Robustness Evaluation Method for Spacecraft Anomaly Detection Models

The anomaly detection of spacecraft is critically important for the safety of spacecraft, and for a specific spacecraft, the most suitable anomaly detection method must be chosen. This paper proposes an evaluation method for the robustness of anomaly detection models. Seven common evaluation metrics are calculated for the prediction results of the anomaly detection models. These metrics can be viewed as a point in a seven-dimensional space. Experiments are conducted using data with varying levels of noise, and metric points are calculated. Subsequently, multiple sets of metric points are clustered. The clustered metric points yield average coordinate points representing each cluster of metrics. Experiments are also conducted using noise-free data to calculate metric points as baseline points. The distance between the baseline points and other metric points is then computed, where a smaller distance indicates better model robustness. Experimental results demonstrate that this evaluation method is applicable for assessing the robustness of anomaly detection models for spacecraft.