generate a fleet of my own size and failure mix?
- Yes. The schema.yaml that produced the published dataset ships with it and runs as-is. Change units, life_mean, the failure modes or the sensor responses and run misata generate. Remaining useful life stays exact because it is solved for rather than annotated.
Generate your own fleet
pip install misata misata generate --config schema.yaml --output-dir ./data
The schema.yaml that produced the published dataset ships inside it and runs unmodified. Change the fleet size, the mean life or the sensor response and regenerate. The countdown stays exact because it is solved, not labelled.