BdezuSZo/turbosens1

TurboSens1是TurboSens涡扇发动机退化基准的第一个场景，这是一个交互式模拟器数据集，它将传感器观测数据与每个飞行周期中发动机的完整十维真实健康状态配对，从而支持任何自监督世界模型的直接逆探测协议。数据集包含三个分割：训练集（400个情节，总计4,000,000个飞行周期）、测试集（40个情节，总计400,000个飞行周期）和困难测试集（40个情节，总计400,000个飞行周期），每个飞行周期的长度上限为10,000。数据以HDF5格式存储，列包括传感器流（形状为(N, 7, 12)）、观测状态（形状为(N, 10)）、动作、事件掩码、事件类型、天气、情节偏移和长度等，并包含文件级属性如场景名称、动作名称、原型名称等。数据集用于时间序列预测任务，特别是涡扇发动机的退化建模和表示学习，通过逆探测协议评估自监督世界模型的性能。注意事项：数据是合成的，不反映真实机队的校准；仅限于单一领域，不能用于跨领域泛化；协议假设线性探测足够表达状态。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. The dataset includes three splits: train (400 episodes, total 4,000,000 flights), test (40 episodes, total 400,000 flights), and test_hard (40 episodes, total 400,000 flights), with each flight cycle capped at 10,000 length. Data is stored in HDF5 format with columns such as sensors (shape (N, 7, 12)), observation.state (shape (N, 10)), action, event_mask, event_types, weather, ep_offset, ep_len, and episode metadata. File-level attributes include scenario name, action names, archetype names, event names, and sensor names. It is intended for time-series forecasting tasks, specifically turbofan degradation modeling and representation learning, and supports evaluation via an inverse probing protocol for self-supervised world models. Caveats: the data is synthetic and not a calibration of any real fleet; it is single-domain and should not be used for cross-domain generalization claims; the protocol assumes linear probe sufficiency for state decoding.