rasinmuhammed/predictive-maintenance-remaining-useful-life

Machine Degradation with Exact Remaining Useful Life

A synthetic run-to-failure dataset: 100 machines, each followed from install to failure, 23,118 hourly readings in total. Every row carries the true remaining useful life, because the failure time was declared before the data was generated rather than annotated afterwards.

That last sentence is the whole point, so it is worth being precise about what it buys you and what it does not.

Why this exists

AI4I 2020, the most downloaded public predictive-maintenance dataset, has a structural problem for prognostics work: it has no unit identity and no time index. Each of its 10,000 rows is an independent snapshot of a different product. There is no machine you can follow, so there is no trajectory, and there is no remaining-useful-life label to predict. It is a very good classification dataset wearing a prognostics costume.

NASA C-MAPSS has trajectories and RUL, and remains the standard, but it is a turbofan simulation from 2008 with a fixed set of conditions you cannot change. If you want a different fleet size, a different failure mix, or a different noise level, you cannot have one.

This dataset is generated from a declaration, so you can have one. The schema.yaml in this repo is the entire specification and it runs as-is.

What is in it

file	rows	what it is
readings.csv	23,118	one row per machine per cycle, sensors plus labels
units.csv	100	one row per machine: its true life, failure mode, split
ground_truth.csv	23,118	the latent damage value behind each reading

readings.csv columns:

• unit_id, cycle: which machine, and how many cycles since install
• rul_cycles: remaining useful life, exact, counting down to 0 at failure
• machine_failure: 1 on the final cycle of each unit
• failure_mode: tool_wear, heat_dissipation, power, overstrain
• tool_wear_min, vibration_mm_s, torque_nm, process_temperature_k, air_temperature_k, rotational_speed_rpm: the observable sensors
• control_type: the machine's quality grade
• split: train (80 units) or test (20 units), split by unit so no machine appears in both

ground_truth.csv gives the hidden damage state, which real telemetry never has. It is there so you can check whether a model recovered the latent process or only fitted the sensors.

What holds, measured on these files

Every line below was computed from the CSVs in this repo, not asserted:

• RUL is exact on all 100 units. It decrements by exactly 1 each cycle and reaches 0 on the failure cycle. There is no smoothing and no clipping.
• Tool wear never decreases: 100% of consecutive steps. Material does not come back. In AI4I, noise alone makes wear fall about as often as it rises.
• Wear correlates +0.849 with cycle. Something is actually progressing toward failure.
• Failure mode is learnable rather than decorative. Units that fail by tool_wear reach a mean 357 minutes of wear against 245 to 261 for the other modes, and heat_dissipation units reach 325.6 K against about 316 K. The mode leaves a signature in the sensors, so predicting it from telemetry is a real task.
• Sensors are correlated but not collinear: mean 0.62, max 0.83 across the four degradation-driven channels. Each unit draws its own susceptibility per sensor, so the fleet does not move as one body.
• Lives range from 128 to 358 cycles, drawn per unit.

INTEGRITY.txt ships the same checks so you can re-run them.

Honest limits

Read these before citing it.

• The physics is not validated. The damage law is a simplified lumped model. This 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 bearing or spindle.
• One damage process per unit. Real machines fail from several interacting mechanisms. Here the failure mode shapes a single underlying process.
• No per-machine attributes beyond the control type. No location, no maintenance history, no operator.
• Not a drop-in replacement for C-MAPSS in published benchmarks. Use it for controlled experiments, for sanity-checking a pipeline, and for cases where you need ground truth that real data cannot give you.

Make your own

pip install misata
misata generate --config schema.yaml --output-dir ./data

Change the fleet size, the mean life, the failure mix or the sensor response, and regenerate. The RUL stays exact because it is solved for, not labelled.

degradations:
  - table: readings
    units: 100
    life_mean: 220
    life_std: 45
    responses:
      - {column: tool_wear_min, baseline: 0, at_failure: 250, monotonic: true}
      - {column: vibration_mm_s, baseline: 0.8, at_failure: 5.2, shape: exponential}

Provenance

Generated by Misata (MIT). No real machine data was sampled and no model was fitted to real data, so this contains no proprietary or personal information. Landing page with a trajectory chart: misata.studio/datasets/machine-degradation.

Citation

@misc{misata_machine_degradation_2026,
  title  = {Machine Degradation with Exact Remaining Useful Life},
  author = {Muhammed Rasin O M},
  year   = {2026},
  note   = {Synthetic run-to-failure dataset generated declaratively with Misata},
  url    = {https://huggingface.co/datasets/rasinmuhammed/predictive-maintenance-remaining-useful-life}
}