Pattern recognition in time series for space missions: A rosetta magnetic field case study. (March 2020)

• Record Type: Journal Article
• Authors: Ostaszewski, K.; Heinisch, P.; Richter, I.; Kroll, H.; Balke, W.-T.; Fraga, D.; Glassmeier, K.-H.
• Is Part Of: Acta astronautica. Volume 168(2020)
• Page Start: 123
• Page End: 129
• Publication Date: 2020-03
• Subjects: Rosetta -- Time series -- Machine learning -- Magnetic field -- Pattern recognition; Astronautics -- Periodicals; Outer space -- Exploration -- Periodicals; Astronautics; Periodicals; 629.405
• DOI: 10.1016/j.actaastro.2019.11.037
• Languages: English
• ISSNs: 0094-5765

Abstract

Abstract: Time series analysis is a technique widely employed in space science. In unpredictable environments like space, scientific analysis relies on large data sets to enable interpretation of observations. Artificial signal interferences caused by the spacecraft itself further impede this process. The most time consuming part of these studies is the efficient identification of recurrent pattern in observations, both of artificial and natural origin, often forcing researchers to limit their analysis to a reduced set of observations. While pattern recognition techniques for time series are well known, their application is discussed and evaluated primarily on purpose built or heavily preprocessed data sets. The aim of this paper is to evaluate the performance of state of the art pattern recognition techniques in terms of computational efficiency and validity on a real-life testcase. For this purpose the most suitable techniques for different types of pattern are discussed and subsequently evaluated on various hardware in comparison to manual identification. Using magnetic field observations of the ESA Rosetta mission as a representative example, both disturbances and natural patterns are identified. Compared to manual selection a speed-up of a factor up to 100 is achieved, with values for recall and precision above 80%. Moreover, the detection process is fully automated and reproducible. Using the presented method it was possible to detect and correct artificial interference. Finally, the feasibility of onboard deployment is briefly discussed. Highlights: Patterns of technical and scientific interest were efficiently identified using machine learning. Compared to manual detection a speedup of 100 was achieved. Artificial disturbances in the magnetic field were automatically corrected. The application is not limited to magnetic fields.