Spacecraft Thruster Firing Test Dataset

WARNING

This version of the dataset is not recommended for anomaly detection use case. We discovered discrepancies in the anomalous sequences. A new version will be released. In the meantime, please ignore all sequence marked as anomalous.

CONTEXT

Testing hardware to qualify it for Spaceflight is critical to model and verify performances. Hot fire tests (also known as life-tests) are typically run during the qualification campaigns of satellite thrusters, but results remain proprietary data, hence making it difficult for the machine learning community to develop suitable data-driven predictive models. This synthetic dataset was generated partially based on the real-world physics of monopropellant chemical thrusters, to foster the development and benchmarking of new data-driven analytical methods (machine learning, deep-learning, etc.).

The PDF document "STFT Dataset Description" describes in much details the structure, context, use cases and domain-knowledge about thruster in order for ML practitioners to use the dataset.

PROPOSED TASKS

Supervised:

Performance Modelling: Prediction of the thruster performances (target can be thrust, mass flow rate, and/or the average specific impulse)

Acceptance Test for Individualised Performance Model refinement: Taking into account the acceptance test of individual thruster might be helpful to generate individualised thruster predictive model

Uncertainty Quantification for Thruster-to-thruster reproducibility verification, i.e. to evaluate the prediction variability between several thrusters in order to construct uncertainty bounds around the prediction (predictive intervals) of the thrust and mass flow rate of future thrusters that may be used during an actual space mission

Unsupervised / Anomaly Detection

Anomaly Detection: Anomalies can be detected in an unsupervised setting (outlier detection) or in a semi-supervised setting (novelty detection). The dataset includes a total of 270 anomalies. A simple approach is to predict if a firing test sequence is anomalous or nominal. A more advanced approach is trying to predict which portion of a time series is anomalous. The dataset also provide a detailed information about each time point being anomalous or nominal. In case of an anomaly, a code is provided which allows to diagnosis the detection system performance on the different types of anomalies contained in the dataset.