Time-invariant properties and principal components of in-situ measurements used for outlier detection in space missions
Jonah Ekelund, Savvas Raptis, V. Toy, Wenli Mo, Drew L. Turner, I. J. Cohen, Stefano Markidis · Journal of Computational Science · 2026
The examination of time series data with multiple features is vital for space missions, where efficient event detection is crucial, especially for automatic onboard analysis. Nevertheless, the constraints of onboard computational resources and data transmission necessitate robust methods for real-time identification of regions of interest. This study introduces an adaptive algorithm for outlier detection, leveraging the reconstruction error from Principal Component Analysis (PCA) for feature reduction, specifically tailored for space mission scenarios. The algorithm operates on a temporal window of data and dynamically adapts to changing data distributions with the use of Incremental PCA, allowing for implementation without a pre-existing model accounting for all potential conditions. A pre-scaling step normalizes the magnitude of each feature while maintaining the relative variance within feature categories. We show that only eight components are needed to preserve over 92.9\% of the variance in the data, accounting for the time invariant changes in the omnidirectional ion spectrum, independent of the length of the time window. We demonstrate the algorithm's utility in detecting space plasma events, including transitions between space plasma environments, such as bow shock or magnetopause crossings, and transient phenomena, such as foreshock transients, utilizing data from NASA's MMS mission. Furthermore, the method is applied to data from NASA's THEMIS mission, successfully identifying a dayside transient with data available onboard. This is an extension of previous work presented in Ref.[1].