TurboSens2

The second scenario of the TurboSens turbofan degradation benchmark, extending TurboSens1 with:

• a two-channel hidden state separating intrinsic mechanical wear $\mathbf{s}_t \in \mathbb{R}^{10}$ from a transient fouling channel $\mathbf{c}t \in \mathbb{R}^{5}{\ge 0}$;
• a phantom observation map $\tilde{\mathbf{s}}_t = \mathrm{clip}(\mathbf{s}_t + M\max(0, -\mathbf{s}_t) + C\mathbf{c}_t)$ so that the sensed state can diverge from the true mechanical state;
• a seventh engine wash action that decays $\mathbf{c}_t$ toward zero without touching $\mathbf{s}_t$ — making the observability of $\mathbf{s}_t$ action dependent;
• six region archetypes (Continental, Littoral, Arid, Equatorial, Highland, Industrial), each characterised by a dominant atmospheric stressor and a different prior over degradation archetypes;
• a hierarchical seasonal weather process and longer episodes (mean ~14k flights, max 20k).

The simpler, two-channel-free scenario is released as TurboSens1 (separate dataset).

Splits

Split	Episodes	Total flights	Mean episode length
train	500	7,010,872	14,022
test	60	830,756	13,846
test_hard	60	854,710	14,245

test_hard skews toward more degraded states and slightly higher extreme-degradation rate; the same generator and region prior as test, distinct seed offset.

Schema (HDF5 columns)

Path	Shape	Description
sensors	(N, 11, 16)	Sensor stream (7 sensors + 4 operating-point parameters) x 16 contexts
observation.state	(N, 10)	Ground truth mechanical wear $\mathbf{s}_t$ (probing target)
action	(N, 1)	Maintenance action index
visit_type	(N,)	Categorical visit type (see below)
component_service_mask	(N, 10)	Boolean mask: which components were serviced at this visit
event_mask	(N,)	Boolean: any event fired at $t$
event_types	(N, 12)	Per event type rising edge flag
weather	(N, 1)	Ambient temperature deviation
valid_flight_mask, repaired_flight_mask	(N,)	Validation flags
ep_offset, ep_len	(N_ep,)	Per-episode offsets
ep_meta/{archetype, archetype_onset, eol_triggered, flights_per_day, is_extreme, region}	(N_ep,)	Episode metadata

File-level attributes

• scenario: "turbosens2"
• n_episodes, n_timesteps
• action_names: [do_nothing, fan_overhaul, hpc_overhaul, turbine_overhaul, full_overhaul, targeted_patch, engine_wash]
• archetype_names: [A_compressor, B_fan_booster, C_turbine, D_balanced]
• region_names: [Continental, Littoral, Arid, Equatorial, Highland, Industrial]
• 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, regional_perm_1, regional_drift_1]
• sensor_names: [HPC_Tout, HP_Nmech, HPC_Tin, LPT_Tin, Fuel_flow, HPC_Pout_st, LP_Nmech]
• context_names, context_phases: per-context labels (16 contexts)

Visit-type encoding

Code	Label	Description
0	do_nothing	No maintenance this step
1	llp_forced	Life-limiting-part cycle ceiling reached
2	proactive	Health-correlated proactive scheduling
3	early_visit	Anticipated visit ahead of an emerging issue
4	opportunistic_repair	Component bundled into another visit after deferred service
5	rerepair	Repeat repair after an imperfect prior intervention
6	wash	Engine water wash — clears $\mathbf{c}_t$ only

Loading

import h5py

with h5py.File("turbosens2_train.h5", "r") as f:
    sensors    = f["sensors"][:]                  # (N, 11, 16)
    state      = f["observation.state"][:]       # (N, 10)
    action     = f["action"][:]                  # (N, 1)
    visit_type = f["visit_type"][:]              # (N,)
    region     = f["ep_meta/region"][:]          # (N_ep,) string codes

Inverse probing + counterfactual evaluation

1. Pretrain any world model self supervised on train's sensor stream alone.
2. Freeze the encoder; train a probe head to predict $\mathbf{s}_t$.
3. Evaluate on test and test_hard.
4. Use the released EpisodeReplayer (companion GitHub repo) to fork any episode under counterfactual maintenance actions and compare the world model's rollout against the simulator's deterministic replay.

Versioning

Generated by turbosens2@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.
• Region archetypes are descriptive labels, not geographic claims; they capture qualitative environmental stressor mixes.
• Single domain. Cross-domain generalisation claims should not be made from TurboSens alone.
• Linear probe sufficiency. Encoders that encode the state in a non-linearly decodable form will appear to fail at probing — informative, not definitive.

Full RAI metadata is in croissant.json.

License

CC-BY-4.0.