---
language: [en]
license: cc-by-4.0
size_categories: [1M<n<10M]
task_categories: [time-series-forecasting]
pretty_name: TurboSens1
tags: [turbofan, world-models, representation-learning, probing, synthetic]
configs:
  - config_name: default
    data_files:
      - {split: train,     path: "turbosens1_train.h5"}
      - {split: test,      path: "turbosens1_test.h5"}
      - {split: test_hard, path: "turbosens1_test_hard.h5"}
---

# TurboSens1

The first scenario of the **TurboSens** turbofan degradation benchmark — an
interactive simulator dataset that pairs sensor observations with the full
ten-dimensional ground-truth health state of the engine at every flight
cycle, enabling a direct **inverse probing** protocol for any self
supervised world model.

TurboSens2 (the second, harder scenario with fouling, phantom observation
and region archetypes) is released as a separate dataset.

## Splits

| Split | Episodes | Total flights | Episode length | Notes |
|---|---:|---:|---:|---|
| `train`     | 400 | 4,000,000 | 10,000 (capped) | 4 archetypes, 10 events |
| `test`      | 40  | 400,000   | 10,000 (capped) | held out, same generator and prior |
| `test_hard` | 40  | 400,000   | 10,000 (capped) | OOD: stronger degradation, modified maintenance |

## Schema (HDF5 columns)

| Path | Shape | Description |
|---|---|---|
| `sensors` | (N, 7, 12) | Sensor stream (7 channels x 12 flight phase contexts) |
| `observation.state` | (N, 10) | Ground truth mechanical wear $\mathbf{s}_t$ (probing target) |
| `action` | (N, 1) | Maintenance action index |
| `event_mask` | (N,) | Boolean: any event fired at $t$ |
| `event_types` | (N, 10) | Per event type rising edge flag |
| `weather` | (N, 1) | Ambient temperature deviation |
| `ep_offset`, `ep_len` | (N_ep,) | Per episode flat offsets |
| `ep_meta/{archetype, archetype_onset, eol_triggered, success_coeff}` | (N_ep,) | Episode metadata |

### File-level attributes

- `scenario`: `"turbosens1"`
- `n_episodes`, `n_timesteps`
- `action_names`: `[do_nothing, fan_overhaul, hpc_overhaul, turbine_overhaul, full_overhaul, targeted_patch]`
- `archetype_names`: `[A_compressor, B_fan_booster, C_turbine, D_balanced]`
- `event_names`: abstract effect-typed labels — `[perm_step_1, trans_drift_1, trans_anomaly_1, perm_step_2, trans_fouling_1, trans_anomaly_2, perm_drift_1, perm_drift_2, sensor_pulse_1, trans_drift_2]`
- `sensor_names`: `[HPC_Tout, HP_Nmech, HPC_Tin, LPT_Tin, Fuel_flow, HPC_Pout_st, LP_Nmech]`

## Loading

```python
import h5py

with h5py.File("turbosens1_train.h5", "r") as f:
    sensors = f["sensors"][:]                  # (N, 7, 12)
    state   = f["observation.state"][:]       # (N, 10)
    action  = f["action"][:]                  # (N, 1)
    ep_off  = f["ep_offset"][:]               # (N_ep,)
    ep_len  = f["ep_len"][:]                  # (N_ep,)
```

Or via the HuggingFace `datasets` library (downloads only; HDF5 must be
parsed with `h5py`):

```python
from huggingface_hub import hf_hub_download
import h5py

p = hf_hub_download(repo_id="<namespace>/turbosens1",
                    filename="turbosens1_train.h5", repo_type="dataset")
with h5py.File(p, "r") as f:
    ...
```

## Inverse probing protocol

1. Pretrain any world model self supervised on the `train` split's sensor
   stream alone.
2. Freeze the encoder; train a small **probe** head on top of the frozen
   embeddings to predict $\mathbf{s}_t$.
3. Evaluate on `test` and `test_hard`. Report per-component Pearson, RMSE,
   $R^2$.

Reference baselines (JEPA, AR-LSTM, RSSM) and probe code live in the
companion GitHub repository.

## Versioning

Generated by `turbosens1@v1.0.0` (deterministic simulator, stamped in
`SIM_VERSION`).

## Caveats and intended use

- **Synthetic data, not a calibration of any real fleet.** Stochastic
  event rates and magnitudes are mathematical abstractions and do not
  reflect operational fleet failure statistics.
- **Single domain.** Cross-domain generalisation claims should not be
  made from TurboSens alone.
- **Linear probe sufficiency.** The protocol assumes a linear probe is
  expressive enough; encoders that encode the state in a non-linearly
  decodable form will appear to fail at probing — informative, not
  definitive about representation quality.

Full RAI metadata is in `croissant.json`.

## License

CC-BY-4.0.