# Spacecraft Anomaly Challenge on ESA dataset — Data

来源：https://www.kaggle.com/competitions/esa-adb-challenge/data

采集时间：2026-10-19T08:54:17Z

## TLDR

The dataset represents several years of real sensor measurements collected from a large spacecraft operated by the European Space Agency. It is a single long continuous multivariate time series data of 87 mission-critical parameters. The dataset was carefully annotated for anomalies and rare events based on the original flight control reports and several iterations of manual and algorithm-driven refinement of labels [1]. 

Your task is to find all anomalous time points in the test set (binary decision, 0 - normal point, 1 - anomalous point, see the `sample_submission.parquet` file) while minimizing the number of false alarms.

## List of files

*   **train.parquet** - the training set; 14 years of ESA-Mission1 data from the [ESA-AD dataset](https://zenodo.org/records/12528696) with annotated anomalous time points (`is_anomaly` column)
*   **test.parquet** - the test set; a never before published 6-months long fragment of ESA-Mission1 data
*   **target_channels.csv** - a list of names of 58 channels monitored for anomalies (other 29 channels are just external auxiliary variables)
*   **sample_submission.parquet** - a sample submission file in the correct format

## More details


### Preprocessing

The dataset is a simplified and preprocessed version of the raw ESA-Mission1 data from the ESA-AD dataset publicly available on [Zenodo](https://zenodo.org/records/12528696). The preprocessing is described together with all other dataset details in the ESA-ADB paper [2] and the preprocessing script is available on the official ESA-ADB GitHub repository [here](https://github.com/kplabs-pl/ESA-ADB/blob/main/notebooks/data-prep/Mission1_semisupervised_prep_from_raw.py). For simplicity, timestamps have been replaced with integer index and annotations have been aggregated across all channels into the single `is_anomaly` column, such that `1` is marked for a time point if at least one channel is affected by an anomaly or a rare event. **Participants are welcome to propose their own preprocessing and annotations aggregation methods**, however, the test set is already preprocessed and aggregated using our described methods.

### Channels 

The dataset contains 87 channels, among which 58 are target channels (monitored for anomalies) listed in the `target_channels.csv` file, 18 channels are auxiliary environmental variables (not monitored for anomalies), and 11 are telecommands (binary control commands sent by operators; marked with the `telecommand_` prefix). The subset of 6 channels 41-46 is recommended as a good starting point for developing and testing algorithms before applying them to the full set. 

These 87 channels are just a small subset of real spacecraft telemetry with thousands of channels. However, this subset was designed by spacecraft operators to be a self-contained representation of the most important mission-critical and auxiliary variables necessary to identify anomalies.

### Events to detect

ESA-ADB dataset contains three categories of events: anomalies, nominal rare events, and communication gaps. The test set does not contain communication gaps. Your task is to detect both anomalies and rare nominal events.

### References

[1] K. Kotowski, C. Haskamp, B. Ruszczak, J. Andrzejewski, and J. Nalepa, “Annotating Large Satellite Telemetry Dataset For ESA International AI Anomaly Detection Benchmark,” in Proceedings of the 2023 conference on Big Data from Space, Vienna: Publications Office of the European Union, Nov. 2023, pp. 341–344. doi: [10.2760/46796](https://dx.doi.org/10.2760/46796).