Data Mining for Early Fault Detection in Artificial Satellites: A Review

Early detection of faults in artificial satellites is crucial for the success of missions, but it is hindered by the scarcity of data on faults and the limitations of traditional monitoring methods. As an alternative, technological trends such as data mining have emerged. This study presents a literature review to provide an in-depth examination of the landscape of data mining applications for early fault detection in satellites. Following the PRISMA protocol, the available scientific corpus from seven scientific databases was reviewed, and 52 primary studies were selected from an initial set of 2726 records published between 2011 and 2024. The results indicate that this is a rapidly expanding field, with an annual growth rate of 35.72%, characterized by a significant geopolitical concentration of research and funding led by China. From a methodological point of view, unsupervised approaches (~40%) predominate, a response to the lack of labeled in-flight data. However, supervised and hybrid models demonstrate superior performance, achieving F1 scores above 97% when selected or simulated data are available. A misalignment was identified in the domain, although research clearly favors the EPS due to the availability of data. Operational statistics, however, confirm that the ADCS system is the primary cause of mission failure. It is essential to note that the limited availability of public datasets and models, with less than 15% of studies providing access, is the main obstacle to reproducibility and progress. The primary conclusion of this work is that the field is expanding, and all stakeholders must contribute to its continued growth. Key actions include establishing public benchmarks that enable transparent evaluation, exploring physics-based models that account for uncertainty to address data scarcity, and concerted efforts to bridge the transfer gap from academic satellite operations to the real world.