{ "@context": { "@language": "en", "@vocab": "https://schema.org/", "citeAs": "cr:citeAs", "column": "cr:column", "conformsTo": "dct:conformsTo", "cr": "http://mlcommons.org/croissant/", "rai": "http://mlcommons.org/croissant/RAI/", "data": { "@id": "cr:data", "@type": "@json" }, "dataType": { "@id": "cr:dataType", "@type": "@vocab" }, "dct": "http://purl.org/dc/terms/", "examples": { "@id": "cr:examples", "@type": "@json" }, "extract": "cr:extract", "field": "cr:field", "fileObject": "cr:fileObject", "fileSet": "cr:fileSet", "format": "cr:format", "includes": "cr:includes", "isLiveDataset": "cr:isLiveDataset", "jsonPath": "cr:jsonPath", "key": "cr:key", "md5": "cr:md5", "parentField": "cr:parentField", "path": "cr:path", "recordSet": "cr:recordSet", "references": "cr:references", "regex": "cr:regex", "repeated": "cr:repeated", "replace": "cr:replace", "sc": "https://schema.org/", "separator": "cr:separator", "source": "cr:source", "subField": "cr:subField", "transform": "cr:transform", "prov": "http://www.w3.org/ns/prov#" }, "@type": "sc:Dataset", "@id": "turbosens1", "conformsTo": "http://mlcommons.org/croissant/1.0", "name": "TurboSens1", "alternateName": [ "turbosens1" ], "description": "TurboSens1 is 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. TurboSens1 ships with health correlated maintenance, four heterogeneous degradation archetypes (A compressor, B fan and booster, C turbine, D balanced), six discrete maintenance actions, twelve flight phase contexts, ten stochastic event types, and imperfect restoration. A second progressively challenging scenario (TurboSens2) adds a fouling channel, phantom observation map, region archetypes, and an engine wash action; it is released as a separate dataset.", "license": "https://creativecommons.org/licenses/by/4.0/", "url": "https://huggingface.co/datasets//turbosens1", "version": "1.0.0", "datePublished": "2026", "isLiveDataset": false, "creator": { "@type": "sc:Organization", "name": "" }, "publisher": { "@type": "sc:Organization", "name": "" }, "citeAs": ". TurboSens1: An interactive turbofan benchmark for probing world model representations against ground truth. 2026. https://huggingface.co/datasets/BdezuSZo/turbosens1", "keywords": [ "world models", "representation learning", "probing", "synthetic dataset", "turbofan", "physical understanding", "POMDP", "counterfactual evaluation" ], "rai:dataCollection": "TurboSens1 is a fully synthetic dataset. All episodes are generated by the OpenDeckSMR thermodynamic turbofan engine simulator under deterministic seeds. No human or sensor data from real engines is collected, processed, or stored. Each episode is parameterised by a (degradation archetype, seed) pair and the simulator advances the latent mechanical wear state, the maintenance policy, and the sensor observation map at every flight cycle. The simulator version is stamped in each released file as turbosens1@v1.0.0; bit identical regeneration is supported via the released generator script.", "rai:dataCollectionType": "Synthetic", "rai:dataAnnotationProtocol": "No human annotation. Ground truth labels (mechanical state, action, event mask, archetype) are produced directly by the simulator at the same step as the corresponding sensor observation, ensuring exact pairing.", "rai:dataPreprocessingProtocol": "Per step records are concatenated into flat HDF5 tensors with episode boundaries given by the ep_offset / ep_len arrays. Sensors are stored at native simulator scale; baseline training pipelines apply z normalisation per channel at load time. Event labels follow an abstract effect typed scheme (perm_step_*, trans_drift_*, trans_anomaly_*, trans_fouling_*, sensor_pulse_*) consistently across all splits.", "rai:dataReleaseMaintenancePlan": "Versioned release on the HuggingFace Hub. The simulator generator and the patch scripts that update HDF5 attributes in place are released alongside in the GitHub repo. Future versions will increment the SIM_VERSION attribute (e.g., turbosens1@v1.1.0); breaking changes to the schema or seed semantics will trigger a new dataset version on the Hub. Bug fixes that preserve schema and seeds will be patched in place with a changelog.", "rai:dataLimitations": "[\"Synthetic data: stochastic event probabilities and magnitudes are mathematical abstractions and in no case reflect what is observed in real life flight settings; calibrating them to operational data is left for future work.\", \"Single domain: the simulator models a single turbofan engine class; cross domain generalisation claims should not be made from TurboSens alone.\", \"Linear probe sufficiency: the inverse probing protocol assumes a small supervised head is expressive enough to read out the state if the encoder has preserved it. Encoders that encode the state in a non linearly decodable form will appear to fail at probing; this is informative, not necessarily a reflection of representation quality in absolute terms.\"]", "rai:hasSyntheticData": true, "rai:sourceDatasets": [ "TurboSens1 derives from no external dataset. The simulator generator (OpenDeckSMR thermodynamic turbofan engine model) is a closed analytical model parameterised by a (degradation archetype, seed) pair; the dataset is fully reproducible from these seeds plus the simulator version stamp turbosens1@v1.0.0. Generator code and seeds are released alongside the dataset in the companion GitHub repository." ], "rai:provenanceActivities": [ "Synthetic data generation: each episode is produced by deterministically advancing the OpenDeckSMR turbofan simulator under a fixed seed and a sampled degradation archetype, recording at every flight cycle the latent mechanical wear state, the maintenance action, the event mask, and the sensor observation map. No human or sensor data from real engines is involved.", "Preprocessing: per-step records are concatenated into flat HDF5 tensors with episode boundaries given by the ep_offset and ep_len arrays. Sensors are stored at native simulator scale; baseline pipelines apply z-normalisation per channel at load time.", "Annotation: ground truth labels (mechanical state, action, event mask, archetype) are produced directly by the simulator at the same step as the corresponding sensor observation, so pairing is exact and no human annotation is involved." ], "rai:dataBiases": "[\"Degradation archetype prior is intentionally skewed (A compressor over represented, D balanced under represented). This reflects the failure mode emphasis chosen for the benchmark and does not represent any real engine fleet's actual failure distribution.\"]", "rai:personalSensitiveInformation": "None. The dataset contains no personal data, no user identifiable information, no sensor recordings from real engines, no operator activity logs, and no proprietary commercial data. All values are simulator outputs.", "rai:dataSocialImpact": "Direct social impact is minimal because the dataset is fully synthetic and not intended for operational decision making on real aircraft. We explicitly note that TurboSens1 must not be used to draw conclusions about real engine reliability, real fleet failure rates, or real maintenance scheduling. Misuse risk: results from TurboSens1 should not be quoted in operational or regulatory contexts as if they referred to real engines.", "rai:dataUseCases": "[\"Probing self supervised world model representations against ground truth latent state.\", \"Evaluating out of distribution generalisation under controlled, well documented distribution shifts.\", \"Studying counterfactual fidelity of trained world model predictors against a deterministic simulator under maintenance interventions.\", \"Benchmarking joint embedding predictive architectures, recurrent latent dynamics models, and reconstruction based world models on a common protocol with a shared ground truth signal.\"]", "rai:safetyMeasures": "The release includes documentation flagging that the dataset is a research instrument, not a real fleet record. Released figures and tables in the accompanying paper are explicit that event probabilities and magnitudes are not calibrated to real failure statistics.", "rai:datasetIntendedUse": "Research on world model representation evaluation. The dataset is intended for academic and industrial research use under the CC BY 4.0 license. It is not intended for operational use in safety critical aviation contexts.", "distribution": [ { "@type": "cr:FileObject", "@id": "turbosens1_train", "name": "turbosens1_train.h5", "description": "TurboSens1 train split: 400 episodes of 10000 flight cycles each, 7 sensors x 12 contexts, 6 maintenance actions, 10 stochastic events.", "encodingFormat": "application/x-hdf5", "contentUrl": "turbosens1_train.h5", "sha256": "7d2fe5311dfdcf9662951b4257dc16d4e78c572e30f3be7458251c82b45d3b57" }, { "@type": "cr:FileObject", "@id": "turbosens1_test", "name": "turbosens1_test.h5", "description": "TurboSens1 in distribution test split: 40 held out episodes of 10000 flight cycles each, drawn from the same generator and archetype prior as the train split.", "encodingFormat": "application/x-hdf5", "contentUrl": "turbosens1_test.h5", "sha256": "0001d18700bb8957ee6228e71cc8cb45427aa91f63480cd343fb8ddb72196033" }, { "@type": "cr:FileObject", "@id": "turbosens1_test_hard", "name": "turbosens1_test_hard.h5", "description": "TurboSens1 out of distribution test_hard split: 40 episodes with stronger degradation and modified maintenance behaviour to assess generalisation beyond the training manifold.", "encodingFormat": "application/x-hdf5", "contentUrl": "turbosens1_test_hard.h5", "sha256": "76fe83dc2cba3265636a36bb2184f6025c49329be8e4fa7a39a10d1f9c47f599" } ], "recordSet": [ { "@type": "cr:RecordSet", "@id": "turbosens1_records", "name": "turbosens1_records", "description": "Per timestep records of TurboSens1. Each row is one flight cycle with paired sensor observations and ground truth state. Episode boundaries are given by ep_offset / ep_len.", "field": [ { "@type": "cr:Field", "@id": "turbosens1_records/sensors", "name": "sensors", "description": "Sensor observation tensor o_t (7 sensors x 12 flight phase contexts) at this flight cycle.", "dataType": "sc:Float", "repeated": true, "source": { "fileObject": { "@id": "turbosens1_train" }, "extract": { "jsonPath": "$.sensors" } } }, { "@type": "cr:Field", "@id": "turbosens1_records/state", "name": "state", "description": "Ground truth mechanical wear vector s_t (10 dimensional). Probing target. Five subsystems x two channels (efficiency, corrected flow).", "dataType": "sc:Float", "repeated": true, "source": { "fileObject": { "@id": "turbosens1_train" }, "extract": { "jsonPath": "$['observation.state']" } } }, { "@type": "cr:Field", "@id": "turbosens1_records/action_idx", "name": "action_idx", "description": "Discrete maintenance action index. 0=do_nothing, 1=fan_overhaul, 2=hpc_overhaul, 3=turbine_overhaul, 4=full_overhaul, 5=targeted_patch.", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "turbosens1_train" }, "extract": { "jsonPath": "$.action" } } }, { "@type": "cr:Field", "@id": "turbosens1_records/event_mask", "name": "event_mask", "description": "Whether any stochastic event fired at this flight cycle.", "dataType": "sc:Boolean", "source": { "fileObject": { "@id": "turbosens1_train" }, "extract": { "jsonPath": "$.event_mask" } } }, { "@type": "cr:Field", "@id": "turbosens1_records/event_types", "name": "event_types", "description": "Per event type rising edge flag. Index into dataset attribute event_names (10 events for TurboSens1).", "dataType": "sc:Integer", "repeated": true, "source": { "fileObject": { "@id": "turbosens1_train" }, "extract": { "jsonPath": "$.event_types" } } }, { "@type": "cr:Field", "@id": "turbosens1_records/weather", "name": "weather", "description": "Ambient temperature deviation Delta T_amb at this flight cycle.", "dataType": "sc:Float", "source": { "fileObject": { "@id": "turbosens1_train" }, "extract": { "jsonPath": "$.weather" } } }, { "@type": "cr:Field", "@id": "turbosens1_records/ep_offset", "name": "ep_offset", "description": "Episode level: starting index in the flat per timestep tensors for each episode.", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "turbosens1_train" }, "extract": { "jsonPath": "$.ep_offset" } } }, { "@type": "cr:Field", "@id": "turbosens1_records/ep_len", "name": "ep_len", "description": "Episode level: length in flight cycles of each episode.", "dataType": "sc:Integer", "source": { "fileObject": { "@id": "turbosens1_train" }, "extract": { "jsonPath": "$.ep_len" } } } ] } ], "prov:wasDerivedFrom": [ { "@id": "https://github.com/OpenDeckLab/OpenDeckSMR/tree/main", "prov:label": "OpenDeckSMR", "sc:license": "CC-BY-NC-ND", "prov:wasAttributedTo": { "prov:label": "OpenDeckLAB" } }, { "@id": "https://huggingface.co/datasets/BdezuSZo/turbosens1", "prov:label": "Synthetic generation, seed 42.", "sc:license": "CC-BY-4.0." } ], "prov:wasGeneratedBy": [ { "@type": "prov:Activity", "prov:type": { "@id": "https://www.wikidata.org/wiki/Q5227332" }, "prov:label": "Simulation generation", "sc:description": "Defining the dataset scenario: archetypes, degradations, contexts and tests to validate the requirements." } ] }